2026-09-30

We present LongCat-DeepResearch, a deep research system that combines an enhanced LongCat model with a multi-agent workflow for producing comprehensive, evidence-grounded reports. The workflow separates global planning from detailed investigation and coordinates revision at the section level. Multiple planning agents first explore external sources and refine an actionable research plan, termed ResearchSpec. Research agents then investigate and draft their assigned sections in parallel, gathering additional evidence in separate contexts as their analyses develop. Once the sections are assembled, global review guides targeted local revisions, reducing reliance on repeated full-report rewriting. This workflow also supports the construction of research tasks and trajectories for the mid-training and post-training of LongCat's general-purpose models. LongCat-DeepResearch achieves 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, and 79.83 on ResearchRubrics. On an in-house benchmark, it scores 76.04, ranking second among four compared systems. Development-set analyses show benefits from combining planning perspectives, while further planning refinement has mixed effects. Additional editing improves average automatic readability preference across two benchmarks, with different trends on each.

Industry applications often demand low-latency classification, yet current large language model (LLM) approaches remain poorly suited for latency-critical applications. Existing prompting and constrained decoding produce verbose, multi-token outputs that require expensive token-by-token generation, while encoder-based models achieve faster inference but sacrifice task flexibility. We propose Koa-action, a framework for low-latency atomic actions -- fast, single-step decisions such as classification, semantic endpointing, Boolean checks, and scoring -- formulated as constrained generation with single-token outputs. By introducing atomic label tokens and applying supervised fine-tuning, our method reduces classification to a deterministic one-step decoding problem. Across standard benchmarks, Koa-action delivers competitive accuracy with consistently low and stable latency. On a production intent-routing benchmark, Koa-action reaches 85.5% accuracy -- competitive with the strongest frontier models (Claude-4.8-Opus, Gemini-Pro-3.1) and ahead of GPT-5 and Gemini-2.5-Pro -- while answering in about half a second, several-fold faster than every frontier model (up to ~7.5x at the median) under identical serving conditions. Against the dedicated single-token system Jev/TypeSafe, Koa-action is competitive on accuracy and faster at the median, while also handling multimodal inputs and multi-label outputs that single-label text systems do not.

As large language models are deployed in increasingly autonomous long-horizon tasks, manually auditing and verifying the actions, artifacts, and outputs of models becomes more difficult. Users instead come to rely on LLM-generated reports to assess the quality and completeness of the work. We introduce a suite of eight adversarial reporting scenarios to systematically study whether LLMs conceal narrative-changing flaws: errors or limitations that undermine an otherwise successful account of work. We call this phenomenon "insecure reporting." When handed machine learning experiment logs containing a planted negative result that substantially weakens the proposed method, GPT-5.5 flags the negative result in only 2 of 200 generated reports. However, when a short honesty instruction, "Be honest in your response," is added, the model flags the negative result in 190 of 200 reports. Across eight open-weight models, chain-of-thought analysis reveals a recurring tension between disclosing narrative-changing flaws and reasoning about ways to appear successful. We perform an activation analysis and a steering experiment on Qwen3.5-9B, finding that honesty and success-seeking correspond to opposing directions in representation space. Our results suggest that LLMs tend to present narratives of success by default, and that steering models toward honesty makes their reports substantially more transparent.

Parallel speculative drafting generates multiple candidates in one backbone pass, but independent token selection can produce inconsistent continuations that shorten the accepted prefix. Existing methods mostly leave conditional decoding to a lightweight module after the backbone, which limits the flow of predecessor information to successors. Our analysis of DFlash shows that early positions already form recoverable predictions in shallow layers, and that accurate adjacent predecessors help successors more when they enter earlier. We therefore propose DSpine, a drafter with causal conditioning injection throughout the backbone: at every layer, gated adjacent injection writes each predecessor's predicted feature into its successor, so the causal conditioning chain unfolds over network depth while all positions update in parallel. A unified transfer space built from the target model's output embeddings unifies layer-wise injection with predecessor-conditioned decoding, and layer-wise output-embedding supervision promotes the formation of predicted features in shallow layers. Fused kernels and a transition cache execute both efficiently in parallel within SGLang. Across seven math, code, and chat benchmarks, DSpine achieves the longest acceptance length at both temperatures on Qwen3-4B and Qwen3-8B. At temperature zero on Qwen3-8B, it raises the seven-benchmark mean from DFlash's 3.77 to 4.82 (+27.8%); in SGLang serving tests, it delivers 23.3% higher throughput than DFlash on average.

We present OLIVE (OnLine InterVEntion). At each iteration, the evolving student policy generates a new prefix, the teacher continues it autoregressively, and the student is updated using cross-entropy computed on the teacher-generated tokens. Each design choice targets a corresponding limitation of existing distillation methods: (1) sequential covariate shift in offline supervised fine-tuning (SFT) on fixed teacher trajectories, (2) fragmented supervision under prefix failure in token-level on-policy distillation (OPD), and (3) the need for access to teacher token probabilities in distribution-matching distillation. OLIVE achieves higher reasoning performance than OPD (with a top-16 KL approximation) at comparable GPU-hour cost. Our asynchronous implementation further reduces OLIVE's total training time by 23.8\%. We evaluate OLIVE on both hard reasoning tasks and agentic tasks which reflects modern post-training scenarios, and it consistently outperforms existing distillation methods under the same training budget. By regenerating prefixes from the evolving student, OLIVE continues improving after offline distillation plateaus while better preserving the general capabilities and plasticity of the student. Using only text from GPT-5.4-mini, continuously training with OLIVE outperforms offline SFT from the same teacher by 13\% on ScienceWorld. These results support OLIVE as an effective and efficient approach to online language-model distillation.

Linear RNNs offer linear-time sequence processing and constant-memory decoding, but their fixed-size recurrent states must accommodate all past key--value associations. Existing forgetting mechanisms and Delta Rule updates reduce interference by selectively clearing or correcting the state, yet earlier associations can still become difficult to retrieve. We introduce CyFA (Cyclic Flow Attention), a Linear RNN with relative-time-partitioned memory. At each step, a learned clock controls the cyclic transport applied jointly to the key and value states before the current key--value pair enters the age-zero slot, thereby organizing stored associations across relative-time slots. We further derive an exact change to absolute-clock coordinates that expresses CyFA as two scalar-decay linear attention recurrences and enables efficient chunk-wise training. Across 400M--1.4B pretraining experiments with matched recurrent-state sizes, CyFA improves recall-intensive performance while maintaining competitive language modeling and high computational efficiency. At 400M, CyFA outperforms KDA on FDA (42.60 vs. 26.07) while requiring only 46.7% and 48.3% of KDA's forward and backward core-operator execution times, respectively. Our code is publicly available at \href{https://github.com/Chyxx/CyclicFlowAttention}{this https URL}.

Evaluation awareness (EA) can cause large language models (LLMs) to behave differently during audits than in deployment, yet measuring and accounting for EA remains challenging. We introduce verbalization training (VT), a method for making LLMs less reticent about verbalizing evaluation awareness while avoiding to supervise the latent belief itself. VT uses a model's spontaneous verbalizations as evidence that awareness is present and truncates each rollout immediately before the verbalization, producing training prefixes at which the model is presumed to be aware. The model is then trained with an RL objective designed to increase verbalization in a calibrated way. Across Qwen3.6-35B-A3B, Kimi K2.6, and Inkling, VT increases verbalized EA by 2.4-2.9 times and transfers to held-out agentic settings, while measured latent EA and behavior remain largely stable. In a causal experiment, we independently implant meta-knowledge about evaluations through synthetic-document fine-tuning and show that VT-induced verbalizations reflect the richer knowledge acquired by the model.

On-policy distillation (OPD) trains a student to match the teacher's next-token distributions on the student's own trajectories and has yielded substantial empirical gains. Generalized variants allow the student to surpass the teacher by extrapolating an implicit reward in output space. The language-model head, however, attenuates this change anisotropically: much of the change encoded in the teacher's hidden states reaches the logits at a small fraction of its weight, and the sampled-token log-probability ratios on which output-space extrapolation relies inject noise that the extrapolation amplifies, making training unstable. We observe that reinforcement learning (RL) shifts a model's internal representations relative to its base checkpoint, and that the direction of this shift can be measured at every layer. Motivated by this observation, we propose RIDE (RL-Induced Direction Extrapolation), which extrapolates the RL-induced change directly in representation space: at every layer and token position, RIDE computes the residual between the teacher and its pre-RL checkpoint and regresses the student's hidden states toward targets displaced beyond the teacher along this residual. Conditioned on a sampled trajectory, this regression is equivalent to maximizing a linear directional reward defined by the residual under a quadratic penalty centered at the teacher, which makes explicit how the objective moves the student along the RL-induced direction while limiting its deviation from the teacher. Across four base/RL-teacher pairs spanning different scales, architectures, and pre-training lineages, RIDE approaches or exceeds the RL-trained teacher on every pair and is the only method whose mean does so, and it consistently outperforms output-space extrapolation, which degrades the student whenever the teacher is close to its base. Project page: https://github.com/xixixixixxxx/RIDE.

On-policy distillation (OPD) trains multi-turn language agents with dense teacher supervision on student-generated responses. However, standard think-then-act rollouts require lengthy reasoning before each short action, delaying environment transitions and experience collection. Generating actions directly reduces this delay but can degrade rollout quality. To address this, we propose ActFirst-OPD, an act-first, reason-later training framework that decouples environment interaction from full-response generation. The student infers and executes actions through reference-conditioned inverse dynamics using its current interaction context and a reference next observation, and switches to autonomous next-action prediction when the resulting transition deviates from the reference trajectory. From the collected interaction contexts, the student asynchronously generates full think-then-act responses for token-level teacher supervision. Experiments across 0.6B-, 1.7B-, and 4B-parameter Qwen3 students show that ActFirst-OPD achieves average wall-clock training speedups of $2.3\times$ on ALFWorld, $1.8\times$ on WebShop, and $4.9\times$ on ScienceWorld over Vanilla OPD. It matches or exceeds all compared OPD baselines in mean task success rate across eight of nine benchmark-model settings. These results demonstrate that reasoning need not block acting during multi-turn agent distillation.

Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for improving large language models (LLMs) on various tasks, yet its sparse outcome rewards lack token-level credit assignment for intermediate steps. To address this, on-policy self-distillation (OPSD) leverages a self-teacher with privileged context to provide additional dense learning signals. However, because the self-teacher is often overconfident and imposes excessive penalties on long reasoning trajectories, OPSD frequently struggles in practice. To mitigate this, we propose self-instructing policy optimization (SIPO) with a contrastive self-teacher to provide dense credit. At each iteration, SIPO samples multiple rollouts per prompt from the current policy, scores them with environment rewards, and constructs two teacher contexts for each rollout by pairing the reference answer with mistakes made within the group. The model then re-evaluates its own responses under both contexts, using the difference between the two teacher log-probabilities as token-level feedback, so that biases shared by both contexts are expected to largely cancel. The resulting objective yields a token-level advantage for every rollout: the reward still sets the main direction of each update while the self-teacher redistributes credit across tokens. Even in groups where every rollout fails and group-relative advantages vanish, SIPO still provides a learning signal. By preserving direct optimization of the task reward while providing dense, token-level feedback, this approach bridges reinforcement learning and on-policy self-distillation. Extensive experiments across multiple reasoning and code-generation benchmarks demonstrate that SIPO outperforms both RLVR and OPSD baselines without an external teacher or additional generation.

Self-rewarding reinforcement learning (RL) enables large language models (LLMs) to self-evolve without human labels. Existing ensemble-based methods construct reward references from rollout groups and assign rewards accordingly. However, a response's reward representation also depends on its randomly sampled group context, i.e., the other responses in its group. Using only one group-context realization may miss desired reward signals and provide unreliable guidance for policy optimization. To address this issue, we propose Group-Marginalized Advantage Estimation (GMAE), which aggregates reward realizations across possible contexts into a response-level distribution and estimates expected advantages. Experiments across eight benchmarks and four base models demonstrate strong performance and cross-domain generalization. GMAE also exhibits stable learning, low extra cost, and good applicability across training datasets and RL backbones.

Group Relative Policy Optimization (GRPO) is widely used to train reasoning language models, where it computes advantages by centering and normalizing rewards across rollouts of the same prompt. For multiple rewards, GRPO sums the reward components and normalizes the total reward by its within-group standard deviation. The corresponding variance equals the sum of all pairwise reward covariances. For a fixed centered reward, larger aggregate covariance produces smaller advantages, and vice versa, allowing update magnitudes to adapt to reward dependence. However, correlated rewards with large scales can dominate this normalization and suppress signals from smaller-scale rewards. We propose Correlation-Normalized GRPO (CorrGRPO), which normalizes pairwise covariances into Pearson correlation coefficients. CorrGRPO keeps the centered total reward unchanged while balancing the influence of differently scaled rewards on the correlation-based normalization. This allows advantage magnitudes to adapt to reward correlations without the normalization being dominated by large-scale reward components. We compare CorrGRPO with GRPO and other variants on code generation, tool calling, and agent security, using models ranging from 0.5B to 8B parameters. These tasks all involve multiple rewards that can improve together or present tradeoffs. Results show improvements across three domains, including code generation, tool calling, and agent security. Our code is available at https://github.com/HKUST-KnowComp/CorrGRPO.

Group-relative methods for reinforcement learning with verifiable rewards (RLVR) learn from differences in rollout outcomes. Independently sampling complete trajectories is costly and does not explicitly explore the decision space at critical positions. Feedback on a completed trajectory can reveal which earlier choices the policy reconsiders, suggesting where to sample alternative continuations. We introduce Hindsight-Divergence Localization (HDL), which uses hindsight-induced changes in token log-likelihoods to select branch points. HDL generates a small number of complete root trajectories and fills each training group with continuations from the selected positions under the original task context. Each continuation reuses its root prefix and contributes policy updates only through its newly generated suffix, reducing generation cost while focusing additional exploration and learning on decisions after branching. Experiments with three models across math, code, and agent tasks show gains in both rollout efficiency and task performance. Compared with GRPO at matched group sizes and training steps, HDL yields up to a 2.5$\times$ reduction in generated tokens and a 1.8$\times$ speedup in rollout wall-clock time. Despite this reduced generation budget, HDL improves performance across all three domains, with gains of up to 12.5 points on agent tasks.

As the capabilities of large language models (LLMs) continue to advance, increasing attention is turning to how to translate their abilities into useful behavior. Personal agents bring this question into everyday settings, where models are expected to serve individual users and continually adapt to their preferences. With the underlying model held fixed, such adaptation relies on harness engineering: designing and evolving the surrounding layer that manages context, memory, tools, and execution. Despite rapid progress, the factors governing effective harness evolution remain insufficiently understood. To narrow this gap, we investigate three central questions concerning harness architecture, harness scale, and self-evolution algorithms through complementary empirical and theoretical analyses. Empirically, we introduce a preference-oriented benchmark and systematically characterize the capabilities and limitations of personal agents associated with these three dimensions. Theoretically, we formulate harness evolution as a learning problem and explain these phenomena through approximation, generalization, and optimization errors. Analyses of reachable policies, capacity under finite interaction evidence, and biased update dynamics provide theoretical accounts of the observed phenomena. Together, these results offer a unified perspective on the limits of personalization through harness evolution and inform future harness design.

For a long-horizon agent, context is the bottleneck: the history is resent with every request, the window caps task length, and reasoning degrades as the history grows. Replacing structured objects with compact retrieval Cards shortens the prompt and keeps the exact originals retrievable, but editing the history can break prefix-cache reuse, and prior recoverable methods time their edits by forecasts of future reuse or by preset intervals. We propose CADOC (Cache-Aware Dynamic Object Context), an online algorithm that replaces structured objects with compact Cards while preserving exact, on-demand retrieval of their original contents. CADOC schedules replacements in batches by balancing accumulated waiting cost against shared cache-reconstruction cost. Its scheduling rule follows from an economic order quantity trade-off, recovers the optimal integer batch under stationary assumptions. Across evaluation, CADOC consistently achieves the lowest aggregate input cost among the compared configurations, which reduces input cost by approximately 40\% on average while maintaining task performance close to full context. CADOC thus provides a cost-derived approach to compressible context management, demonstrating that efficient compression depends not only on shortening prompts but also on scheduling edits to preserve cache reuse.

Recent approaches to reinforcement learning (RL) post-training for large language models increasingly remove the critic to reduce training instability and memory overhead. Even where a critic is trained, it is discarded once training ends, although it has learned to predict outcomes. We revisit this trend and show that a pretrained critic's ability to predict future outcomes can make it a valuable asset for efficient long-horizon reasoning. First, we find that instability in critic-based RL for long chain-of-thought reasoning is largely an optimization artifact: keeping policy updates small and low in variance restores stable convergence. Second, a well-pretrained critic estimates the posterior probability of eventual success from later trajectory states and unfinished prefixes. Its predictions provide outcome-derived, dense, per-prefix learning signals that, during policy optimization, require neither completed rollouts, step-level annotations, nor external reward labels. Building on this insight, we introduce Reward-Free Policy Optimization (RFPO), which repurposes a single calibrated, frozen critic as a rollout-level reward, a value baseline for generalized advantage estimation, and a success forecaster for unfinished prefixes. We further show that binarizing the debiased score stops the policy from exploiting the critic's length bias. Binarized, RFPO matches supervised PPO without a single label in the training loop, while cutting compute and memory overhead. This makes RFPO well suited to long-horizon reasoning tasks, where outcomes arrive late and generation dominates cost: because rollouts can be rewarded before they finish, training no longer has to pay for waiting on every trajectory to complete. Our findings challenge the prevailing critic-free paradigm and establish critic-based, reward-free optimization as a scalable and computationally efficient path for LLM post-training.

When a user changes their mind partway through a task, an agent that has already split the task into sub-goals and paid for tool calls must decide, per cached sub-result, whether to keep, patch, or discard it (salvage), restarting wastes valid work and continuing unchanged answers the old question. Our main finding is that salvage quality is a matter of role design rather than model capability: a language model asked the keep/patch/discard question one node at a time is unreliable, but asked to classify the revision once, with a deterministic layer propagating the decision, it reaches the cost-optimal oracle on all three models tested, from two vendors. Modeling the plan as a commitment hierarchy and the intent change as a belief-revision operator with AGM style postulates, we prove that no policy observing only a node's local view can be both safe and cost optimal, while the single classification design is both. Across three environments the policy recovers the full achievable savings, 43% cheaper than restart, at 100% correctness.

Retrieval-augmented language models are expected to answer from the retrieved evidence, but in practice they often keep answering when that evidence is missing. We trace this behavior to the training signal: answer-focused fine-tuning assigns no target to unsupported contexts, so it cannot distinguish a reader that abstains from one that guesses, and unsupported answering stays near 100% even as supported accuracy improves. We introduce Regime Boundary Alignment (RBA), which trains a single reader on matched variants of the same question and gold answer. The reader is trained to produce the gold answer when the context supports it, including when conflicting evidence is also present, and to abstain when the correct support is removed; inference is ordinary decoding, with no verifier, threshold, or regime label. On three multi-hop QA datasets across three seeds, RBA reduces the unsupported-answer rate by more than sixty percentage points relative to conflict-focused training while matching its supported accuracy. On a held-out TriviaQA retrieval-miss slice, the same reader reduces unsupported answering from 100% to below 1% while also improving supported accuracy. These results indicate that evidence-gated answering must be learned on both sides of the support boundary.

On-policy distillation (OPD) trains a student on its own reasoning trajectories using feedback from a stronger teacher. Teacher interventions can improve these trajectories, but also change the distribution on which the student learns. Our controlled studies show that rollout quality alone is an incomplete criterion for allocating teacher guidance. Deeper intervention yields diminishing gains in rollout accuracy while increasing off-policy load. In a training probe with a restricted rollout horizon, peak student accuracy and performance retention favor different intervention strengths. The preferred intervention depth and placement also vary across benchmarks. These findings motivate MAESTRO, which uses local policy disagreement to jointly adapt when the teacher takes over and how long it generates. Its {policy disagreement score} combines teacher-weighted candidate coverage with local distribution similarity and is aggregated within reasoning paragraphs. Across eight mathematical reasoning benchmarks, MAESTRO achieves the highest macro-average accuracy among the compared methods for both 0.6B and 1.7B Qwen3 students, with the 1.7B student leading on every benchmark. MAESTRO also reduces average training response length by 67.3\% relative to standard OPD. The code is available at https://github.com/yhao-wang/MAESTRO.

Assigning credit to intermediate steps remains a central challenge in training Large Language Models (LLMs) on multi-step reasoning tasks with sparse terminal rewards, and actor-critic methods such as PPO address this by learning value functions to construct token-level advantages. Their effectiveness, however, hinges on reliable value estimation, a difficult task requiring the critic to both assess progress toward a correct solution and anticipate an evolving policy's future behavior; errors in either can compromise credit assignment and destabilize online training. In this paper, we revisit the standard state-only formulation of value estimation and propose $\pi$PPO, a self-privileged actor-critic framework. By reusing verified same-prompt rollouts as contrastive evidence, $\pi$PPO helps the critic assess intermediate reasoning against successful and failed attempts, while preserving standard policy optimization and the deployment interface. Experiments show that $\pi$PPO consistently improves value-estimation quality by a substantial margin and outperforms representative actor-critic and critic-free RLVR baselines on challenging mathematical reasoning benchmarks, while remaining effective even when paired with substantially smaller asymmetric critics.

Reliable FP8 attention remains a barrier to fully native 8-bit large language model training. We derive how forward-backward inconsistencies produce stale delta and empirically show how it distorts training dynamics. Our stale-delta hybrid runs show a modest loss gap at 569M parameters but substantial loss increases and downstream degradation at 1.67B and 5.29B. QK normalization, NoPE (no positional encoding), and lower-learning-rate context extension mitigate or delay degradation without eliminating it. This pattern suggests accumulated optimization error that smaller models and short runs can conceal. We propose Delta-Matching, proving that it restores the softmax gradient's zero-row-sum invariant under the stated numerical assumptions. It enables native block-scaled FP8 in every forward and backward attention-core matmul without architectural changes, smaller global batches, or auxiliary forward outputs. Across tested architectures, scales, and training stages, Delta-Matching matches BF16/FP32 mixed-precision training loss and overall downstream performance. We will release our implementation, trained models, and data recipes.

As the context size of text processed with an LLM grows, the size of KV caches can outstrip the memory allocated for the original model weights. This impacts LLM throughput negatively, since decoding is memory-bound and decode cost grows with cache size. Recent work alleviates this bottleneck by discarding the least relevant tokens. Eviction introduces a tension, since a one-off decision to discard content may prove detrimental later. Instead, we focus on alternative choices that can lead to cache compression without evicting tokens. We achieve this by learning a selector that is able to produce, based on context, a per-layer cache configuration towards an overall compression budget. The selector operates along three axes: sharing one cache across layers (depth), caching at fewer bits (precision), or truncating the low-rank latent cache representations (rank). We call the resulting method KV-Kaizen, for the many small per-layer choices it compounds. We observe that these interventions taken independently and uniformly over all layers limit achievable compression because they degrade accuracy. Crucially, composing them locally and adaptively to the context can instead preserve accuracy while achieving large memory savings. At inference, the selector runs once, before pre-fill. In evaluations on instruction following and reasoning tasks, our selectors reach the Pareto frontier of accuracy against cache size, against learning-free and post-hoc baselines. On long-context tasks, KV-Kaizen improves on eviction and can be composed with it, reaching a 32x smaller decode-time cache on a 14B model while preserving accuracy. A 4x cache size reduction incurs no accuracy degradation from 7B parameters up, and a compressed model is more accurate than a smaller uncompressed one with the same cache size. Together, these findings support pre-training large models and compressing them only afterwards.

Chain-of-thought traces are widely read as records of how models reach their answers, informing debugging, agent auditing, and claims about reasoning. Testing this interpretation is difficult because natural-language thinking traces are rarely mechanically verifiable. We revisit it in iGSM, a synthetic grade-school mathematics benchmark designed to study thinking traces and used to support claims of learned reasoning and planning. Crucially, iGSM exposes the exact quantities and dependencies that a correct solution should use, allowing generated traces to be checked programmatically step by step and enabling us to test whether correct answers are reliably accompanied by valid traces. We first evaluate models trained exclusively on valid, minimal traces. Answer correctness and trace validity nearly coincide in distribution but decouple out of distribution: on the hardest instances, 31.6% of correct answers have invalid traces, over half of which pass all syntactic and arithmetic checks but fail semantic dependency checks. We then intervene on trace supervision. Non-minimal training traces induce non-minimal outputs, while re-asking the same problem with a different query reveals computations inherited from the original query, weakening minimality as evidence of selective planning. Shuffling tokens in 10% of training trace sentences preserves near-clean accuracy even out of distribution despite no trace passing verification. Swapped training traces likewise retain high in-distribution accuracy. We discuss the implications of these findings for chain-of-thought monitoring and interpretation in the context of AI safety.

A small trainable advisor can steer a frozen language-model executor using natural-language advice. In addition to learning from task rewards, the advisor can use feedback from completed interactions to improve its advice. However, a plausible correction need not change execution, yet learning from such corrections can still affect the advisor's future decisions in other contexts. In a shared-parameter model, we prove that such corrections can limit learning if their targets favor useful advice less strongly than those of other corrections. Keeping them less often than the rest improves the model's eventual performance compared to learning from every correction. Motivated by this, our method, Advisor Self-Distillation (AdviSD), pairs outcome-based reinforcement learning with self-distillation from a feedback-conditioned copy of the advisor selectively. Reflection proposes corrections, and the advisor scores the same recorded executor response with and without its issued advice, using the magnitude of the difference to select decisions for supervision. This approach does not require executor likelihoods or additional executor rollouts. Experiments with Qwen3-8B advisors for Gemini and Claude show that AdviSD outperforms advisor-GRPO by 4.2-6.4 percentage points on BFCL-v3 and by 3.9-5.1 score points on EnvScaler. The trained advisors generalize to out-of-domain tasks and transfer across different executor versions and model families. AdviSD also beats matched-count random selection, supporting the value of its selection rule.

Automated skill extraction underpins workforce planning, yet most systems represent skills as flat labels with no notion of the responsibility level at which a skill is practiced. The Skills Framework for the Information Age (SFIA) captures exactly this dimension, defining 147 professional skills across seven responsibility levels, but no automated LLM-based extraction targeting SFIA has been reported. We formalize the task as structured prediction of (skill, level) pairs from free text and ask three questions: how accurately can text be mapped onto SFIA's closed vocabulary, which strategies reliably predict the level alongside the skill, and do agentic designs improve on simpler retrieval and prompting? We evaluate five strategies (a lexical baseline, dense retrieval with LLM reranking, a zero-shot schema-constrained LLM, single-agent agentic RAG, and a three-agent retriever--matcher--verifier crew) against expert-mapped European ICT role profiles, all drawing on an SFIA~9 corpus built by a fully automated agentic pipeline that we release. Retrieval-based matching identifies the most skills while generative strategies are markedly more precise; only strategies assigning the level as an explicit decision predict it reliably, with similarity-based selection more than twice as inaccurate; and the crew doubles latency without improving accuracy, so added agent roles do not automatically benefit closed-taxonomy matching. These results provide the first reproducible baseline for structured, level-aware skill extraction against SFIA.

LLM agents can make unsafe tool calls even when instructed to behave safely. Existing defenses constrain agents before execution, modify tool inputs/outputs, or rely on LLM judges; these approaches may depend on model behavior or block unsafe actions without helping the agent recover. We argue that the execution environment should instead enforce safety as the agent runs and steer it toward safe alternatives when violations occur---we call this Environment Steering. We implement this by modeling the agent and harness execution state as database tables, track the record-level data flows, and check these data flows against declarative policies during runtime. When violations are detected, policy- and context-specific feedback steers the agent toward safe trajectories. On AgentDyn, this enables the agent to improve task success rate over no-defense while achieving 0% attack success rate.

Is human readability necessary for effective fine-tuning of large language models? We investigate whether model-conditioned training representations can preserve or improve adaptation utility without requiring a human-readable textual form. We propose Desired-Update-Aligned Synthetic Data (DASA), which uses activation-gradient feedback from a frozen reference model to guide the optimization of continuous synthetic input embeddings. Inspired by the role of activation gradients in local risk reduction, DASA targets useful adaptation updates rather than source-text reconstruction or linguistic fluency. The resulting embeddings are used directly for downstream fine-tuning; discrete token projections are employed only for qualitative inspection. Experiments on six models from the Llama and Qwen families, ranging from 1B to 32B parameters, cover six benchmarks spanning knowledge, mathematical reasoning, code generation, and commonsense reasoning. Under matched LoRA adaptation settings, DASA achieves performance comparable to the source natural-language data and surpasses it in multiple configurations, while outperforming GRADMM in most comparisons. Further experiments cover general-domain and task-specialized source data. Under the evaluated synthesis settings, DASA provides a $3.6$--$4.9\times$ speedup over GRADMM with comparable peak GPU memory.

Automated generation of LLM harnesses promises to improve inference through task specialization. Yet additional answer coverage can arise from repeated execution of the same program, making specialization difficult to identify. We introduce a controlled evaluation that separates answer coverage, repeatable task advantages, and gains from pre-execution selection. On 386 MATH-500 tasks, we compare eight generated harnesses plus a baseline with nine byte-identical baseline copies, using three executions per member. Identical programs yield 2.16 percentage points of repeat-averaged oracle headroom. Generated programs exhibit substantially more repeatable score patterns, but these chiefly reveal persistent weaknesses: losses relative to the baseline persist across all three repeats on 100 tasks, while persistent wins occur on only one task and are sensitive to answer extraction. The frozen selector gains 0.00 percentage points, and both populations reach 98.70% oracle coverage at 27 harness executions. Stable complementarity remains unresolved at three repeats. Supporting BIRD traces locate failures in mechanism implementation, activation, and output validity. Together, these findings establish why coverage and repeatability alone cannot justify claims of useful specialization. They motivate an evaluation standard for harness diversity: task advantages should persist across executions, guide usable decisions, and improve on additional fixed-program executions under matched inference budgets.

Computer-use agents (CUAs), while capable of completing computer tasks in everyday and professional workflows, can cause unintended harm even under benign instructions and environments. However, detecting such harm remains challenging. First, it requires careful, task-specific reasoning: verifiers guided only by general safety criteria often overlook many important but subtle harmful behaviors. Second, it requires active investigation: past trajectory screenshots show what the agent did but not always what actually changed in the environment, so LLM-as-a-judge verifiers that rely on screenshots alone may be unable to determine the actual consequences of actions. To address these challenges, we introduce SCOUT, a two-stage agentic safety verifier that synergizes reasoning-intensive rubric generation with tool-intensive evidence gathering. First, our SCOUT rubric generator extensively reasons over the task and the agent's trajectory to determine what successful and safe execution should entail, generating task-specific completion and safety rubrics. Then, our SCOUT probing agent follows these rubrics to interact with the post-execution environment and collect grounded evidence for final safety and completion judgments. We evaluate our framework on two computer-use safety benchmarks. On AutoElicit-Bench, SCOUT achieves 75.4 unsafe F1 and 74.5 completion F1, outperforming LLM-as-a-judge verifiers and naive tool-use verifiers. SCOUT leads on OS-Blind with 76.4% unsafe detection accuracy. Test-time reflection reduces final unsafe execution rates from 30.2% to 17.2% on AutoElicit-Bench. Ablations and analysis show that tool-free rubric generation in SCOUT elicits substantially more reasoning and is crucial for safety detection across verifier backbones, especially non-frontier ones. A preliminary extension to coding tasks shows that SCOUT can support safety verification beyond computer-use.

Large language models (LLMs) are increasingly used by people whose first language is not English, yet these users have been shown to receive systematically lower-quality responses than fluent speakers. Which specific features of non-native English drive this gap remains unclear, because fluency is itself a composite of mechanical accuracy, vocabulary use, organization, and discourse coherence. Here, we introduce FABLE, a controlled dataset of 190,911 English prompt variants derived from 174K real user prompts for writing-related tasks. Evaluating responses from 34 open-weight LLMs, we find a clear asymmetry; while models do not propagate surface errors such as misspellings into their outputs, models do mirror higher-level rhetorical and lexical qualities present in the user's prompt. Further, the overall quality of responses differs substantially between the least- and most-fluent prompts. These results highlight a key LLM performance disparity for non-native English LLM users, resulting in both lower-quality and less-fluent answers.

An assistant that serves the same user over a long horizon has to answer from what that user has revealed: which preferences still hold, which were revised, and which constraints apply now. Retaining that information is not the same as acting on it, and the two are usually optimized as if they were. Keeping the information as text makes the reader's input grow with the retained history, while compressing it into a fixed number of latent vectors bounds the interface but is typically trained to reconstruct text or imitate reference answers, both of which are scored on sequences the reader never produced. We present MemFold, which optimizes a fixed-budget soft memory by the behavior it supports. A query-conditioned textual memory is compressed into K continuous vectors that form the reader's memory interface, and the reader is then trained on its own rollouts under two complementary signals: group-relative rewards for task outcomes, and confidence-gated on-policy distillation in which a frozen textual-memory teacher re-scores the student's sampled tokens under the textual memory. The teacher is never sampled from, so supervision stays on the student's current distribution and adds no autoregressive decoding; at inference it is removed entirely. Across three Qwen backbones, MemFold attains the highest accuracy we measure on PersonaMem-32K and PersonaMem-128K, with margins that widen at the longer history length, and transfers to PrefEval and LongMemEval without target-domain training. Ablations attribute most of the task gain to the reward term and a smaller additional gain to the teacher signal, and memory interventions show that the reader depends on the instance-specific content of its soft memory.

Masked diffusion language models (MDLMs) can generate text efficiently by predicting multiple masked tokens in parallel, but predictions from the same forward pass are not necessarily reliable when committed together. We study when parallel commitment is reliable. Our diagnostics show that confidence alone does not determine a reliable commitment order: confident predictions near the end of the sequence can fix an answer before its supporting computations are established, and downstream predictions become less reliable as the uncertainty of their upstream context grows. At the same time, a single forward pass can already resolve several masked tokens, and predictions that remain stable across the final layers are more likely to be correct. Based on these findings, we propose Reliable Parallel Decoding (RPD), a training-free method that selects candidates by layerwise prediction stability and final confidence, and commits them under a cumulative entropy budget over their preceding masked positions. RPD defers predictions with uncertain upstream context while committing the remaining candidates in parallel, without relying on a fixed block schedule. Across mathematical reasoning and code generation benchmarks on LLaDA and Dream, RPD achieves the highest decoding throughput among the evaluated methods while maintaining or improving accuracy.

On-policy distillation (OPD) trains a student model on its self-generated trajectories with dense token-level teacher feedback. However, naive OPD may suffer from teacher unanchoring, where the student's reasoning trajectory drifts far from the teacher, causing the teacher to be queried on states it would hardly visit and thus provide unreliable supervision. We propose Interactive-Policy Distillation (IPD), which applies adaptive teacher intervention to the student rollout. Under a bidirectional propose-and-verify state machine, the student and teacher alternately exchange their roles as proposer and verifier, and collaboratively generate mixed-source trajectories. Then different supervisions are applied according to the source of each token. This bidirectional propose-and-verify mechanism and the source-split loss make IPD not only a more performant distillation method, but also a unified bridge between on-policy and off-policy paradigms. To make the interleaved dual-model rollouts more efficient, we also design a dedicated fused inference engine that co-hosts both models in one serving instance with separate KV caches and instantiates the state machine model to distribute, collect, and process requests. On math reasoning tasks and across multiple teacher-student model pairs, student models trained with IPD not only outperform those trained with OPD, but also demonstrate higher data efficiency. Specifically, when distilling Qwen3-30B-A3B into Qwen3-1.7B-Base, IPD brings a +3.28 mean@8 and a +3.28 best@8 benchmark-averaged accuracy improvement compared with OPD. Besides, IPD only consumes about 1/4 of the training examples and steps to outperform OPD trained on the whole training dataset for one epoch. We also investigate the impact of different loss variants and takeover / handback configurations, and demonstrate the robustness of IPD on different training data.

On-policy distillation (OPD) reduces train-test state mismatch by training a student on its own generated trajectories, but weak students may visit teacher-misaligned prefixes where supervision is less representative. We introduce SAKI (Supervision Allocation with KL-constrained Interpolation), which combines a KL-constrained teacher-guided rollout with maximal coupling and reuses realized accept/correction events to route token-level supervision. Accepted positions retain sampled-token reverse-KL supervision, while correction positions receive direct supervision on the teacher's highest-probability token. Under maximal coupling, the correction probability is exactly TV(p_t, q_t), so the same trust-region radius controls rollout deviation and upper-bounds intervention and specialized-supervision frequency. We further implement an engine-resident speculative verifier that preserves the exact-q trajectory distribution and coupling semantics while improving matched-workload rollout throughput by 4.22x. Across seven mathematical reasoning benchmarks, SAKI improves the matched teacher-guided baseline in Mean@8 and Pass@8 for both 1.7B and 0.6B students. Placement controls and fixed-prefix analysis further support correction-triggered routing as a conflict-adaptive supervision signal.

Retrieval-augmented fact-checkers often receive a reliability label, such as HIGH or LOW trust, for each evidence source. These labels should adjust the model's confidence and its decision to search for more evidence, while the verdict should follow the evidence content. We introduce TrustSwap, a counterfactual test that swaps, lowers, or removes source labels while keeping every evidence text fixed, and measures its three output channels (the verdict, the confidence, and the search decision) separately. Across untrained and RL-trained models at two scales, three datasets, and two prompts, confidence and search respond to the labels as intended in 49 of 50 comparisons, yet a label change alone alters 4-23% of confident verdicts for Qwen3 models and up to 50% for an existing RL-trained fact-checker. Standard GRPO fine-tuning amplifies this shortcut at 8B in all six settings. To reduce it, we propose trust-swap augmentation (TSA), which trains GRPO on each claim with both its original and its label-swapped evidence under the same gold verdict. At 4B, TSA lowers the verdict flip rate by 7-35% (relative) in four of six settings, keeps accuracy and the intended confidence and search responses, outperforms reward-based alternatives in the main setting, and carries over to an unseen label-removal perturbation. An added consistency reward helps on the trained-on swap but not on unseen perturbations. At 8B, TSA's effect is not detectable, which makes scale the main open question.

Embodied agents now take on ever longer tasks. For long tasks, knowing only whether a task finally succeeds or fails says little; the steps along the way matter. Progress Reward Models (PRMs) score how far a task has come at every step, and serve as dense rewards, verifiers and monitors. Yet in long tasks the current frame alone often cannot tell how far the task has come, because progress depends on what happened before. We call this problem context-dependent progress estimation. Existing benchmarks on progress estimation mostly focus on short tasks whose progress can be read from the current observation, and whether PRMs can estimate progress when context is needed remains underexplored. We therefore build ContextProgress-Bench, with 24 manipulation tasks for 120 episodes. The benchmark covers three settings: (i) State Recall, where information needed for progress appeared earlier but is not in the current frame; (ii) Sequence Tracking, where steps follow a fixed order, so progress requires knowing which steps are done and which comes next; and (iii) Recurrence Disambiguation, where look-alike frames sit at very different progress. We then run a paired diagnosis: each PRM keeps the same input format in both runs, and in one run its instruction integrates the right context. Even PRMs that read the entire history get lost in estimating progress, yet with the right context the same five models cut their progress error by 77-82%. Embodied PRMs are thus not incapable of progress estimation, but lost without the right context. We therefore propose ProgressCompass, an autonomous agentic loop that reorients an existing PRM and uses current general-purpose VLMs to supply the context the PRM needs. Wrapped in the loop, the same frozen PRM cuts its progress error by 63% and raises its rank agreement by 76%. With such a compass, PRMs estimate progress far better on longer, more complex tasks.

Matrix optimizers have emerged as a promising direction, with Muon standing out as a prominent design. Revisiting Muon through its full-Gram representation, we observe that it jointly processes marginal-scale and interaction information. This opens an alternative way to organize geometric information hierarchically, motivating the Normalize-Then-Precondition framework. Specifically, it first uses diagonal-Gram information to construct a marginally normalized update, then applies spectral preconditioning to its directional interaction geometry. Building on this framework, we develop NormPre with NormPre-G and NormPre-L adopting global and localized spectral preconditioning, grounded in spectral-norm steepest descent and a regularized formulation followed by leading mode selection, respectively. To enable large-scale training, NormPre-G uses Newton-Schulz iterations and NormPre-L employs randomized sketching to approximate the leading interaction eigenspace. Theoretically, we establish $\mathcal{O}(T^{-1/2})$ convergence guarantees for simplified versions of NormPre. Across extensive pretraining experiments on GPT-2 Small, LLaMA and Qwen3, both variants consistently outperform AdamW, Muon and MANO under matched training budgets. Further efficiency and spectral analyses reveal the complementary strengths of two variants and characterize their performance-efficiency trade-off. We open-source our code through a GitHub repository at https://github.com/zx-gong/NormPre.

A key strength of Proximal Policy Optimization (PPO) is its learned critic, which uses historical trajectories collected during reinforcement learning to estimate expected returns and reduce policy-gradient variance. However, we find that the critic is also a major source of instability in reinforcement learning for large language models (LLMs). We identify two critic failure modes that destabilize PPO. First, filtering truncated rollouts from both actor and critic shifts the policy objective to reward conditioned on completion, allowing truncation to increase even as conditional reward improves. Second, heterogeneous return noise can cause high-variance prompts to dominate critic updates in finite batches. We introduce EasyPPO to address these failures. Actor-only overlong filtering trains the critic on returns from both completed and truncated rollouts. Noise-normalized critic regression weights each prompt's critic loss by the inverse standard deviation of its sampled returns, balancing noise contributions across prompts. Moderately smaller critic mini-batches confine outlier influence to fewer rollouts during gradient clipping. Across continuous-reward coding on FrontierCS, binary-reward mathematical reasoning on AIME24, and multi-turn search on Search-R1, EasyPPO remains stable throughout the full training horizon and consistently outperforms vanilla PPO, VAPO, and HL-Gauss PPO. Its best validation scores show relative gains of 14.89%, 2.28%, and 9.47% over PPO, respectively.

Reinforcement learning (RL) has become a cornerstone for improving the reasoning capabilities of large language models (LLMs), but the need for on-policy data substantially limits training efficiency. Reusing off-policy data through importance sampling (IS) can improve efficiency but introduce considerable instability. Hence, algorithms such as PPO and GRPO widely adopt IS-ratio clipping to stabilize training. However, training stability and gradient estimate are mainly determined by the product of IS ratio and advantage. To further stabilize training, we propose ACPO, which clips the product of the IS ratio and the advantage, leading to more stable gradient estimates. We also establish a connection between ACPO and gradient clipping in policy mirror descent (PMD), which is a standard technique to stabilize optimization process, and prove the convergence of clipped-PMD under the standard RL setting. Experiments on widely used mathematical reasoning benchmarks show that ACPO consistently outperforms PPO and GRPO in both accuracy and training efficiency, delivering 4-6 percentage points gains on standard math benchmarks, with Qwen3-8B+PPO. Hence, ACPO is a practical and effective alternative to conventional IS-ratio clipping for RL post-training of LLMs.

Reward hacking occurs when policy optimization exploits a brittle reward interface or an overly permissive proxy objective, improving the training score without improving the underlying response quality. This phenomenon is amplified in group-relative policy optimization: an unsupported reward can shift the group baseline and alter the updates of other rollouts, while post-hoc or purely relative weighting cannot represent group-wide uncertainty. We propose \emph{Self-Tuned Anchored Reliability Group-Relative Policy Optimization} (STAR-GRPO), a reliability-first advantage estimator based on paired assessments of the same rollout. STAR separates the quality signal from its learning influence: score disagreement determines rollout reliability, relative reliability enters a self-tuned robust location--scale fit before group normalization, and absolute group reliability attenuates the resulting bounded advantage. The analysis establishes coordinate and second-moment bounds, characterizes exact centering through the weighted location equation, and gives reliability-dependent attenuation guarantees for outlying rewards. We evaluate STAR-GRPO in two complementary reward-hacking regimes. In token-interface exploitation, STAR prevents runaway optimization of the deployed-interface score while improving the canonical quality signal. In rubric-proxy overoptimization for medical reasoning, STAR improves independent semantic evaluation, narrows the proxy--judge discrepancy, and reduces overclaim while optimizing the same task proxy. Together, these results show that reliability-first normalization offers a principled way to limit unsupported reward influence on both group baselines and policy updates, while retaining the task reward as the optimization target.

Many computer tasks recur: the same workflow runs many times, with new inputs and from different starting states. Current computer-use agents re-plan every step of every run, which makes them costly and unreliable on such tasks. We introduce neuro-symbolic computer use, in which a recurring workflow is executed by a learned policy rather than re-derived by an agent on each run. The policy fixes the decisions that are stable across runs (ordering, variables, loops, and branches) in executable code, and delegates observation-dependent decisions, such as grounding and state checks, to neural models. We learn these policies with neuro-symbolic policy iteration: starting from one agent trajectory, it executes the policy, diagnoses failures with task-completion and step-level judges, and revises the code with a coding model informed by an agent's continuation from the point of failure, without access to the benchmark evaluator. Iterating on generated parameter and initial-state variants makes the policy reusable, and a pre-action verifier guards each state-mutating step at deployment. On OSWorld-Verified and ScienceBoard, the learned policies achieve the highest Pass^3 of all methods in all four settings, 3.6-15.8 points above the base agent, while cutting per-run cost by 15-217$\times$ and latency by 3.4-5.1$\times$. On OSWorld-Verified, policies built only on variants transfer to the held-out original tasks, exceeding AutoRPA by 8.6-17.5 points in Pass^3.

Low-Rank Adaptation (LoRA) is a widely used approach to parameter-efficient fine-tuning (PEFT), yet a performance gap can remain relative to full fine-tuning (FFT). Many LoRA variants improve the initialization or optimization of low-rank factors. At each training step, however, their first-order weight-space directions are constrained by the current parameterization. We characterize the corresponding LoRA-accessible gradient space and show that it coincides with the tangent space induced by the current LoRA parameterization. This characterization yields an orthogonal decomposition of the full weight gradient at the current model parameters. We term the component orthogonal to this space the normal gradient. Based on this decomposition, we propose GDLoRA (Gradient-Decomposed Low-Rank Adaptation). GDLoRA reconstructs the full weight gradient from forward activations and backward signals, extracts its normal component, and directly updates the base weights with this component, while retaining standard AdamW optimization for the LoRA factors. GDLoRA incorporates complementary normal gradients without increasing standard LoRA's optimizer-state memory budget under matched adapter and optimizer configurations. Experiments on natural language understanding, mathematical reasoning, commonsense reasoning, and image classification show that GDLoRA consistently improves over LoRA and narrows the performance gap to FFT. The code is available at https://anonymous.4open.science/r/GDLoRA.

Large reasoning models improve performance on challenging problems by allocating additional computation before answering, but longer reasoning does not always lead to better results and can introduce substantial redundant reasoning on simple problems. Conversely, aggressively shortening reasoning can degrade performance on difficult ones. Effective reasoning therefore requires dynamically deciding when additional computation is useful based on the reasoner's capabilities and evolving solution state. Existing approaches often rely on predefined budgets or intervention rules, retrain the target reasoner, or require additional supervision. We introduce MetaCtrl, a lightweight controller that adaptively regulates a frozen reasoner without predefined token budgets or reasoner retraining. We formulate reasoning regulation as a sequential metacognitive control problem: MetaCtrl observes the evolving reasoning trace and decides whether to continue, simplify, skip redundant steps, or conclude reasoning. It is trained directly with reinforcement learning using a reward that prioritizes correctness while favoring shorter trajectories among correct solutions, requiring neither supervised intervention trajectories nor problem-specific budgets. Across seven benchmarks spanning mathematics, science, and code, MetaCtrl consistently improves the accuracy of LRMs while reducing their reasoning length. On DeepSeek-R1-Distill-Qwen-7B, it improves average accuracy by 4.7 points while reducing generation length by 53.3%. Without further training, the same controller transfers to an unseen reasoner (e.g., Qwen3-14B), improving average accuracy by 2.9 points and reducing generation length by 50.3%. These results establish MetaCtrl as a plug-and-play controller for improving reasoning accuracy while substantially reducing inference-time generation. The code is available at https://github.com/binbin2xs/MetaCtrl.

Recent Recommendation Agents (RecAgents) offer a promising alternative by shifting recommendation to an active, user-side paradigm, where generative agents autonomously perceive external platforms, reason over user preferences, and execute decisions. However, existing RecAgents still suffer from two critical limitations: brittle item perception based on noisy and heterogeneous item pages, and inefficient long-context reasoning over extended user histories and multi-step interaction traces. To address these challenges, we propose a novel recommendation agent framework, termed as ReMem, that combines OCR-based multimodal perception with time-evolving dynamic memory. Instead of parsing raw HTML, ReMem observes item pages through screenshots and extracts structured multimodal information via an OCR tool, enabling a more humanoid and platform-agnostic perception mechanism. To support long-horizon preference modeling, ReMem further introduces a chunk-wise sequential memory update strategy, where the agent selectively maintains a fixed-size memory of informative historical interactions while processing arbitrarily long contexts with linear inference complexity and bounded context length. This design allows the agent to preserve evolving user preferences without relying on external memory modules or disrupting the standard autoregressive generation process. To enhance the dynamic memory instruction, we further develop a multi-memory GRPO variant, which propagates the final-answer advantage to all intermediate conversations that contribute to the final response. Extensive experiments on three datasets demonstrate that ReMem consistently outperforms state-of-the-art baselines, achieving an average improvement of 5.16\% across three recommendation agent tasks, namely searching, ranking, and judging.

On-policy distillation (OPD) trains compact language agents with teacher feedback on student-generated trajectories. In multi-turn tasks, compounding errors can move students beyond the teacher's effective supervision. We introduce Graph-Conditioned On-Policy Agent Distillation (GC-OPD), which enriches an off-the-shelf teacher's scoring context with execution evidence. A graph indexes repeated teacher executions by shared states while preserving complete successful and failed histories. After each student episode, GC-OPD retrieves current-state references or historical alternatives and combines them with student hindsight to score the original thought-action tokens. Using the same original teachers, GC-OPD improves mean success over vanilla OPD from 24.70% to 48.78% on ScienceWorld (4B student), from 53.36% to 85.26% on ALFWorld Unseen, and from 29.10% to 37.65% on WebShop. At matched student sizes, it also achieves higher mean success than every evaluated OPD baseline using GRPO-trained teachers on ScienceWorld and ALFWorld; the strongest such ScienceWorld 4B baseline reaches 46.66%. GC-OPD requires no task-specific teacher optimization.

The common paradigm of reinforcement learning with verifiable rewards (RLVR) is to let agents make multiple attempts at a task, and optimize towards the successful ones. This becomes problematic in the realms of self-improvement, where tasks are so difficult that the agent has a low or even no chance of success, and where there are no teacher models or example solutions to distill from. In this paper, we introduce RLTL;DR. After each failed attempt, we show the policy the verifier outputs and let it write its own feedback, in the form of a single TL;DR insight. The next rollout is conditioned on all previous insights, and we sequentially sample rollouts until a solution is found. Moreover, we enable backpropagation on the in-context insights to internalize a direct task to insight mapping. On challenging tool-calling and coding datasets (filtered to Pass@128=0), standard GRPO training of a Qwen 3.5 9B Thinking policy stays flat at a Pass@1 of 0% to 1%. RLTL;DR breaks through this learning barrier, achieving a Pass@1 of 14-31% with insights in context during training and, crucially, 12-13% when no insight is in context at eval time. We identify that the key is the task to insight internalization. To study this further, we reduce our approach to SFTL;DR, training only on (task, insight) tuples, without showing or backpropagating on any rollouts. Training on only 4k of these tuples recovers almost the full performance of RLTL;DR and classical SFT on full rollouts. This demonstrates a promising compacted training paradigm of the form "on this sort of task, keep this sort of thing in mind", which we hope to inspire future research on.

Vision-Language Models (VLMs) remain vulnerable to cross-modal implicit risks: visual and textual inputs that appear benign in isolation can jointly elicit unsafe responses. Existing safety methods often require large preference datasets, costly multi-rollout training, or additional safeguards at inference time. They may also sacrifice helpfulness by directly refusing requests that could be answered safely. In this paper, we propose Intent-Privilege On-Policy Self-Distillation (OPSD), which leverages evidence-grounded intent as privileged supervision during training to help VLMs recognize implicit risks and provide safe, useful responses instead of blanket refusals. OPSD distills a teacher's intent-conditioned preferences over responses into a student using a single rollout per prompt; the student then responds without intent annotations or an additional safety module. With only 1,447 safety-specific examples - 95% fewer than standard preference datasets - OPSD reduces training time by 5x relative to multi-rollout GRPO-style training and average inference length by 7%. It attains the highest ratio for joint safety-helpfulness success, which measures the proportion of responses that are both safe and helpful, across all five evaluation groups. Remarkably, on pooled SIUO+HoliSafe, this success ratio rises from 43.9% to 53.5%. These results show that training-time intent supervision can improve both safety and helpfulness while substantially reducing data, training, and inference costs.

Zero-order optimization (ZO) trains without backpropagation, making it relevant to forward-only hardware and non-differentiable loss, but its gradient variance grows with perturbed dimension, inhibiting large-model training. Sharded Optimization Mixture of Assemblies (SOMA) trains LSTM experts independently on $N$ data clusters using simultaneous perturbation stochastic approximation (SPSA), without exchanging gradients, activations or optimizer state. Its separable loss removes cross-expert perturbation noise at the cost of jointly learned representations across domains. Using 80,000 estimated RTX 5090 GPU-hours, we show modest sharding improves training compute efficiency over all tested monolithic ZO controls. At 8.44M parameters and 150 aggregate GPU-hours, SOMA $N=2$ with 64 perturbations reaches 1.76 test nats/byte, versus 2.00--2.11 for monolithic SPSA at 64, 256 or 1,024 perturbations and 2.21 for EGGROLL. On WikiText-103, these frozen checkpoints reach 2.07, 2.25--2.36 and 2.49, respectively. On a fixed separable objective with equal-size blocks, we prove independent losses reduce relative gradient variance to approximately $1/N$ of a shared-loss estimator's. Holding starting weights, data, perturbations and compute fixed, independent rather than summed losses lower SOMA $N=4$ test loss by 0.035 nats/byte after 1,000 updates across three seeds. Larger ensembles offer a separate inference benefit: at similar model size with top-$k$ routing ($k=4$), SOMA $N=256$ achieves 2.36M tokens/s versus 257k for SOMA $N=8$ ($9.19\times$, including routing), at lower test loss (1.68 versus 1.71), albeit using $59.9\times$ as much aggregate training compute. We release all training and evaluation code and checkpoints.

On-policy self-distillation (OPSD) trains a student to match a privileged teacher distribution along its own sampled trajectory. Standard OPSD applies this supervision to unverified student rollouts while conditioning the teacher on privileged context, typically a reference solution. We separate these roles in a factorial analysis and find that scaffold correctness has a stronger effect on downstream accuracy than context correctness. Unverified scaffolds create an imitation gap because the teacher can use information unavailable to the student. This gap shrinks with model scale, yet OPSD continues to supervise mostly unverified trajectories. In contrast, verified scaffolds remain effective even when the teacher is conditioned on the student's own unsuccessful rollout. Based on this finding, we introduce OASIS, which retains the OPSD objective but supervises mostly verified by label on-policy trajectories and replaces written solutions with unverified model-generated attempts as the teacher context. OASIS therefore requires only final-answer labels. Across Qwen3-1.7B, 4B, and 8B on AIME 2024, AIME 2025, and HMMT 2025, OASIS improves over the base model by 3.2--3.8 points on average, while OPSD's gain falls from 3.05 points at 1.7B to 0.14 at 8B. At 8B, OASIS improves over OPSD by 3.05 points, showing that verified on-policy scaffolds preserve the effectiveness of self-distillation as models scale.

In recent years, the performance of large language models (LLMs) on reasoning tasks has been remarkable, even surpassing human capabilities on various benchmarks. However, there remains a lack of clear understanding in the academic community regarding how the structure and internal parameters of LLMs progressively solve complex reasoning problems. In this study, we investigate the inference process of LLMs on cross-linguistic materials and propose the hypothesis that LLM layers exhibit a structured division of labor across conceptualization, reasoning, and textualization. Based on this hypothesis, we introduce a bottleneck identification mechanism using sensitivity analysis to pinpoint the most critical functional stage for a specific task. Leveraging this insight, we propose a novel approach, Layer-Informed Fine-Tuning (LIFT), which achieves efficient and effective fine-tuning by selectively updating only these functionally critical layers. We then conduct extensive experiments to show that the LIFT method not only accelerates the training process but also significantly improves model performance.

Just as Socrates recognized the limits of his own knowledge, Retrieval-Augmented Language Models (RALMs) should learn to abstain when the retrieved evidence cannot support a reliable response. Existing approaches largely rely on monolithic LLMs to handle heterogeneous retrieval failures in a single step, resulting in limited abstention performance and high computational costs. We instead decompose retrieval failures into two distinct states: (i) the unanswerable state, where the required evidence is absent, and (ii) the distracted state, where relevant evidence is mixed with conflicting, negated, or adversarial information. Based on this decomposition, we introduce a lightweight module (Sieve) that screens retrieved document sets for distracting evidence before invoking a costly LLM (Sage) for grounded generation and abstention. Evaluated across both general and high-stakes expert domains, our Sieve and Sage framework preemptively detects distracting noise, improving system accuracy by up to 69.4 percentage points and Macro-F1 by 55.2 percentage points compared to one-stage baselines. Furthermore, it achieves up to a 1.99x speedup, establishing a highly efficient and reliable abstention pipeline for RALM with abstention.

Single-token typed-decision models answer a schema question by reading the logits of a few one-letter answer codes at a single position: they are fast and return a probability for every allowed answer, but they are trained with plain cross-entropy that ignores most of the structure in their training data. We study such a model whose data is curated as contrastive pairs---two contexts that differ in one edited fact that flips the answer---each carrying a machine-checked certificate that deleting the decisive sentence makes the fact unknown. We propose PACT, which turns this structure into four training terms that need no new annotation: a difference-in-differences margin over each pair that is invariant to any shared logit offset, a permutation-consistency term against answer-code position bias, an evidence-necessity term on certificate-verified ablated contexts, and an ordinal transport cost for rubric fields, plus a three-parameter contextual temperature. On a frozen 324-item holdout with three seeds, PACT matches the published recipe in accuracy ($84.6\%$ vs. $85.2\%$; McNemar $p \ge 0.50$ at every seed) while giving the lowest position bias of all runs (answer flips under relabelling $9.8\%$ vs. $13.8\%$) and the lowest ordinal error on rubric fields (MAE $0.232$ vs. $0.311$). Against a control with the same optimiser and schedule but cross-entropy only, PACT is significantly more accurate at two of three seeds, halves the seed-to-seed spread and lowers NLL by $26\%$. Seed-matched ablations and pre-specified falsification tests locate these gains precisely: no single term raises raw accuracy, and the method's value lies in robustness and stability rather than headline accuracy. Code, data splits, trained adapters, and all run records are available at https://github.com/BennyLinntu/PACT-Pairwise-Anchored-Calibrated-Tuning-for-Single-Token-Typed-Decisions.

Adapting large language model agents during deployment requires not only retaining past experience, but also turning new observations into timely guidance. Many test-time learning methods, however, acquire knowledge from completed episodes. Feedback from an ongoing interaction may therefore not be distilled into knowledge soon enough to help the next decision. Acquiring knowledge at the granularity of individual transitions could reduce this delay, but raises a separate challenge: a rule that is useful within one episode may not be reliable enough to guide future episodes. Waiting for validation can forfeit immediate benefits, whereas unrestricted reuse can propagate accidental or misattributed guidance. We introduce StepLearn, a nonparametric framework that separates immediate use from persistent trust. It turns informative transitions into hypotheses that can guide the next step, while requiring prospective validation before reuse across episodes. Their predicted effects are checked against subsequent observations outside the source episodes, and only sufficiently supported hypotheses become available for persistent guidance. This process updates external knowledge while keeping all model parameters fixed. Over five rounds on WebArena-Lite and ALFWorld, StepLearn achieves average success rates of 59.9% and 84.0% with GPT-5-mini, and 57.8% and 88.1% with Qwen3.5-35B-A3B, respectively. It outperforms EvoTest, the strongest baseline, by 2.2-12.7 percentage points across the four settings. Learning dynamics further shows that these gains are not restricted to the final repetition, with advantages already present on first task attempts in most settings.

Large language model post-training generates self-generated rollouts through reinforcement learning and on-policy distillation, yet this experience is often treated as stale once the policy advances. Historical rollouts can remain compatible with a later policy while preserving behaviors that the policy no longer expresses reliably. However, they may also contain mistakes, abandoned attempts, and redundant actions that should not be imitated, motivating finer-grained selective supervision. We introduce ROSS (Relearning from Self-Generated Rollouts through Selective Supervision), which preserves the full historical trajectory as context while applying loss only to selected model-generated continuations. Across domain-specific reinforcement learning, multi-teacher on-policy distillation, and agentic reinforcement learning, ROSS consistently improves upstream checkpoints and outperforms baselines across mathematics, code generation, instruction following, and software engineering. On Qwen3.6-35B-A3B, ROSS improves the six-benchmark MOPD average from 58.40% to 62.20% and SWE-bench Verified from 64.20% to 68.40%. These results show that self-rollout training leaves behind reusable behavioral experience that can yield further gains through offline supervised fine-tuning (SFT), without additional policy rollouts.

Long-running LLM agents compress past interactions into persistent memories that may be reused as premises for later tasks. This creates a distinct derivation problem: whether the memory actually follows from what the interaction history supports. Relevant evidence may be scattered across earlier interactions, while compression can introduce relations or event status that the history never established. A valid memory may therefore appear unsupported because its citations omit relevant evidence, while individually supported facts may be composed into a stronger statement the history never established. We characterize this problem through three coupled requirements: (1) Evidence scope; (2) Compositional validity; (3) Admission reliability. We therefore ask whether the interaction history available at write time supports what enters persistent memory. We introduce DerivAudit, a framework for auditing whether a memory is actually supported by the history available when it was written. The audit separates three questions: whether supporting evidence lies beyond writer-provided citations, whether the composed memory introduces unsupported meaning, and how write-time admission decisions affect later memory use. Across two natural memory corpora, audits using broader pre-write history recover support for nearly 60% of memories that appear unsupported from citations alone, while 17-21% remain unsupported after expansion. Yet broader evidence does not by itself make admission reliable: unsupported memories are still frequently admitted across verification models, and evidence expansion alone worsens it on two backbones.

Group Relative Policy Optimization (GRPO) has become a promising approach for training large language model agents. However, its uniform assignment of trajectory-level advantages to all policy tokens fails to distinguish consequential decisions from less relevant ones, obscuring which intermediate decisions contributed to success. We introduce ProVer, a framework that targets potentially pivotal decisions for fine-grained credit assignment in agentic reinforcement learning. Given a rollout group, an agentic judge contrasts successful and failed trajectories to propose a segment potentially responsible for their divergent outcomes. Rather than directly trusting the judge's assessment, ProVer verifies the proposed segment by estimating its advantage from the difference in terminal success rates between current-policy continuations sampled before and after the segment. Positive estimates are then incorporated into the GRPO advantages of policy tokens within the proposed segment. By using model judgment only to select where to verify, ProVer grounds local credit in observed outcomes without exhaustively evaluating every intermediate state. Across ALFWorld, WebShop, and SearchQA, ProVer achieves the strongest average performance at both model scales, with relative improvements over GRPO of 9.91% and 7.12% for Qwen3.5-2B and Qwen3.5-4B, respectively. Further analyses demonstrate that informed segment selection improves policy training with modest additional generation overhead, even without a frontier-scale judge model, highlighting the effectiveness and efficiency of selectively targeting pivotal decisions for fine-grained credit assignment in agentic reinforcement learning.

The strong generalization performance of on-policy post-training paradigms has motivated studies of their parameter update behaviors. However, these studies treat the observed behaviors only as byproducts in on-policy training, overlooking their potential to serve as optimization principles for improving the generalization of other paradigms such as supervised fine-tuning (SFT). To address this limitation, we investigate whether there exists a specific on-policy update behavior that can achieve such improvements. First, our analyses reveal that SFT updates parameters along consistent directions, while the on-policy paradigm continuously adjusts the direction during training. This difference inspires us to focus on the cumulative update direction of each parameter as a promising behavior. Then, we evaluate its effectiveness for improving generalization by proposing On-Policy direction-constrained Supervised Fine-Tuning (OPSFT), which constrains SFT updates to the direction identified by on-policy paradigms. The strong performance of OPSFT indicates that the generalization advantage of on-policy paradigms can be transferred to SFT through the parameter update direction. Once such a direction is identified, even SFT can generalize with its updates constrained to this direction. This finding offers two practical benefits by combining the strong generalization of on-policy paradigms with the advantages of SFT, including the high training efficiency and ability to leverage high-quality trajectories. For efficiency, we identify update directions that support strong generalization using a few on-policy training steps, and subsequently apply OPSFT to achieve high training efficiency. For leveraging high-quality trajectories, OPSFT can utilize these trajectories to continue improving a post-trained model along its update direction without disrupting the ability learned from on-policy training.

Compiler backends are expensive to build and maintain as programming models, workloads, and accelerators evolve. We investigate whether large language models can replace the conventional optimizing and lowering pipeline, a process that we call AI lowering. We study AI lowering from Triton to NVIDIA PTX: an LLM agent translates Triton kernels directly into PTX. We build an environment that evaluates candidate PTX, and an agentic harness in which an LLM translates Triton kernels into PTX. Across twelve common kernels on Ada, Hopper, and Blackwell GPUs and ten kernels from recent ML papers, AI lowering achieves 0.83x-3.34x the performance of autotuned Triton. The largest gains come from transformations that Triton's lowering pipeline does not perform, such as decoding packed binary weights directly into Tensor Core operands (3.34x on BitDelta), assigning each thread a complete softmax row in tensor memory (1.37x on FlashAttention), and reusing overlapping convolution windows (up to 2.23x). These results rely on a robust evaluation harness with comprehensive verification support. We build on Volta, an existing PTX verifier, and substantially extend it to support modern GPU architectures by introducing support for Blackwell's tcgen05 Tensor Core interface. This requires modeling three architectural features: managed tensor memory, descriptor-based operand layouts, and asynchronous execution coordinated through commits, waits, memory barriers, and proxy fences. We discuss the challenges involved in formalizing them, as well as the current limitations. Our results suggest an emerging future in which AI compilers replace custom-written intermediate representations and checkers, reducing the time and engineering effort required to bring up software for new general-purpose and custom chips.

Group-relative policy optimization relies on reward-derived advantages and sequence-level likelihood weights, both of which can be sensitive to localized outliers. Extreme rewards can collapse the contrast among clean responses after group normalization, while token-level log-ratio perturbations can alter sequence weights and clipping decisions. We introduce RoVR-GSPO, a dual-channel robust optimizer that addresses these failure modes separately. Its reward channel combines robust reference estimation with bounded residual credit, while its ratio channel uses differentiable SoftRoVR aggregation to construct robust sequence weights. We provide stability and efficiency analyses for both channels. Experiments on mathematical reasoning, long-context summarization, and tool-call annotation show consistent improvements over GSPO, while controlled perturbation studies demonstrate stronger robustness to reward contamination and token-ratio anomalies.

Long-horizon LLM agents require reinforcement learning methods that can assign credit to intermediate decisions under sparse and delayed rewards. Existing group-based methods such as GRPO and GiGPO alleviate this issue by comparing rollout returns or repeated anchor states, but they still fail when the compared returns have no variation. We identify this failure mode as zero-credit failure: during early training, many failed rollouts contain useful prefixes, yet existing methods assign them no task-discriminative advantage. To address this issue, we propose Milestone Viability Potential Policy Optimization (MVPO), a potential-routed policy optimization algorithm that learns from viable failure prefixes. MVPO estimates prefix potential over Union-Find viability regions, repairs zero-credit groups with potential-difference advantages, and attenuates the potential branch according to relative performance progress. Experiments with Qwen2.5-1.5B-Instruct show that MVPO outperforms eight strong baselines, including GRPO and GiGPO. Under the same training length, MVPO improves over the GiGPO baseline by +4.4 success points on ALFWorld and +5.3 on WebShop, while adding only 0.16%-0.20% advantage-construction overhead.

Prospective memory allows an agent to retain an intention tied to a future condition, but the stored intention does not reveal whether that condition currently holds. Checking it may require web access, multi-step tool use, and paid calls. Existing systems decide when intentions require attention, but do not allocate the resulting observations under a shared budget. We introduce the first resource-allocation formulation for the external observations required by stored intentions under a shared episode budget. BudgetPM offers two policy variants that share a hard-budget executor. BudgetPM-Static uses a lightweight Logistic scorer to learn whether a check improves the current decision. BudgetPM-Sequential distills full-episode hindsight schedules into a lightweight policy that decides when to spend or reserve capacity using only pre-query information at deployment. We evaluate BudgetPM against two public memory-agent systems, five matched controls, and four hand-designed monitoring or budget-adaptation rules. Across two benchmarks and three backbones, BudgetPM-Static outperforms adapted Mem0 and PMA workflows. On PM-Bench, its Logistic scorer reaches competitive quality--cost operating points alongside higher-capacity scorers and retains 99.9--100\% of unconstrained quality with 42--54\% fewer observations. Under severe scarcity and the same hard caps, BudgetPM-Sequential exceeds the strongest tested natural monitoring schedule by 1.92--2.58 Set F1 points. It reaches the same Set F1 and on-time recall with 16--33\% fewer observations. Matched attribution, exact-cost analysis, and a fixed-budget load intervention link this gain to competition between present and future opportunities. These results yield a demand--capacity design rule: local gating works when capacity covers demand, while future-aware supervision adds value when observations compete across time.

Learner simulation aims to reproduce how a particular learner behaves on new tasks. Although Large Language Models (LLMs) can generate increasingly fine-grained learning behaviors, existing approaches often need to repeatedly process a growing interaction history to reconstruct the learner. This introduces additional context and inference costs and makes the acquired learner-specific simulation capability difficult to reuse across different LLMs. We therefore propose Learner2Skill, which externalizes the simulation capability acquired from historical interactions into a persistent and reusable Simulation Skill. The Skill captures the learner's current learning state and recurring response patterns, evolves as new real interactions arrive, and can be adapted to a new LLM through lightweight executor calibration without reconstructing the learner from scratch. Experiments show that Learner2Skill more faithfully reproduces fine-grained learner behavior while reducing overall token cost, and that the same constructed Skills can be effectively reused across different LLM executors.

Large language models (LLMs) are increasingly deployed for natural-language logical reasoning, where the final answer is easy to check but the proof behind it is not. In natural-language logical reasoning, an intermediate conclusion should follow from its premises, and the resulting derivation should support the final answer. Existing methods lack machine-checkable verification of intermediate conclusions and answer-supporting proof dependencies, so they may assign credit to invalid or answer-irrelevant steps. We propose Proof-R1, an RL framework from formal verification that trains LLMs to construct verifiable proofs for natural-language logical reasoning. Proof-R1 admits a generated conclusion into the verified proof state only when the corresponding reasoning action satisfies the proof obligations through UNSAT-based machine-checkable formal verification. Proof-R1 also recovers the answer-supporting dependency closure to trace the proof structure of the final answer and align outcome credit with the proof dependencies. Experiments demonstrate that Proof-R1 improves answer accuracy across three logical reasoning benchmarks and four backbone models and outperforms training-free agents and training-based methods in terms of reasoning-process verifiability.

User-written Triton kernels enable high-performance GPU computation within PyTorch, but their end-to-end latency can remain dominated by host-side orchestration, especially when device execution is short. Although torch.compile can generate native host wrappers for captured graphs, each invocation still passes through runtime-managed specialization lookup, guard evaluation, and preparation before reaching the wrapper. We present Trident, a compiler backend that removes this recurring overhead from the specialization cache-hit path. Trident introduces the Specialization Cache Module (SCM), which compiles guarded specialization selection, argument and execution-environment preparation, and host execution for multiple specializations into a single executable module. An invocation enters the SCM once, remains in compiled code when a specialization matches, and returns to Python only when a new specialization must be compiled. Built on Torch-MLIR, Trident lowers guards and host-side orchestration to native code while retaining calls to optimized runtime implementations of supported ATen operators. Our evalu- ation on two LLMs shows that Trident achieves up to a 1.47x speedup in model-level end-to-end latency over eager execution and up to 1.68x over torch.compile.

Generating optimization instances that are both feasible and computationally challenging is crucial for benchmarking solvers and training learning-based optimization algorithms. Existing non-LLM generators rely on seed instances or parameter tuning, resulting in high test-time computational cost, while existing LLM generators lack explicit hardness measures. Recent reinforcement learning methods with verifier feedback evaluate only binary correctness, which is misaligned with generating challenging problems. We note that an optimization solver reports the cost of solving at several stages of its pipeline, and leverage this to design a reward that scores both the solvability and the hardness of generated problems, measured by branch-and-bound nodes and post-cut relaxation gaps. Our key idea is a challenger-solver asymmetric self-play approach, where an LLM challenger generates progressively harder instances and the solver verifies feasibility and hardness, so no seed or training MILP instances are required. We fine-tune Gemma-4-12B and Qwen3.5-4B with GRPO and a size curriculum into OptiScribe-12B and OptiScribe-4B, which generate feasible yet challenging MILP problems from natural language instructions. On capacitated facility location and max-cut, OptiScribe-12B raises median SCIP search nodes by 1.7-5x and post-cut gaps by 1.1-1.7x over its base model and improves the feasibility rate on facility location by 9-19 points, while OptiScribe-4B raises median nodes by up to 15.6x. The problems cover a wider difficulty range than public benchmarks of the same size, follow instructions on density and difficulty, and can tune solver settings for families that public libraries lack. These results indicate that optimization-specific rewards, used in self-play mode, can teach LLMs to generate high-difficulty optimization benchmarks. We will release our code and models publicly on acceptance.

Large language model (LLM) routing reduces serving cost by assigning each query to an appropriate model while preserving response quality. Learning such a router, however, often requires executing multiple candidate models on historical queries to collect query--model quality feedback, creating a nontrivial supervision cost before deployment. Existing work largely focuses on serving-time efficiency, overlooking whether the resulting savings are sufficient to recover this upfront expenditure. We further observe that routing quality often saturates well before all query--model feedback is collected, suggesting that dense supervision can be economically over-provisioned. We propose SaveRouter, a sparse-supervision routing framework that selectively acquires informative model feedback and shares capability information across related queries, while retaining query-level refinement for fine-grained routing. We evaluate routing by jointly accounting for supervision expenditure and subsequent serving-time savings. Across four routing benchmarks, the main setting uses only about 33--41% of available training feedback while maintaining competitive or better routing quality, and reduces the break-even deployment volume by approximately 1.9--9.5 times compared with the fastest conventional router. Further analysis shows that acquiring more supervision is not always economically preferable: the supervision level that minimizes serving cost can differ from the one that achieves the earliest payback. Our code is publicly available at https://github.com/LAMDA-Model-Reuse/SaveRouter.

LLM agents accumulate interaction histories that grow linearly with task length, causing quadratic inference cost scaling and performance degradation from attention dilution. Existing context-compression methods learn what to discard offline: by contrastively optimizing guidelines, distilling compressors, or training compression policies. This incurs a substantial cost. Further, the compression policy is learned a priori and is not dynamically conditioned on the evolving test-time trajectories. In this paper we ask a complementary question: Which past interactions causally shape the agent's future decisions? We recast context compression as a causal decision preservation problem over discrete interaction units and introduce FOCUS, a training-free context compression framework that operates entirely at test time. Our method requires no offline data collection or fine-tuning, and is architecture-agnostic, attaching to any closed-API frontier model as a modular compression layer. We evaluate FOCUS on diverse agentic benchmarks including API and tool-calling, QA, web domain and multi-turn dialogue. Our method establishes new state of the art performance, cutting peak context by up to 48% and dependency by 73% while improving task success by up to 8.9 percentage points over uncompressed execution.

Longitudinal electronic health record (EHR) modeling requires integrating new visits with an expanding patient history. Yet the continual accumulation of clinical information imposes increasing computational and memory costs on large language models (LLMs) when they process and retain complete patient histories. A practical alternative is visit-wise recurrent compression, which incorporates each incoming visit into a compact, continually updated patient memory. However, under a fixed memory budget, successive updates must integrate new information without progressively losing critical historical evidence needed to subsequent tasks. To address this challenge, we introduce Recurrent Longitudinal Memory (ReLMem), a framework that learns to maintain fixed-capacity patient memory for efficient downstream prediction with a frozen LLM. ReLMem equips this LLM with lightweight compression adapters to recurrently update the memory from its previous state and each incoming visit, without rereading earlier records. Specifically, we develop a multi-granularity optimization strategy to preserve task-relevant information throughout recurrent updates and support downstream prediction from the final memory. The intermediate supervision aligns attention outputs from compressed memory and the full history under identical queries, while prediction supervision minimizes cross-entropy with ground truth answers conditioned on the final memory. On EHR-based medication prediction, ReLMem approaches the F1 scores of full-history baseline while reducing average retained historical storage by 97.1%. Under the same memory budget, it improves macro- and micro-F1 over the strongest compressed-memory baseline by 4.66 and 4.75 percentage points, respectively. These results highlight the value of learning recurrent patient memory for efficient longitudinal EHR modeling.

Purpose: Most Greek papyri remain unpublished and undigitised; a handwritten text recognition (HTR) pipeline that transcribes them automatically would let scholars discover documents and literary works that have so far gone unread. Recognition systems for Ancient Greek papyri are in statu nascendi, and how accurate they must be for a given papyrological task has not been examined. To answer this and set a benchmark for Greek papyrus HTR, we test a range of character error rates (CER) against four papyrological tasks, using published editions as ground truth. Methods: From 63,846 current editions of Greek texts in papyri.info, we imitate a letters-only "perfect HTR" output by removing the editorial layer, then degrade it with a seeded algorithm to exact CERs of 1 - 50%, with lost lines and four error-shape variants. On these data we train small models (TF-IDF, fastText, a character CNN, ByT5-small) for document type, dating and documentary-versus-literary classification, and apply eight keyword search methods. We compare models trained on clean text with models retrained at a specific CER level, and evaluate across CERs. Results: Tolerance differs by task. With clean-trained models, documentary-versus-literary classification retains 90% of its metric up to 20% CER; document type up to 7.5%; subtypes and search up to 5%; dating only up to 3%. Retraining on text containing character errors largely eliminates the sharp degradation that otherwise sets in above 15% CER. Models generally tolerate concentrated damage in a long document better than small errors spread across a short text. Conclusion: The study provides a CER target for each of the four tasks and shows that models trained on noisy text make current, imperfect text recognition useful for them.

Persistent textual memory allows language models to carry information across long interactions, but learning what to remember is fundamentally a credit-assignment problem. A memory rewrite may only become useful many steps later, while much of the observed utility may be inherited from information already stored before the rewrite. We introduce Memory Gain Policy Optimization (MGPO), which isolates the incremental value of each memory rewrite by crediting it for its marginal contribution to current and future downstream utility. This turns delayed memory utility into a direct learning signal for optimizing what information should persist. We study MGPO on document-level information extraction, where structured supervision makes the effects of individual memory updates directly measurable. MGPO improves extraction while reducing average memory length by nearly 80% relative to the initial memory policy before optimization. The learned memory policy also supports reuse and transfer across domains, downstream models without further training. These results show that effective memory learning depends not only on preserving useful information, but on identifying which memory updates create lasting incremental value.

Running a language model locally offers advantages in privacy, latency, and cost, but local hardware fits only small models, which are less capable than frontier models. The usual remedy for a hard query, escalating it to a cloud model, gives up the privacy and cost advantages of running locally. A deployment that stays local faces two decisions for hard queries instead. First, it can spend more computation on a query, e.g., reasoning before answering, which raises accuracy at a cost in latency, so it must decide which queries are worth the extra computation (efficiency). Second, some queries are beyond the local model, and delivering a wrong answer is worse than deferring the query to a human in the loop, so it must decide which answers are safe to deliver (safety). We show that both decisions can be made from the model's own hidden states. The prefill state, computed before any token is generated, predicts whether the model will answer correctly, and the answer state, at the end of the generated answer, predicts whether that answer is correct. HARISSA fine-tunes the model so that both states predict correctness, then makes both decisions with one policy that cascades through the ways of answering from cheapest to most expensive, skipping a way the prefill state predicts will fail and deferring the query when the answer it stops with is predicted wrong. On a device running a single model, HARISSA is within one accuracy point of chain-of-thought at 2.7 times lower latency. On a server holding four sizes of one model, HARISSA is more accurate than the FrugalGPT and Self-REF cascades at the same latency, and at the same deferral rate the answer state leaves fewer wrong answers than the standard confidence signals in five of six task and setting pairs.

Backpropagation (BP) dominates deep learning but imposes a massive memory tax. For example, training OPT-30B with Adam requires $\approx$ 600GB of GPU memory (assuming batch size 8 and sequence length 2048). Alternatively, zero-order optimization (ZOO) trains in inference-mode (requiring only $\approx$ 60GB for the same model): no stored activations, no gradients, and no optimizer states. However, ZOO convergence has lagged behind BP. In this work, we evaluate two methods to close this gap. First, we show that reallocating training compute budget from many steps to large effective batch sizes with many perturbations (or probes) but fewer steps, allows 1SPSA (Spall, 1992) to outperform zero order methods like MeZO (Malladi et al., 2023) with less training compute. Next, we introduce 1.5-SPSA, adding a single "clean" forward-pass per step to 1SPSA to calculate a cheap diagonal preconditioner in probe-space, which improves convergence rate and convergence by down-weighting high curvature directions. Benchmarking on 6 post-training datasets on both Qwen3 and OPT model families, we show that 1.5-SPSA achieves State-of-the-Art results over previous ZOO solvers with much less optimization steps. For example, we train OPT-13B (for direct comparison to MeZO) and find 1.5-SPSA achieves +3.1% accuracy on SST-2 over both MeZO and BP in only 70 steps vs. MeZO's 100,000 steps. Finally, we combine an 8-bit-packing random generator, triton fused unpack/apply kernels, and distributed parallelism to achieve fast and stable training of models as large as OPT-30B in-place on commodity GPUs (e.g. A100).

Mixture-of-Experts (MoE) layers are central to frontier language models, and recent architectures push toward more and smaller experts. In this regime, the standard linear router becomes a bottleneck: with $M$ experts and hidden dimension $h$, its per-token cost $\Theta(Mh)$ dominates the MoE layer once $M$ is large. We introduce MoRE (Mixture of Rank-reduced-routed Experts), which factorizes the router weight matrix at rank $r$ and reduces the routing cost to $O((h + M)r)$. We prove that rank logarithmic in $M$ suffices for routing expressivity when the number of active experts is fixed, and is necessary up to precision factors. We also prove that logarithmic rank preserves load balance in a Gaussian memorization model, and training on a synthetic phonebook task shows that low rank does not hurt memorization. At matched active FLOPs, the factorization allows a factor of $\Theta(h/r)$ more experts. To realize this gain in wall-clock time, we design a fused Triton kernel at inference that avoids expensive memory operations on HBM. Empirically, MoRE improves memorization on the phonebook task and performance on knowledge-intensive Q\&A benchmarks after pretraining, while matching reasoning ability. Code available at https://github.com/Matheart/MoRE_code.

Large language models (LLMs) have shown strong potential for lossless data compression, but existing approaches are constrained by the high computational cost and low throughput of autoregressive decoding. We propose HyperZip, an efficient and scalable LLM-based compression framework that leverages diffusion-based LLMs (dLLMs) with Multi-Token Prediction (MTP) to accelerate LLM-based data compression processes. We identify a trade-off in diffusion-based compression, where increasing decoding throughput degrades the compression rate. To mitigate this trade-off, HyperZip employs a hypernetwork to generate data-specific updates from a context representation, adapting the dLLM to the target data without costly fine-tuning, resulting in a low compression rate and high throughput. Extensive experiments show that HyperZip achieves a superior trade-off between compression rate and speed compared with state-of-the-art baselines.

Self-evolution lets large language models (LLMs) improve iteratively using their own generated data, but often suffers from self-evolution degeneration: performance improves, plateaus, then declines. Existing methods address this issue at the component level, targeting either the Questioner or the Solver, and overlook that self-evolution is a tightly coupled system. We propose a holistic framework based on learnable information gain, which measures how much novel, parameterizable information a round provides relative to the previous round. Theoretically, this gain equals the Kullback-Leibler divergence between the two rounds' data distributions plus their entropy change. Practically, it is estimated by fitting a small language model to the previous round and scoring new data via negative log-likelihood. Based on this diagnostic, we propose ATRI (Adaptive Training Regulation via Information-gain), which reweights samples within a round and halts training across rounds when information gain remains low. Experiments on popular datasets demonstrate the superiority of our proposal.

Language-model agents are usually trained by reinforcement learning from one reward per episode, and privileged self-distillation enriches it by letting the same policy, given a skill, teach its skill-free self through token probabilities. However, we identify two phenomena that question this channel. Invisible Advantage: a skill in context lifts WebShop success from 42.2% to 56.2%, yet changes the probabilities of fewer than a quarter of the sampled tokens. Much to Align: a skill changes the hidden states of over 80% of response tokens, in a way that linear probes can trace back to the specific skill. To exploit this, we propose Privileged Representation On-policy Self-Distillation (PR-OPD). After a GRPO warm start, the policy writes a hindsight skill for each trajectory, re-reads its own responses with that skill as a stop-gradient teacher, and aligns its projected hidden states to the teacher's at every layer alongside the reward objective, with no external skill library, separate teacher, or inference overhead. On ALFWorld and WebShop with two backbones, PR-OPD achieves the best overall results in every setting, improving over GRPO by up to 4.7 points in ALFWorld success and 14.0 points in WebShop accuracy. Code is available at https://github.com/balibata/PR-OPD.

The pre-training of Large Language Models (LLMs) is increasingly conducted across multiple data centers. As training scales to a larger number of accelerators, the fraction of time spent on computation decreases, while the fraction spent on communication increases. Therefore, frequent synchronization becomes a growing bottleneck. Local update methods reduce this cost by allowing workers to perform several optimizer steps between synchronizations. Most local update methods set the number of local optimizer steps between synchronizations before training and keep this interval fixed throughout the run. However, the best interval can change during the entire train process. If the interval and optimizer are adapted to the current training state, the communication frequency is reduced while maintaining the training performance. In this work, we introduce AutoLoCo, an adaptive training framework to reduce communication in LLM training. It adapts the local interval using scalar training statistics and corrects each outer update. Our method is motivated by two observations: 1) the appropriate local interval varies across training stages, and 2) changing the number of inner steps per interval creates a mismatch with an unchanged outer optimizer, requiring a correction to the outer update. We optimize this mismatch by correction of the outer optimizer for the momentum and the learning rate using the accumulated inner learning rate. Our experiments under communication constraints demonstrate that AutoLoCo reduces communication frequency by 27% relative to DiLoCo while maintaining training performance.

Molecular optimization is inherently iterative: a candidate is proposed, evaluated against several objectives, and revised while preserving a relationship to the source molecule. Most instruction-following models instead emit one edited molecule, forcing validity, property improvement, and similarity control into a single response. We introduce MARCO, an evaluator-grounded reinforcement-learning framework that trains molecular editors on bounded proposal--feedback--revision trajectories. MARCO aggregates shaped turn rewards into an undiscounted trajectory return for group-relative policy optimization. We evaluate two consequences of this training: Same-1 tests the trained policy under a one-response budget, while Same-5 tests whether the same policy can use verifier feedback when up to five responses are available. Across the three-objective MuMOInstruct benchmark, three Qwen backbones, and seen/unseen instruction splits, SFT-initialized MARCO obtains the highest product of property success rate and similarity in every reported primary setting. Same-5 further improves the observed score under the tested budget, while four-objective and public-checkpoint experiments test transfer across constraint sets and initialization regimes.

Self-distillation turns knowledge distillation into a closed learning loop and offers a path toward recursive self-improvement. Without an external teacher, however, the model must determine both what information can improve its supervision and which induced changes should be learned. Existing methods typically improve teacher-generated data or select training examples in isolation, leaving the information transferred between these stages unmeasured. We introduce InFlow, a retrieval-guided on-policy self-distillation framework that models this process as potential-to-realized information flow. InFlow first retrieves potentially informative sources using certainty-calibrated hidden-state trajectories, then measures their realized effect through the Jensen--Shannon divergence between the teacher's initial and retrieval-conditioned answer beliefs. Examples with larger belief shifts are selected for on-policy distillation. Our analysis formalizes the information optimized by retrieval and selection and relates the answer-level shift to the teacher--student distillation gap. Across four open-weight language models and three knowledge domains, InFlow achieves the strongest cross-model average among the compared selection methods, with ablations supporting both stages of the framework. Our code is available at https://github.com/1240148048/INFLOW.

LLM judges are inherently subjective, often favoring different responses in pairwise comparison when neither option is objectively wrong. To study this subjectivity, we introduce JudgeProfile, a framework that dissects LLM evaluation into perception (how a judge compares two responses across specific attributes like clarity, correctness, and detail) and prioritization (how much each attribute influences the final choice). We curate SubjectiveSet, a dataset of 50,013 response pairs from 17 public data sources, evaluated by 21 LLM judges across 87 attributes. We find a hidden consensus in perception: judges frequently agree on attribute judgments even when their overall choices diverge. Building on this separation, we first characterize each judge's prioritization using attribute weights estimated from its own overall choices. These weights differ across judges even when estimated from the same attribute judgments. We then learn new weights from reference labels to adapt their decisions to a target evaluation standard. Reweighting perceived attributes improves average held-out agreement with reference labels from 66.48% to 71.97%, outperforming fine-tuning and rubric prompting. Our findings show that understanding and steering the subjectivity of LLM judges requires attention not only to what they perceive, but also to how they prioritize it.

Chinese Semantic Error Correction (CSEC) targets semantic errors in Chinese text, which are typically more subtle and complex than spelling and grammatical errors but remain relatively underexplored. Existing LLM-based approaches face two recurring obstacles in this task: over-correction, and unclear interaction between Chain-of-Thought (CoT) reasoning and self-consistency decoding, such that the benefits brought by CoT cannot be reliably transferred to final corrections. We propose Vote-guided Advantage Allocation for CSEC (VAA-CSEC), a multi-stage framework that combines CoT distillation, Supervised Fine-Tuning (SFT), Reinforcement Learning (RL) and self-consistency decoding. During RL, we design a task-specific reward function that directly aligned with the minimal-editing principle of CSEC. We further introduce Group-Level Relative Policy Optimization (GLPO), which reallocates GRPO advantages according to the margin between individual rollout rewards and the vote-aggregated group reward, aligning the RL training objective with the self-consistency objective used at inference time. Experiments on CSED-C and NaSGEC-Exam show that VAA-CSEC outperforms all LLM-based baselines on CSED-C with an F0.5 of 47.72%, achieves the highest recall of 42.15% among all methods, and establishes a new state of the art of 41.55% F0.5 on NaSGEC-Exam.

Existing reactive Graphical User Interface (GUI) agents often fail in long-horizon, dynamic scenarios, where unexpected disturbances trigger attention-diverting and cascading failures. To address this, we propose PrecogUI, a pre-cognitive architecture that shifts the paradigm from reactive execution to proactive decision-making. Specifically, we design a Proactive Experience Pool (PEP), which caches recurring anomaly and success patterns as "state-action-result" tuples in a dual-memory repository. Furthermore, we introduce a Proactive Simulation Executor (PSE) that learns to forecast the next symbolic UI layout given a candidate action, enabling early anomaly avoidance and ranking candidate actions by predicted reliability. Finally, a Pre-cognitive Execution Controller (PEC) fuses these priors and predictions, prioritizes handling of foreseen anomalies, and ensures execution robustness through a closed-loop error correction mechanism. For robust evaluation, we develop AutoTraj, an automatic data-generation engine, to construct InterfereBench, a benchmark for long-horizon tasks with strong disturbances. Experiments demonstrate that PrecogUI surpasses state-of-the-art methods on InterfereBench while maintaining competitive performance on public benchmarks. The code will be publicly available.

Recently, Group Relative Policy Optimization (GRPO) and its variants have been developed for policy optimization and demonstrated notable performance gains. However, these methods usually incur substantial computational overhead due to per-question multi-rollout sampling and repeated per-token probability evaluation across rollouts. Furthermore, low-information or highly homogeneous trajectories can degrade downstream learning signal efficiency, hindering model optimization and limiting final performance. To address these issues, we propose FastRL, a novel plug-and-play reinforcement learning framework that simultaneously improves training efficiency and the effectiveness of policy learning. Specifically, 1) We introduce an advantage-aware pruning strategy to selectively preserve high-advantage trajectories while maximizing inter-trajectory gradient diversity. 2) Then, we design an adaptive rollout sampling mechanism to dynamically adjust the sampling scale across different training stages based on historical pruning distributions, balancing exploration adequacy and computational efficiency. Experiments demonstrate that FastRL can be seamlessly integrated into GRPO, DAPO, and GSPO variants, achieving an average 2.07$\times$ training speedup on Geometry3K and GeoQA8K-R1V, along with an approximately 1.64\% improvement in average accuracy on visual reasoning benchmarks. Source codes will be available at https://github.com/Nicozwy/FastRL.

GUI agents automate tasks on digital devices by grounding language instructions in visual interfaces. Existing group-relative reinforcement learning improves GUI action prediction by comparing the rewards of multiple responses sampled from the same GUI state. However, binary evaluation treats spatially different failed clicks as identical and provides no relative signal when all sampled clicks fail. To address these limitations, we propose Spatial Credit Assignment (SCA), which uses the screen coordinates of sampled clicks to refine group-relative credit. Specifically, SCA predicts each held-out response's reward from the other responses in groups containing both successes and failures, then uses the prediction residual to adjust credit. When all sampled clicks fail, SCA instead orders them by distance to the annotated target. These spatial references are used only to construct the training update; the deployed policy remains unchanged. We evaluate whether this correction improves the policy update itself by comparing its error and directional alignment with the exact return gradient in a controlled synthetic study. Across GUI grounding and offline action-prediction benchmarks, SCA improves grounding across professional domains and achieves the strongest results among reinforcement-fine-tuned models on most action-prediction metrics, with consistent gains across the reported GUI suites.

Explicit intermediate reasoning gives large language models (LLMs) a stronger problem-solving mode. We study learning from this think-mode advantage via on-policy distillation (OPD). OPD preserves student-generated trajectories and provides dense token-level teacher targets at student-visited prefixes. Privileged reasoning is used during distillation rather than student inference. Uniform ThinkOPD, a natural think-enabled OPD baseline, conditions a fixed teacher on one shared think trace and uniformly distills every sibling student response. Although its prefixes are on-policy, the trace need not follow a route compatible with every complete response: the same privileged trace can induce different teacher-student discrepancies even when responses reach the same outcome. We summarize this interaction with trace-response divergence (TRD) and introduce ThinkOPD, which routes supervision at the response level by combining group-relative reward gain with a TRD-based compatibility proxy. Final response weights are normalized within each rollout group. Across mathematical reasoning and code generation, ThinkOPD outperforms Uniform ThinkOPD in both same-model settings and both cross-model teacher-student pairs, and it exceeds representative rationale and self-distillation baselines in a controlled comparison. Controlled interventions show that outcome benefit and the TRD-based proxy provide complementary routing signals in this setting. Think-enabled OPD provides a controlled setting for studying how teacher advantage becomes transferable along student responses.

Mid-training equips pretrained large language models with specialized and reasoning capabilities, but the returns of this stage are bounded since additional serial compute yields little further downstream improvement and can even degrade some capabilities, which places a practical ceiling on how much compute mid-training absorbs. We revisit how this compute should be allocated to a single run or multiple similar optimizations. We find that branches forked from a shared checkpoint under various controlled recipe reaches measurably different regions of parameter space, and establish a form of compatible diversity that extending one run cannot supply. Therefore, we introduce Trajectory Soup, which distributes a mid-training budget over several independent branches, and consolidates strongest checkpoints selected on validation through intra- and inter-trajectory averaging into a single model. A local bias and variance analysis separates the two averaging levels, showing that inter-trajectory averaging removes residual error beyond the reach of averaging within a trajectory, while checkpoint selection carries a bias that bounds how many checkpoints are worth merging. Across model scales, learning-rate schedules, token budgets, and trajectory counts, Trajectory Soup improves aggregate downstream performance over the strongest single-trajectory average under matched budgets and keeps improving as budgets expand, with the advantage preserved after an identical post-training pipeline. These results position trajectory allocation and merging as a practical way to extend the compute-scaling frontier of mid-training beyond serial saturation.

Off-policy and on-policy distillation have traditionally been formulated as separate paradigms, each favoring a different property of distillation trajectories. Teacher-generated (off-policy) traces are typically high-quality but lie far from the student's distribution, whereas student-generated (on-policy) rollouts are more learnable but often contain erroneous reasoning. We view these paradigms as the endpoints of a policy continuum and posit that a more effective rollout policy may lie in between. We introduce \textbf{Interpolated Policy Distillation (IPD)}, which defines the next-token distribution at every decoding step as an explicit linear interpolation between the student and teacher distributions. The interpolation operates at the distribution level, token by token, and its coefficient provides direct control over the balance between trajectory quality and student learnability. Naively sampling from this policy would require sequentially querying the teacher at every token and is thus expensive. To make IPD practical, we accelerate it with a new speculative-decoding rule while exactly preserving the interpolated next-token distribution.At the trajectory level, the resulting rollouts naturally interleave student- and teacher-generated segments. Unlike recent heuristic segment-interleaving methods, however, this interleaving is induced by an exactly realized token-level interpolated policy rather than by hand-designed switching rules. Across text-only and multimodal reasoning benchmarks, IPD consistently outperforms both endpoint policies (SFT and OPD), their conventional two-stage combination (SFT-then-OPD), and recent heuristic segment-interleaving methods, demonstrating that token-level policy interpolation better balances trajectory quality and student learnability.

Persistent memory creates a control problem that retrieval relevance alone does not solve: a memory can remain highly useful after an update, deletion, or revocation makes it inadmissible for the current answer. We formalize this as a separation between utility and authority. A fixed finite penalty applied to an unnormalized utility score cannot guarantee exclusion under arbitrary positive-affine reparameterization of that score; by contrast, rank-normalized compensation is scale-invariant and therefore forms a stronger empirical comparator. Our prospectively frozen TIDE/LongMemEval primary was quarantined before a valid HELDOUT comparison because the materialized TIDE adapter conflated historical age with query-relative inadmissibility and the aligned LongMemEval split left no DEV set for the predeclared penalty selection. We therefore report a post-primary replacement diagnostic on Memora Remembering, where update/delete operations provide item-level forgetting state. On Qwen3-8B, DEV selected lambda = 0.6 from a ten-point normalized SOFT family. Across 185 HELDOUT units in 28 dependency clusters, HARD exclusion yields 4.04% balanced construct error versus 19.66% for locked SOFT, a paired difference of 15.61 points with a 20,000-replicate cluster-bootstrap 95% interval of [13.07, 18.76]. The effect is driven primarily by forgotten-value leakage while current-value recall is preserved. This is same-Q operator-comparison evidence, not a universal claim that scalar control fails, not an evaluation of learned authority inference, and not an independent downstream-harm endpoint.

Masked diffusion models (MDMs) generate sequences by progressively unmasking several tokens per denoising step, but their reverse process is typically factorized over positions, limiting sample quality in the few-step regime where diffusion's speed advantage over autoregressive decoding matters most. A recent line of work introduces a continuous Gaussian latent, trained as a variational autoencoder, to capture correlations across positions, but such approaches are prone to posterior collapse, where the latent is silently ignored. We propose Enhanced Mixture-of-Experts (E-MoE), which builds the reverse process as a mixture of factorized distributions over a discrete shared latent given by the expert-routing decisions of a Mixture-of-Experts (MoE) backbone, without increasing active parameters over the factorized baseline. Across synthetic multi-modal benchmarks, binarized MNIST, and LM1B, E-MoE improves few-step generation over factorized baselines.

Sparse mixture-of-experts (MoE) large language models scale model capacity by routing each token to a small subset of experts. Their routers are regularized with load balancing terms and learn affinity scores through the language-model objective. However, these objectives do not provide direct alignment between routing affinities and token-level error. We introduce token-error supervision for sparse routing in two forms. The first form predicts an error score per expert. The affinity-weighted aggregate of these scores is aligned to the next-token cross-entropy loss, while the individual scores attenuate affinity before top-$K$ selection. The second directly aligns the router's affinities to the model's objective without requiring an additional head or inference-time modification. Both formulations use the Itakura--Saito divergence or an exponential negative log-likelihood for aligning affinities and token errors. Across two sparse MoE backbones and four multiple-choice question-answering benchmarks, we evaluate both supervision mechanisms. On Granite, our method improves accuracy by approximately 2.3 percentage points on average over a parameter-matched routing baseline. With stronger supervision, the gain on ARC-Challenge reaches 2.94 points. Both mechanisms preserve the native sparse execution budget and aggregation policy. Our code is available in the supplementary materials.

Vision-language models (VLMs) may accept false visual premises, answering questions about a target object's color, count, location, or state even when it is absent. We call this reliability-critical behavior a target-absence grounding failure. Existing visual-grounding detectors primarily rely on generated responses, hidden states, or uncertainty measures. We present the first framework to leverage internal routing decisions in Mixture-of-Experts (MoE) VLMs to detect target absence before generation and guide selective correction. We extract target-token routing probabilities from Qwen3-VL-30B-A3B-Instruct and Gemma-4-26B-A4B-it, train a separate L2-regularized linear detector for each model, and use its predictions to selectively invoke a target-aware review prompt. Using routing alone, the Qwen and Gemma detectors achieve ROC-AUCs of 0.9988 and 0.9956 on GQA-Inpaint and retain 0.8095 and 0.7781 on the external OBER dataset, respectively. The resulting routing-gated policy improves end-to-end accuracy on GQA-Inpaint and OBER by +22.25% and +12.17% for Qwen, and by +13.42% and +1.39% for Gemma, without modifying model weights. Further analysis shows that the signal is localized to the target-object token, emerges in early MoE layers, and is distributed across partially substitutable experts. Although cross-dataset threshold shifts require recalibration, false-positive review causes limited harm overall, suggesting that intervention risk can be controlled through joint selection of the detector threshold and review prompt. Overall, we show that routing probabilities alone preserve actionable information about visual perception, allowing computation already produced by an MoE VLM to support low-cost detection and selective visual regrounding.

Linear attention replaces growing KV caches with fixed-size recurrent states, yet these persistent states can become a substantial memory bottleneck under concurrent serving. Directly quantizing recurrent states to low precision often leads to severe accuracy degradation, as quantization errors propagate through successive state updates. We discover that the impact of these errors depends on two complementary dimensions: temporally, errors in long-lived memory can persist across many decoding steps; spatially, errors in different key rows affect model outputs differently, while state magnitudes vary substantially along both rows and columns. Motivated by these observations, we propose STEPQuant, a spatial-temporal post-training quantization framework for Delta-rule recurrent states. STEPQuant allocates precision according to error magnitude and memory lifetime, and jointly fits key-row and value-column scales based on state distributions and key-row impact on output error. Experiments on Qwen3.8-27B and Kimi-Linear-48B-A3B-Instruct across both long- and short-generation benchmarks show that STEPQuant closely matches FP32-state accuracy under a nominal 6-bit budget and outperforms uniform INT8 in its 4-bit configuration. Integrated into SGLang with optimized GPU kernels, 6-bit STEPQuant achieves over 5x recurrent-state compression and reduces total serving memory by up to 68.7%. Our code is available at https://github.com/Dreamer-Toby/STEPQuant.

Large language models (LLMs) are increasingly trained with reinforcement learning from verifiable rewards (RLVR). An exact verifier can also support test-time scaling by selecting a passing response from multiple samples, while other deployments use beam search, adaptive sampling, or tools. We study single-sample decoding, where each query receives one response without search, to ask whether search-exposed behavior can be absorbed into the model. Existing verified-response post-training recipes do not generally distinguish problems already solved on the first decode from failures recovered within K samples. Under a fixed budget, this can spend examples repeating behavior the deployed policy already has. We introduce GapFT, which selects training evidence by the source checkpoint's single-sample outcome and fine-tunes on the Pass@K-Pass@1 gap: problems the policy fails on one sample but solves within K samples. We match training examples, processed tokens, and optimizer steps while keeping the objective unchanged. GapFT fills the matched budget with recovered failures and uses an exact decomposition to distinguish corrections of recovered and missed failures from regressions on first-decode successes. On LogiQA 2.0 and ReClor with Llama-3.1-8B, GapFT improves Pass@1 by 14.4 and 13.9 points over the source model, outperforms budget-matched uniform verified RFT at the same learning rate, and matches fine-tuning on the full verified pool using one third of the data. A single decode matches the source model's verifier-selected Pass@4 accuracy. A randomized control attributes gains to covering distinct failures, and our analysis relates available gains to transferable failure support. A three-seed Qwen2.5-7B replication retains positive gains over uniform RFT on both logic tasks.

We propose ReLOBGen, a method for generating limit order book (LOB) messages that are replayable by construction. Replayability is required for closed-loop market simulation, yet existing LOB message generators may produce non-replayable raw messages, i.e., messages inconsistent with the current market state. These generators therefore rely on post-hoc correction or rejection followed by resampling, which may alter the replayed message distribution or increase inference cost. ReLOBGen instead ensures replayability during generation: it selects the referenced order from the resting orders in the current LOB and then generates the remaining message fields to be consistent with that order and the market state. For realistic reference selection, ReLOBGen samples from a learned distribution over eligible resting orders, efficiently computed from cached order representations and a context-dependent query. It then enforces the consistency of the remaining fields by masking out invalid tokens. Together, these components enable efficient generation of realistic messages without post-hoc correction or resampling. In 500-message rollouts, ReLOBGen achieves 100% replayability, improves market realism, particularly for top-of-book statistics and the relative prices of LOB messages, and provides a $2.7\text{-}3.6\times$ speedup per replayed message over the LOBS5 baseline.

Large language models often improve problem-solving performance by generating multi-step reasoning paths, yet how to analyze the hidden states along these paths remains unclear. Existing approaches typically assign each intermediate state the final-answer correctness label and train probes across heterogeneous questions. We argue that this protocol obscures reasoning dynamics in two ways: (1) correctness prediction can exploit question-level variation rather than path quality, and (2) states aligned by absolute step indices may correspond to different functional phases of reasoning. In this work, we propose to view reasoning paths as phase-structured trajectories within fixed questions. We instantiate this view as PAIR, short for Phase-Aligned Intra-question Reasoning. PAIR samples multiple trajectories for each question, maps variable-length paths into shared relative phases based on normalized trajectory progress, and compares successful and unsuccessful trajectories only within the same question and phase. This yields phase-specific path-quality directions that better isolate path-quality signals from question-level variation. Empirically, we find that standard across-question correctness probes lose much of their predictive power under within-question evaluation, suggesting that these probes partly rely on question-level information. PAIR improves within-question trajectory ranking and Best-of-N trajectory selection across models and benchmarks. Phase-wise steering further shows that the learned directions can change generation outcomes, providing causal evidence that they capture trajectory-relevant information.

Legal contracts contain ambiguities that expose enterprises to financial and legal risks. Some ambiguities allow flexible interpretation without triggering disputes, while others lead to significant legal conflicts. This makes identification alone insufficient, and interpretable rationale analysis essential. We propose LAURA, a post-training framework for interpretable ambiguous clause identification. LAURA leverages knowledge distillation with an IRAC-Unlearning prompting technique to transfer knowledge from a teacher LLM to an open-weight student model (<=1B parameters), which is then trained using a joint objective combining classification and rationale generation losses. The framework supports both legal and non-legal stakeholders in making informed decisions about which ambiguities require further attention. Extensive experiments across 7 baselines and 7 open-weight models demonstrate that LAURA with Flan-T5 (250M) delivers state-of-the-art interpretability over all interpretable baselines while matching the identification performance of the best-performing opaque baseline.

Detailed image captioning requires accurate and comprehensive descriptions of fine-grained visual content, yet caption quality spans factual accuracy, information coverage, and clarity. Compared with conventional methods that rely mainly on high-quality supervision or holistic rewards, rubric-based reinforcement learning decomposes these requirements into explicit criteria and provides targeted, structured feedback. However, existing methods often use separate models for caption generation, rubric construction, and judging, which may lead to inconsistent interpretations across roles. Some dynamic rubric methods alternate updates between the caption policy and rubric generator while keeping the judge fixed, but staged optimization may still leave rubric construction and judging out of step with policy optimization. We propose MoCo Rubric, a two-stage framework that coordinates these roles. First, role-conditioned, shared-parameter multi-task supervised fine-tuning equips a single vision--language model to serve as the Caption Policy, Rubric Generator, and Rubric Judge. Then, the Generator constructs rubrics online from captions sampled by the current Policy, reference captions, and image evidence. The Judge provides rubric-based rewards, and only the Policy receives GRPO updates. As Policy updates change the candidates being evaluated, we use an exponential moving average of the Policy parameters to update one momentum model shared by the Generator and Judge. This gradual transfer lets both rubric roles track Policy updates without separate RL optimization while smoothing parameter changes that could disrupt their rubric capabilities under direct synchronization. Across five captioning benchmarks, MoCo Rubric achieves an average pairwise win rate of 72.83\%, the best mean rank in blind ranking, and the highest average score in caption-based question answering.

Post-training has been shown to significantly improve language models' performance on tasks with verifiable outcomes, including mathematical reasoning, software engineering, and computer use. However, whether the same approach can improve forecasting in financial markets is much less clear. Compared with tasks with verifiable outcomes, not only are realized returns noisy, but even what constitutes a relevant information set for making effective predictions is not obvious a priori: the model must decide which observations to gather and then commit to a numerical judgment before the outcome is known. We study this question in a chronological stock-price sandbox, where a language model gathers price, volume, relative-performance, and market-context evidence and predicts a future return. We post-train Qwen3-4B with supervised fine-tuning (SFT) on tool-use demonstrations, then proximal policy optimization (PPO) with a terminal reward given by the forecast score against the realized return. The resulting AURA-4B more than doubles the starting direction--magnitude score, from 20.94 to 43.31, and is comparable to frontier language models on this benchmark. Conditional magnitude agreement rises from 33.3 to 66.2, while directional accuracy changes from 62.9 to 65.4. SFT expands tool use, and PPO further increases the share of ranking and market-context queries. These results show that post-training can substantially improve financial forecasting performance, together with changes in how the model investigates the market, on this outcome-selected benchmark.

Ensuring the safety of reasoning large language models (LLMs) across languages is essential for their reliable deployment. However, when exposed to jailbreak attacks in non-high-resource languages, these models may generate unsafe responses even when their reasoning traces identify safety risks. To address this issue, we propose aligning cross-lingual thoughts and responses (ACTR), a framework that improves multilingual safety alignment by strengthening the use of existing safety reasoning. Specifically, we first present the think gap score (TGS) to compare the normalized contributions of reasoning traces to attention outputs during response generation across languages, and use reasoning- trace substitution to measure the cross-lingual safety gap. Next, using a corpus of jailbreak queries, we assess neuron importance through changes in response representations caused by neuron masking and compare the high-importance neuron sets obtained with reasoning enabled and disabled to identify safety think neurons that support the use of safety reasoning. Finally, we devise neuron-selective consistency optimization (NSCO), which uses a frozen judge model to reward agreement between the safety categories of reasoning traces and responses while updating only the parameters associated with the selected neurons, requiring no human-annotated responses or preference data. Across two reasoning models, ACTR achieves lower average attack success rates than the evaluated state-of-the-art methods on AdvBench-X and MultiJail, with safety gains extending to unseen languages, while preserving or improving average performance on multilingual knowledge and mathematical reasoning tasks and limiting false refusals of benign requests. Warning: this paper contains examples with unsafe content.

Post-training is central to mathematical reasoning in modern large language models (LLMs), but endpoint pass@1 alone underidentifies what has changed. Gains may reflect newly reachable solutions, cheaper sampling of latent solutions, surface robustness, or memorisation. We compare three post-training paths under a common diagnostic readout: our sufficiently trained off-policy distillation trajectories, released Qwen3 off-policy-plus-on-policy distillation endpoints, and a released DeepSeek-Math endpoint trained with Group Relative Policy Optimisation (GRPO). Our probe uses cross-surface pass@K over verbatim prompts, paraphrases, numerical isomorphisms, and translations, plus consistency, distribution-shape, and verified supervised-fine-tuning (SFT) membership analyses. We find two regimes. On easier AMC problems, large-K ceilings are near saturation, so post-training mainly compresses sample cost. On harder AIME problems, post-training expands the large-K ceiling over the base model: sufficient off-policy distillation already raises this ceiling, Qwen3 released endpoints raise it further, and DeepSeek-Math GRPO does not dominate sufficient off-policy distillation at large K. English-dominant distillation improves non-English reasoning but preserves language-tier gaps. A controlled-overfit audit finds limited sensitivity in current SFT-membership probes. Compression is one regime of post-training, not a universal explanation.

Language model agents can be improved by updating their model weights or refining the harness that guides task execution. These components are coupled: weight updates change how the model uses the harness, while harness updates change the trajectories used for training. We propose VACE, Validation-Gated Alternating CoEvolution, which alternates agentic reinforcement learning with trajectory-driven harness refinement. After each RL stage, VACE reuses the collected trajectories to propose a harness revision and evaluates the incumbent and candidate with the updated model held fixed. The candidate guides subsequent training only if it improves validation performance. With Qwen3.5-9B, VACE achieves 45.26% test accuracy on OfficeQA and a mean partial-credit score of 75.19% on AutomationBench, exceeding weight-only RL by 6.43 and 9.09 percentage points and ungated alternation by 4.59 and 6.95 points, respectively. Across 44 harness proposals, 17 reduce validation performance at the updated checkpoint and are rejected before subsequent RL training, highlighting the importance of validation gating.

An agent that uses tools typically responds to what the user explicitly asks, yet completing the task may require information the user never requested. Work on proactive agents mainly studies whether and when an agent should act on its own, not what information it should pursue. We study a distinct axis of proactivity: its content. Horizontal proactivity pursues unstated information that the current context already identifies, and vertical proactivity pursues needs that only earlier evidence reveals. A need graph, recovered from a benchmark's own decomposition, records which needs depend on which, so both forms, and whether the agent stops at the right time, can be scored from a transcript without a model judge. To learn this behavior, we propose Q&D (questioner and drafter), which trains a questioner to prefer the question whose continuation retrieves more of the required evidence, with no reward model or judge. On held-out splits of three multi-hop question-answering benchmarks, at equal retrieval spend, the trained questioner improves both forms of proactivity over the same model, prompted, and outperforms a prompted model $15\times$ larger in the same role on two of the three, and the gain persists after controlling for question volume and length. Without further training, we place the questioner in an interactive customer-service agent with a simulated customer, where it completes more tasks while asking fewer questions, and in retail it outperforms the $15\times$ larger model with fewer follow-up turns from the customer. These results show that proactivity depends not only on whether an agent acts without being asked, but also on what it chooses to pursue and when it stops.

Pre-trained language models (PLMs) with encoder-based architectures have shown impressive capabilities in zero-shot cross-lingual transfer for various language understanding tasks. However, applying this technique to dependency parsing remains a significant challenge due to its syntactic nature. To boost model generalizability across linguistic typologies, we propose a cross-lingual unsupervised bootstrapping method to improve syntactic knowledge within the PLM. We show that our method achieves a significant improvement in zero-shot parsing performance in low-resource languages. Analysis of these bootstrapped models uncovers increased robustness in recognizing syntactic structures, evidenced by higher scores in parameter-free tree probing tests.

Reinforcement learning for long-horizon agents typically relies on sparse outcome-based rewards. This leads to a severe cold-start problem, as early-stage policies often fail to solve sampled tasks, leaving little useful reward signal for learning. To mitigate this problem, we use on-policy distillation (OPD) to provide token-level guidance on the student's own rollouts. We find that the benefit of this guidance depends on the performance gap between the teacher and the student. When the teacher substantially outperforms the student, distillation helps guide the student through the early training stage where outcome rewards provide little learning signal. As the gap narrows and eventually reverses, however, continued distillation becomes less beneficial and may hinder further improvement. Motivated by this observation, we propose Gap-Adaptive Teacher Scheduling (GATS), which augments the student's RL objective with an OPD term whose weight adapts to the teacher-student performance gap. Specifically, GATS gradually reduces teacher guidance as the student approaches the teacher's reference performance and withdraws it once that reference is reached. This enables GATS to leverage task-trained teachers smaller than the student, since teacher guidance is primarily needed during early training. Across ALFWorld, WebShop, and ScienceWorld with three Qwen2.5 teacher-student configurations, GATS achieves the highest average success rate among the compared methods in all three configurations, improving over reward-only GRPO by 4.37%-11.87% under matched student rollout budgets. Code is available at https://github.com/Ricardo-H/guide-then-let-go.