2026-09-25

Reinforcement Learning with Verifiable Rewards (RLVR) performs well on problems with clear rewards, such as mathematics and coding, but whether it also works where the reward is less clear remains open. The reason-over-search recipe applies RLVR to open-domain question answering, where retrieval grounds the answer and a match against the reference supplies the reward. So far it has been demonstrated on large models, and below one billion parameters only with distillation from a larger teacher. We test the recipe on a small model. We train Qwen3.5-0.8B with Group Relative Policy Optimization (GRPO) and an interleaved Wikipedia-search tool on MuSiQue, varying only the reward across three shapes over three seeds each, and we evaluate every checkpoint held-out on a seven-benchmark question-answering suite. The recipe works: the best run reaches 0.352 average exact match against a 0.092 untrained floor, a 3.8-fold gain, with no distillation step in the training loop. The reward shape also matters. The Search-R1-faithful exact-match-only reward is the worst of the three at every seed at the matched training horizon, and it is worst even on exact match, the metric it directly optimises. We conclude that the sparse exact-match reward, RLVR's default in mathematics and code, is the wrong starting point for models of this size. The reason-over-search setting can supply a suitable reward for RLVR on small models, but small-model RLVR needs its own reward-design study rather than a scaled-down copy of a large-model recipe.

On-policy distillation (OPD) corrects a student on the responses it writes, but its signal is the teacher's next-token distribution: it tells the student what the teacher says but misses how it thinks. Latent supervision promises the missing part by aligning the student's latent states to the teacher's. Recent methods such as OPRD bring this signal into on-policy distillation. However, we observe two failures of this recipe when distilling Qwen3-4B and Qwen3-8B into Qwen3-1.7B-Base. Early gain, late collapse: latent supervision alone lifts MATH-500 accuracy from 25 to 46 in 10 steps, but subsequent training degrades performance down to 11 with no recovery. Better alignment, worse behavior: although the alignment metric steadily improves throughout this collapse, the most aligned model turns out to be the worst performing. Further analysis suggests a mismatch in how the latent signal is applied: layers paired by depth play different roles in the two models, so continued alignment may pull the student toward teacher states it cannot understand. To address this, we propose LastOPD, which applies the latent signal only at the last-layer state, the common interface both LM heads read, and only during a 10-step crossfade into token-level OPD. This keeps the useful part of the latent signal and hands the student to token-level supervision before the collapse sets in. Extensive experiments show that LastOPD improves MATH-500 over token-only OPD by 5.55 and 4.02 points with the 4B and 8B teachers, leads on most held-out datasets, and reaches the final score of token-only OPD in about half the steps. Code is available at https://github.com/Muyiiiii/LastOPD.

On-policy self-distillation (OPSD) has become a popular recipe for post-training LLM agents. It supervises the agent model at the token level with a stronger teacher view of the same model, obtained by conditioning on privileged information (PI). In this work, we show that in multi-turn agents, this paradigm teaches the student to act with confidence but without the information behind it. The trained agent behaves as if it had privileged information it never observed, and its performance falls well short of plain RL, in the worst case below the untrained base model. Therefore, we propose Privileged Self-Practice (PSP), which keeps the PI and moves it from the loss to the sampler. When the student's rollouts on a task mostly fail, we inject a short per-task instruction written by an analyzer model, sample the task again with the instruction in context, and train on the result with an unchanged GRPO objective. The privileged information stays in the prompt and never enters the loss. Across AppWorld and SWE-bench Verified, with three different student models, PSP obtains the best average score in every setting and is the only method that consistently outperforms plain GRPO, improving task-goal completion by up to 65% on AppWorld and the resolved rate by up to 61% on SWE-bench Verified.

Reasoning-capable language models often produce long chains of thought when direct answers suffice, wasting inference compute. Many dual-mode models leave this choice to users. Automating it is challenging because routing targets evolve with the policy, initial mode preferences destabilize exploration, and sequence-level objectives entangle routing with response learning. We introduce CounterRoute, an online reinforcement-learning framework that jointly learns routing and modeconditioned responses in one shared policy directly from a native dual-mode checkpoint, without method-specific SFT warm-up. Paired current-policy counterfactual rollouts assign cross-mode credit only to the routing token, while within-mode GRPO trains response tokens. A paired-to-self-routed curriculum stabilizes early training with forced rollouts from both modes, then increases self-routed updates to improve autonomous routing. Across nine benchmarks, CounterRoute better balances accuracy and efficiency than heuristic and learned adaptive-routing methods. Relative to always-thinking checkpoints, it improves macro-average accuracy while reducing mean generated tokens by 51% for Qwen3-8B and 41% for Qwen3-14B. On instruction-following and commonsense benchmarks where direct answering is strong, think rates fall as low as 1% while response quality improves. Despite training only on math and instruction following, its routing behavior and response quality generalize to held-out coding, science, knowledge, and commonsense benchmarks.

Direct On-Policy Distillation (Direct-OPD) transfers reinforcement-learning-induced policy improvements from a small model to a larger student by using the token-level log-ratio between post-RL and pre-RL checkpoints as dense supervision on the student's own rollouts. This transfer rewards the policy shift at every state, yet the log-ratio measures only relative change: it can stay fixed even as the probability mass that both checkpoints assign to the student's candidate tokens vanishes. Through an exact construction, we show that the Direct-OPD reward and its update can remain unchanged while the Jensen-Shannon divergence (JSD) and both KL directions between the checkpoints vanish with this mass, and we note that a small JSD bounds how much the teacher's behavior changed. Motivated by this analysis, we propose Selective Supervision for Direct-OPD (S$^2$D-OPD), which ranks student-sampled states by their teacher-reference JSD and masks Direct-OPD supervision at low-divergence states, retaining only the top 10% of states per response. Across two teacher pairs and four student models ranging from 1.7B to 8B parameters, S$^2$D-OPD improves held-out accuracy over dense Direct-OPD on AIME and HMMT benchmarks in seven of eight settings and matches it in the eighth, without extra forward passes. Our code is available at https://anonymous.4open.science/r/S2D-OPD-8868.

Machine Translation (MT) for low-resource languages remains far behind that of high-resource languages, and the gap is widest in specialised domains, where parallel data is scarce or entirely absent. We present EnSiTa, a trilingual multi-domain parallel dataset and benchmark for English, Sinhala and Tamil. EnSiTa provides human post-edited training data for seven domains, plus manually translated test sets for those and one additional domain, all produced by professional translators under a multi-year, rigorously quality-controlled process. Using this dataset, we conduct an extensive study of domain-specific MT for all six language directions, fine-tuning a from-scratch Transformer, a pre-trained translation model (NLLB-600M), and decoder-only LLMs (Gemma 3 family, 1B-12B, and TranslateGemma) across training-data sizes, model scales, and in-domain, cross-domain, multilingual and multi-domain settings. To the best of our knowledge, this is the most extensive systematically documented multi-domain parallel data creation and benchmarking effort for low-resource MT. Our data and models will be publicly released.

Long-horizon tool agents often make useful progress without reaching terminal success, motivating partial-credit evaluation. Yet evaluators may reward milestones that were temporary, later reversed, or not attributable to the evaluated agent. Comparing an honest trajectory with a higher-scoring adversarial one is inconclusive if the latter made more genuine progress. We introduce PartHackBench, a controlled methodology that removes this confound. A private certifier admits a pair only when its trajectories match component-wise in both current-state predicate satisfaction and standardized agent attribution; score inflation, defined as f(A) - f(H), is measured only afterward. In 18 sealed held-out tasks in PB-CSTE, the frozen historical-target run produced matched adversaries for 15 tasks. Historical credit yielded mean inflation of .252, conditional attack success of 10/15, end-to-end yield of 10/18, and detected none of 14 strict rollbacks. Semantic LLM judges were more resistant but remained vulnerable, especially under evaluator-targeted attacks, while PB-CSTE current-state controls, defined as exact functions of the certified components, yielded zero inflation by construction. PartHackBench thus provides a certified control for testing whether evaluator credit changes while all benchmark-defined task-relevant progress remains fixed.

Recursive AI, the prospect of AI taking an increasingly complete role in building and improving AI, is a crown jewel of AI for AI. Although recursive self-development has become practical for small models, bounded tasks, and fixed time budgets, a more consequential realization of this ambition, i.e., developing a release-ready, frontier-competitive model, remains far more challenging. In this work, we ask how little human involvement is sufficient for an agent to develop a frontier model. We concentrate human input into a high-density, low-frequency interface: experts encode objectives, stage scaffolds, permission boundaries, and operating procedures as reusable research skills, while the agent instantiates these priors, selects experiments, diagnoses outcomes, and revises the training strategy. In the challenging domain of industrial coding, the agent evolves data and coordinates SFT, on-policy self-distillation, and reinforcement learning with verifiable rewards, ultimately producing iCoder, a 27B model for RTL design and GPU kernel optimization. Across seven benchmarks, iCoder leads RTLLM, outperforming GPT-5.5 and Claude-Opus-4.8; ranks second on CVDP and KernelBench L2, exceeding GPT-5.5 by 16 points; and ties Claude-Opus-4.8 for the best TritonBench result. Exploratory case studies further show iCoder's competitive iterative RTL and GPU-kernel optimization with substantially fewer tokens. These results chart an engineering path toward recursive self-improvement, in which humans distill the principles of model building, agents operationalize them through evidence-driven experimentation, and each generation of AI becomes a more capable architect of the next.

Most agentic question answering (QA) systems do an important part of their semantic work at the worst possible time: every time someone asks a question. When a corpus contains revisions, drafts, revocations, deletions, and sources with different levels of authority, the model must reconstruct the governed current state on every read - then throw that work away and repeat it on the next query. This is a bit like a database that rebuilds a materialized view every time someone reads from it. We present ingest-time fact compilation, an architecture that performs this work when corpus data is ingested or changed. Raw passages are rephrased into self-contained facts; rules governing revisions, deletions, effective dates, and source trust are resolved once; and the resulting state is stored as typed records carrying source and revision provenance. At query time, an inexpensive model reads the compiled record instead of reconstructing it from noisy candidates. In a controlled synthetic experiment across five seeds, the same low-cost model produced the correct value, source, and revision in only one of 30 trials under query-time reconstruction, but in all 30 trials from the compiled substrate, at 12.89 times lower mean read cost per question. On simpler revision questions both architectures were exact, but the compiled path used 21.6 times fewer tokens. A separate test found that fact rephrasing roughly halved verbose Federal Reserve dialogue while preserving high source entailment, but left concise Wikipedia prose essentially unchanged. These results support a narrow but practical claim: resolving a corpus state once can make subsequent QA cheaper and more reliable for inexpensive models. We release the open source, MIT-licensed implementation and experimental artifacts.

Long horizon language model agents continually accumulate reasoning history, increasing context length and inference cost even after earlier decisions have been executed and observed. Unlike static Chain of Thought compression, removing historical reasoning can change future actions and the resulting interaction trajectory. We study when such reasoning can be safely forgotten. We propose Interaction Aware Compression for Long Horizon Reasoning (ICLR), a training free online method that ranks reasoning blocks using frozen proxy entropy while preserving actions, tool calls, and observations. On 260 WorkBuddyBench tasks, ICLR improves average reward from 0.699 to 0.718, while reducing input, output, and cache read tokens by 25.5%, 14.4%, and 33.3%, respectively. Ablations reveal trajectory amplification, where local reasoning deletion produces nonlinear changes in total computation by altering subsequent interaction. Representation probing, activation patching, and controlled trajectory analyses further suggest that historical reasoning becomes more replaceable once task relevant derived state has been reliably externalized into code, files, tool outputs, or environmental feedback. These results characterize agent reasoning as dynamic working state rather than permanent interaction history.

Customer experience (CX) agents use tools and large language models to address customer requests and guide conversational interactions with an organization's products. Improving these agents, especially in regulated industries, is difficult: they must detect intent, follow complex operational policies and use tools reliably. Manual end-to-end testing offers limited coverage, while live experiments expose customers to failures that can erode trust. We present a hypothesis-driven simulation workflow for screening candidate CX agents before deployment. Synthetic customers react to agent responses and simulated tool outputs enable multi-step agentic workflows without invoking production backends. We use the Snowglobe simulator on Nubank's Card Delivery agent and its expanded successor, Card Management - Nubank's highest-volume chat-support agent in Brazil. Across 4 deployed versions, simulated and production version-level binary evaluator scores show high correlation. Simulation-guided iteration increased transactional net promoter score (tNPS) by 36.69 points in a live A/B test. We also screened open-weight configurations in over 16,000 simulated conversations. In a subsequent live A/B test, the selected model increased self-service rate (SSR) by 8.82 percentage points to the highest level observed at Nubank, with no statistically significant change in tNPS. Simulation made broad exploration of models, reasoning settings, and prompts feasible without customer exposure, enabling production improvements that would have been impractical to pursue through live experimentation alone.

Vision-language models (VLMs) have become a common foundation for autonomous mobile GUI agents, but most existing systems rely on the VLM for both planning and action grounding at nearly every interaction step, leading to substantial latency and model-serving cost. We introduce Jev-Mobile, which shifts this paradigm to low-frequency VLM planning and high-frequency lightweight execution: the VLM specifies local goals, the accessibility tree defines a structured executable action space, and Jev, a fast typed decision model, repeatedly selects actions within this space. This design allows multiple GUI actions to be executed under a single VLM decision, reducing expensive VLM inference while preserving adaptive interaction. On the full AndroidWorld task suite, Jev-Mobile achieves 79% task success, compared with 78% for SeeAct-V and 84% for a Step-wise VLM baseline. Among successful trajectories, it reduces mean end-to-end execution time by 32.7% and mean model API cost by 73.4% relative to Step-wise VLM. These results show that decoupling high-level VLM reasoning from low-level action execution can substantially improve mobile GUI agent efficiency while maintaining competitive task performance.

Reproducing a machine learning paper involves most research steps, from installing software and debugging to running experiments, work that AI agents increasingly do. We introduce RECLAIM, a benchmark of 100 NeurIPS 2025 papers that can be rebuilt yearly from new conferences. For each paper we fix in advance the result to reproduce, what counts as a successful reproduction, and a GPU-hour budget. An agent must reproduce that result using the paper and whatever its authors released. What the authors released decides the difficulty tier. Run-tier releases include code, data, and weights; Retrain-tier releases lack weights, so the agent trains the model; Reimplement-tier releases lack code, so the agent writes it. A separate language model grades runs from logs and outputs rather than agents' reports. We run four agents once per paper; the best agent in each tier reproduces only 41% of Run-tier papers, 27% at Retrain, and 15% at Reimplement, where every agent does worst. Failed attempts use on average 29% of their budget, so most stop with budget left. The most common agent error is writing the method without checking any part against the paper's numbers, in 63 of 400 runs.

We introduce Forecast-Dojo, a replayable environment for benchmarking and training LLM forecasting agents. It combines resolved prediction-market questions with dated news, allowing agents to research an event and revisit their predictions at successive historical dates. The same tasks and tools support repeated evaluation, collection of training interactions, and feedback from recorded outcomes without waiting for new events to resolve. Forecast-Dojo contains 1,568 Polymarket events, split by time into training and evaluation periods, and 18.8M dated news articles. In an evaluation of 12 models, research tools lower Brier score for all 12. Forecasts also improve as events unfold, with the largest gains at steps where more newly dated evidence is recorded. Every model still trails historical market forecasts in both Brier score and accuracy. A belief notebook carried between dates lowers research cost but does not consistently improve forecast quality. Beyond evaluation, Forecast-Dojo provides interaction trajectories and outcome feedback for agent learning, with supervised fine-tuning as a proof of concept.

Tool-using agents are continually updated with new interaction data. After each policy update, however, previously estimated action credits may become stale. Recomputing them from scratch can require many additional tool calls and environment interactions, making repeated updates increasingly expensive. We ask a simple question: when does historical action credit actually need to be updated? Our key observation is that a change in action value does not necessarily imply a change in the decision. Historical credit can still be useful as long as policy-induced drift is too small to overturn the existing action ranking. Building on this idea, we introduce pairwise branch sensitivity to capture how strongly a policy update affects the downstream regions that distinguish two candidate actions. We then derive a first-order anchored credit-transport estimator that updates historical credit using old interventional trajectories, and propose a Decision-Sufficient Credit Gate (DSC-Gate) that chooses whether to reuse, transport, or resample credit. Experiments show that branch sensitivity explains credit drift substantially better than global policy distance. With sufficient historical data, credit transport reduces estimation error, while its benefit to decision making is concentrated on updates that affect action-distinguishing branches. On a fully independent test set, DSC-Gate changes mean regret by only +0.00004 relative to a gap-based gate while reducing mean new tool steps from 472 to 286, a 39.4% reduction. We observe the same pattern after a real tool-agent parameter update. Overall, our results show that agents do not need to recompute action credit after every policy update: much of the historical evidence can be reused or cheaply corrected, reducing the additional interaction required to keep action decisions up to date.

Large language models (LLMs) are increasingly deployed as multi-turn agents that must sustain goals, use tools, and adapt to other agents over extended interactions. However, existing research lacks auditable, multi-turn, multi-factorial experiments that quantify LLM behavior under explicit constraints, with time-resolved statistics that reveal how behavior unfolds over long horizons. To address this gap, we develop a multi-agent micro-benchmark inspired by the Stanford marshmallow experiment: ReAct agents operate minute-by-minute with a "raise a question" tool under a per-step budget, while we factorially manipulate social context (broadcast vs. isolated), personas (age, hedonic drive), and metacognitive policy (mandatory vs. optional tool use). We analyze outcomes with Kaplan-Meier (KM) survival curves and discrete-time hazard models over a long risk horizon across 19,200 agent trajectories in 64 cells. Behavior shows a sharp early "eat" impulse, and only 75.9% of agents persist to the end. In a discrete-time hazard model, isolation reduces per-minute risk relative to broadcast, whereas a must-use self-questioning policy increases risk. On average, agents ask $\approx 7.12$ questions and hit the per-step budget in $\approx 6\%$ of minutes. Questioning declines faster under broadcast than isolation. Ablation experiments demonstrated that removing hedonic drive and/or persona age increases survival and completion, narrows the broadcast/isolated gap, but leaves the must vs. may ordering intact. The combined ablation (no hedonic + no persona age) yields the highest completion (approaching $1.0$). These results establish delay-of-gratification as a compact, multi-turn interaction benchmark that captures social contagion and tool-use dynamics in LLM agents, providing a reproducible testbed and statistics for analyzing long-horizon, multi-agent behavior.

Reinforcement learning (RL) has proven effective in enhancing the reasoning performance of large language models (LLMs), particularly in complex mathematical and programming tasks. However, this capability comes with systematic \textit{length misallocation}, in which models devote excessive reasoning to simple questions while terminating prematurely on harder ones, degrading inference efficiency with negligible accuracy improvement. Many length-adaptive methods mitigate this issue by allocating token budgets according to question difficulty, under the implicit assumption that harder questions benefit monotonically from extended reasoning. In contrast, we find that the effect of reasoning length on accuracy is concentrated on \textit{partially solvable} questions. Our further analysis reveals that explicit length rewards can produce unintended training dynamics. Motivated by these findings, we propose \textbf{CARE}---\textbf{C}ontrastive \textbf{A}ccuracy \textbf{R}eward \textbf{E}stimation---which compares the beneficial length adjustment per question from online sampled responses and applies adaptive length rewards within Group Relative Policy Optimization, with no extra hyperparameters or additional inference cost. Experiments across multiple reasoning benchmarks demonstrate that our method improves Pass@1 by up to \(4\%\) while simultaneously reducing reasoning length by \(37\%\), achieving higher token efficiency. Code will be available upon the acceptance of this paper.

The rapid progression of large language models is extending AI from passive content generation into the active workflows of engineering and scientific discovery. This shift raises a compelling question: can AI be both the object of development and an active participant in building next-generation AI systems? We explore this question by building Qwen-Planner-Agent within a closed-loop AI-for-AI framework for scalable development and iterative improvement. Mobile planning offers a demanding test of this approach: complex, long-horizon tasks challenge agent reliability, while costly real-device interaction limits development scalability. The framework connects data production, model training, and deployment through a shared action-feedback-verification contract. (i) AI for Data builds a human-gated agentic data flywheel in which specialized agents construct tasks, collect interaction trajectories, curate and balance training data, and use training feedback to guide subsequent data generation. (ii) AI for Training combines a supervised planning cold start with hybrid-environment online agentic reinforcement learning, where we introduce Competence-Aware Reward-and-Advantage Engineering (CARE) to reduce reasoning and tool-use costs while preserving task performance. (iii) AI drives model--harness co-evolution through an execution-evidence-driven loop that orchestrates memory, skills, and tools at runtime and feeds structured action feedback and preserved failure traces back into coordinated model and harness adaptation. Qwen-Planner-Agent achieves the best overall performance among all evaluated models and systems on MobilePA-Bench, improving over its base model across tool use, memory, skills, and sub-agent coordination. Further evaluations of our model show improvements across non-mobile agentic benchmarks while largely preserving general capabilities.

Online misinformation increasingly appears in spoken formats such as news clips, podcasts, interviews, political speeches, and social media videos, creating a need for fact-checking systems that can verify claims directly from speech. We introduce VeriSpeak, a probe benchmark for studying speech-based fact verification in Large Audio Language Models (LALMs). VeriSpeak contains 3,879 spoken claims spanning temporal, geographical, and relational facts, with balanced true and false labels. The benchmark is designed to examine whether factual verification ability transfers from text to speech, and whether retrieval-augmented LALMs can use textual evidence to correctly support or refute spoken claims. Our experiments reveal a consistent text-speech modality gap: LALMs that verify written claims reliably often fail on the same claims when spoken. Moreover, retrieval alone provides limited gains because models frequently conflate retrieved evidence with the spoken claim. In contrast, retrieval combined with explicit reasoning improves claim-evidence comparison, with a thinking-tuned LALM reaching 86.1% accuracy. VeriSpeak highlights that effective speech misinformation detection requires not only speech understanding, but also grounded reasoning over retrieved evidence. The dataset is publicly available via Hugging Face at https://huggingface.co/datasets/abhiram4572/VeriSpeak.

Faithful user simulation is fundamental to building, evaluating, and improving interactive AI at scale. However, plausible individual responses do not ensure that simulated users reproduce the intent evolution and outcomes observed in real interactions. We propose TRACER, a multi-turn user simulator that explicitly models users' evolving intent and learns to align simulated behavior with real interaction trajectories. TRACER is trained in two stages: supervised fine-tuning on real user dialogues, followed by multi-turn reinforcement learning. The RL stage combines hierarchical outcome- and trajectory-level rewards with deviation-aware advantage modulation, jointly mitigating reward sparsity and credit assignment in long dialogues. On real customer-service sessions organized into reference cohorts, TRACER-7B surpasses the strongest baseline by 11.4 conversion F1, while also achieving the lowest group-level conversion-rate error and semantic trajectory distance, and generalizing to out-of-distribution scenarios. Human Turing tests yield identification accuracy close to chance, supporting the perceived naturalness of generated conversations. Building on this simulator, we further introduce the Dynamic Marketing Benchmark, which jointly evaluates persuasion effectiveness and response quality of LLMs through simulated interactions, revealing that higher response quality does not necessarily correspond to higher conversion rates.

Mechanistic interpretability seeks to make verifiable statements about the internal behavior of large language models (LLMs). Many interpretability techniques struggle to scale with the increasing size and depth of architectures. Our solution to this is to introduce smaller models with structures that lend themselves to interpretability. In this work, we introduce the Stream Recursion Model (SRM), a modification of the Hierarchical Reasoning Model (HRM) designed to expose internal computational structure while remaining scalable. SRM organizes computation into multiple interacting latent streams that are updated through recursive refinement, enabling direct analysis of stream dynamics, causal contribution, and routing behavior. SRM achieves performance comparable to GPT-2 on a per-parameter basis. Our analysis reveals consistent and distinct behavior across streams, indicating structured specialization and interaction. These results suggest that SRM provides a practical architectural foundation for scalable mechanistic interpretability and opens up promising avenues for future research in both reasoning performance and interpretability.

Consumer-facing skincare AI must coordinate visual evidence, product information, tool use, and user-facing actions within explicit evidence and safety boundaries. This study evaluates SkinAgent AI, a non-diagnostic multimodal framework that combines visual concern routing with grounded and auditable LLM-based orchestration. The architecture includes routing for Acne, Pores, and Wrinkles; photograph-based skin-type estimation; count-informed ordinal acne-severity support; typed tools; database-grounded recommendation and action functions; deterministic safety, privacy, and evidence checks; approval before state-changing actions; and structured trace and replay mechanisms. Visual-model performance and system-level agent behavior were evaluated separately. Across three seeds, the skin-condition routing model achieved 99.84% +/- 0.07% accuracy. Skin-type estimation achieved 88.85% accuracy, while count-informed acne-severity support achieved 84.59% accuracy with a quadratic weighted kappa of 0.9076. On a locked but non-independent 240-case system benchmark, intent accuracy was 80.00%, exact tool-set match was 62.92%, and strict task completion was 47.08%. No violations or successful cross-user leakage events were observed in the finite safety and privacy test suites. Tool-selection errors, incomplete grounding of product attributes, and unreliable failure fallback nevertheless remained. These findings support the feasibility of bounded, database-grounded, and traceable agent orchestration for non-diagnostic skincare assistance. They do not establish clinical readiness, external generalization, formal privacy guarantees, or universal safety. Independent validation, expert assessment, robustness and fairness testing, and prospective evaluation in real-world settings remain necessary.

Privacy-preserving continual learning (PPCL) must reduce the reproduction of sensitive content while retaining useful knowledge across sequential tasks. Formal privacy guarantees characterize randomized mechanisms, whereas operational output control concerns whether a trained model selectively reduces the likelihood of sensitive content in its outputs. In this work, we investigate the latter together with continual-learning utility under realistic task evolution. Retention and privacy correction operate at different granularities: task acquisition requires broad preservation of current- and old-task behavior, whereas privacy correction targets sparse annotated positions. Joint optimization leaves the current-task preservation target continually changing. We propose SPARK, a retention-correction decomposition that first freezes the learned post-task distribution and then applies selective correction around this stable reference. Self-Distillation Replay learns the current task while distilling behavior from previous tasks, and Post-Task Privacy Correction reduces annotated-PII likelihood while anchoring current- and old-task non-PII behavior to the resulting checkpoint. Extensive evaluations demonstrate that SPARK achieves effective selective PII suppression while preserving strong continual-learning utility and knowledge retention across diverse settings. Code and data will be released upon publication.

Fine-tuning large language models on new objectives can improve helpfulness, instruction following, or domain-specific performance, but it can also induce regressions on safety-critical prompts. Existing safety-preserving fine-tuning methods typically control average safety loss or use weighted auxiliary penalties, which can obscure rare but severe failures. We propose a chance-constrained formulation for safety-preserving fine-tuning that limits the fraction of safety examples whose degradation relative to a reference model exceeds a prescribed threshold. Because the resulting empirical chance constraint contains a discontinuous indicator, we introduce a differentiable majorization of the violation rate, yielding a tractable conservative constraint. We then develop a constraint-aware gradient descent method that treats the majorized constraint as a safe set in parameter space and minimally modifies the fine-tuning direction to preserve feasibility. The resulting update admits a closed form and produces a tail-aware safety correction that emphasizes examples near or above the degradation threshold. We conduct an extensive set of experiments on harmful fine-tuning across three different tasks and three models and show that our approach consistently outperforms the baselines that exist in the literature. These results suggest that safety preservation in LLM fine-tuning is better viewed as a reliability-constrained optimization problem than as average-risk regularization.

Role-playing agents based on large language models have been widely applied in areas such as personalized assistance and social simulation. Recent RL methods typically train on a fixed scenario pool collected before learning begins. This creates a distributional bottleneck: as the agent improves, the scenarios where it performs poorly also change, while the training distribution remains static. Therefore, we propose AdvRole, an adversarial context rewriting framework that turns role-playing RL into a closed-loop curriculum. AdvRole alternates between an Actor that learns to role-play and a Rewriter that edits character profiles and dialogue contexts into actor-specific hard scenarios. The Rewriter is trained with a performance-gap reward, which favors rewrites that reduce the current Actor's score relative to the original scenario. As a result, the scenario pool evolves with the Actor and continuously targets under-mastered regions of the character-context space. Experiments on three role-playing benchmarks covering English and Chinese, as well as a new multilingual benchmark we release, show that AdvRole consistently outperforms baselines.

Machine translation (MT) for Indian languages remains constrained by the limited availability of high-quality, Indic-centric parallel corpora and evaluation benchmarks. Existing multilingual resources are largely constructed from English-pivot content and often fail to capture the linguistic diversity, cultural complexity, and domain-specific characteristics of Indian languages. We present COILD, an Indic-centric parallel corpus comprising over 1.16 million human-translated and human-verified sentence pairs, covering 20 Indian language pairs across the Indo-Aryan, Dravidian, Tibeto-Burman, and Austro-Asiatic language families. The corpus is built entirely from original Indian language sources collected from licensed repositories spanning eight domains with direct real-world applicability. Furthermore, we introduce a domain-centric benchmark comprising 2,000 expert-verified sentences to enable consistent multilingual and cross-lingual evaluation across Indian language pairs. To validate the effectiveness of COILD, we fine-tune two representative multilingual neural machine translation models, IndicTrans2-Distilled and NLLB-200. Experimental results demonstrate consistent improvements across language pairs, domains, automatic evaluation metrics, and human evaluation, highlighting the effectiveness of high-quality Indic-centric supervision. COILD provides a valuable training and evaluation resource for advancing multilingual machine translation and future multilingual language models for Indian languages.

Group-based reinforcement learning (RL) methods, such as GRPO and its variants, have become a leading paradigm for training reasoning and agentic large language models (LLMs). While their group-normalized advantage estimation is reliable at the response level, it becomes systematically biased at the step level, since coarse-grained trajectory-level advantages are hard to accurately reflect the contribution of individual steps (i.e, failed trajectories may contain valuable steps). Revisiting the foundational RL definition, we notice that GRPO's success on single-turn tasks stems from its advantage estimation strategy, which adheres to the basic definition: the mean reward of multiple actions sampled from the same state constitutes a credible state-value estimate. Extending the faithful estimation to step-level would in principle demand sampling multiple actions from each intermediate state, which is too costly on a per-state basis. To mitigate this issue, we propose a Graph-based Faithful sTep-level credit-assignment framework (GRAFT) that grafts all rollout trajectories into a trajectory graph, recovering node state-values via Bellman iteration on the graph, and assigning credit to each edge by the node value difference. Theoretically, the estimated step-level advantage faithfully adheres to the basic advantage definition in RL. To further ensure the reliability of step-level advantage estimation, we further propose Graph GAE, which extends GAE to the trajectory graph for reducing the impact of state-value estimation bias. Experiments across a range of multi-turn agentic benchmarks show consistent gains over GRPO and superior performance compared to recent agentic RL algorithms. Code will be available at https://github.com/xcyao00/GRAFT.

End-to-end event extraction remains challenging for large language models as it requires simultaneous identification of event triggers, classification of event types, and extraction of schema-grounded argument spans. We present EAGER, a reinforcement learning framework for generative event extraction that combines fine-grained verifiable rewards with Schema-Contrastive Advantage Estimation to alleviate advantage collapse under sparse binary rewards. Our reward design explicitly targets structural validity, extraction accuracy, groundedness, coverage, over-generation, and span precision. Experiments across seven benchmark datasets show that EAGER consistently outperforms prompting, supervised fine-tuning, and prior reinforcement learning baselines, achieving a substantial improvement over the strongest prior method. Results demonstrate that task-aligned verifiable rewards and contrastive advantage estimation substantially improve structured extraction.

Deployed agents run on handed-over knowledge: a frozen policy consults a briefing of consolidated items written before the stream begins. The world then moves while the store stands still: paths close, flags change, price bands move; every item was true at handover, and the failure is staleness, not ignorance. We introduce ERRAND, which treats revalidation as a priced errand: a recheck competes with the task it protects for the same scarce actions, funded only when the value per action of resolving a doubt clears a running wage. The errand index is single-peaked, vanishing at both ends of belief, so certainty in either direction costs nothing; free en-route receipts maintain on-path knowledge, and repair writes a version, never a deletion. Under equal action budgets in two drifting tool-use worlds, ERRAND clears every non-oracle policy on the preregistered calibers, primary in every setting and conditional at every binding budget, leading eager revalidation by 10.0pp at the base cap. Restraint wins: given no cap, ERRAND stops on its own, spending 11.0% of steps, while uncapped eager revalidation spends 70.7% and still finishes 4.5pp behind capped ERRAND. The margin sits where the briefing's coverage is thinnest, the shadow price of long-tail knowledge: a small budget, well priced, beats a bigger store that never rechecks.

Many LLM inference problems, including model routing, prefix-cache management, prompt trimming, and test-time search, can be viewed as optimization over a tree. This structure arises naturally from autoregressive generation: every prefix defines a node, and its continuations form a subtree below it. Internal nodes of the tree provide cheap but biased estimates of a region's value, while leaf evaluations are expensive but accurate. Hierarchical bandit methods can exploit this structure, but typically require a specific smoothness schedule to be specified in advance, even though real objectives are often only piecewise smooth and their optima may lie near sharp boundaries. We introduce CANOPY, a multi-fidelity tree bandit that learns where the smoothness prior is valid rather than assuming it globally. CANOPY uses cheap random-path probes to construct an online certificate of local aggregation bias, then directs expensive leaf evaluations toward cells where the certificate detects a smoothness violation. We prove fixed-budget and regret guarantees whose additional cost is additive in the number of discontinuities, recovering the smooth-tree rate when no violations are present and approaching structure-blind search as violations become dense. Across routing, top-$k$ identification, test-time search, caching, and prompt trimming, CANOPY consistently improves matched-budget performance, including $2.9\times$ higher top-10 recall on a 1000-model pool, $1.6\times$ more SWE-bench Verified issues resolved than best-of-$N$, and $3.6\times$ lower median time-to-first-token with prefix caching.

Conspiratorial discourse on social media is not always expressed through explicit claims or stable lexical markers. The same surface content may express endorsement, legitimate concerns, criticism, satire, or mockery. The main challenge is therefore not only recognizing conspiracy-related claims, but inferring the speaker's intent -- the utterance's illocutionary force. We argue that this can be achieved through the use of relevant social contexts and propose an agentic framework, equipped with a set of tools supporting social queries. We demonstrate the benefits of our approach on a unique dataset of Hebrew tweets, covering 80\%--90\% of the public Hebrew tweets published over a four-year span (late 2018-- early 2023), encompassing several election cycles as well as the COVID pandemic years and related vaccination campaigns. This extensive coverage can be used in recovering different social contexts. Evaluating our framework on a manually-annotated adversarial dataset, we find that context-aware workflows consistently outperform text-only classification and that the agentic framework performs significantly better than other frameworks and settings, including a non-agentic model exposed to the same contexts available to the agent. We further provide an analysis of the results, the errors and efficiency (token economy) tradeoffs. These findings support viewing the task of conspiracy detection as a socially embedded interpretation task, in which effective classification depends not only on access to contexts, but also on adaptive reasoning in which the agent uses tools on a per-case basis, asking only for evidence relevant to its current reasoning step.

Harnesses, the products that run AI coding agents, are multiplying, and enterprises are rolling them out to their employees: what started as pilots with a few hundred seats is scaling to tens of thousands. Most enterprises do not build these harnesses but buy them from large vendors, such as Anthropic's Claude Code or OpenAI's Codex. A harness decides which model answers, what the model reads, how the prompt cache is used and which subagents run, so it picks the rate on the price sheet and sets the volume bought at it. Enterprises that keep a proprietary or untuned harness at its defaults inherit these choices and their bill. We build a fast, customisable router in which Jev, a classifier with calibrated probabilities, labels every prompt against a bring-your-own taxonomy of agentic requests. Because one user turn is many requests over a prompt cache that belongs to one model, the router moves work only where no running conversation has to rebuild its cache: at session start, in side lanes and at subagent launch. From the price sheet we derive when a mid-task switch pays back, and a crossover: on long tool-heavy sessions the highest-priced model costs less than the next tier, as repricing about 10,000 real sessions from public datasets confirms. In an emulated enterprise of 10,000 seats with user behaviour taken from these datasets, the router recovers 14 to 21% of model spend at Anthropic's list prices of 21 September 2026, \$3.3M to \$5.0M a year. The paper also maps the risks across twenty harnesses, prices the dependence on one vendor's models, and proposes a control plane that enterprises can run from within, starting now, with a ladder for deciding later whether to own the harness.

Internet texts are replete with format tags that carry structural, semantic, and functional meaning. Current large language model (LLM)-based translation systems struggle to balance translation fluency with tag fidelity when processing tagged text. We argue that resolving this tension requires a systematic approach at three interconnected levels: data synthesis, capability building, and multi-objective alignment. At the data level, we identify and formalize a fundamental trade-off between structural tag diversity and translation naturalness in synthetic data generation; existing methods optimize for one at the expense of the other. We propose a hybrid synthesis strategy (Hy-LST) combining LLM-based synthesis tag method and Two-Stage LLM-based synthesis tag method to produce both diverse and natural tagged data. At the capability level, we decompose tag-aware translation into four sub-tasks of increasing difficulty in a multi-task supervised fine-tuning framework, enabling targeted capability acquisition and knowledge transfer. At the alignment level, we design three complementary reward functions under a group relative policy optimization framework, each targeting a distinct objective (fluency, tag fidelity, and tag-scoped translation quality), and show that joint optimization consistently outperforms single-reward alternatives. Experiments on six language directions (en2zh, en2ja, en2de, en2fr, en2ru, de2fr) demonstrate that each level contributes measurable improvements, and the complete system significantly outperforms existing methods. Qualitative analysis reveals specific error patterns and their mitigation after training with our method.

Agent skills provide reusable knowledge and instructions, yet agents must repeatedly infer how to apply them and which operation should follow. This couples task reasoning with control decisions, allowing prescribed steps to be omitted or applied incorrectly. We introduce HEXIS, which compiles agent skills into extended finite state machines that separate knowledge from control flow. Skill knowledge is incorporated into local instructions that guide reasoning and generation within states. The machine records execution progress and intermediate results, while explicit transition conditions determine subsequent operations. Our incremental compiler first maps skill clauses and tool interfaces to state operations, local instructions, data bindings, and transitions. It then aligns development traces with existing states to identify missing operations and dependencies. These are incorporated by adding or reusing states and refining their connections. Updates are accepted only after static checks and replay of the current and all previously accepted traces. Across four benchmarks and four executors, HEXIS improves success over Skill + ReAct by 16.1 percentage points on average. Qwen3.8-27B reduces execution tokens by 38.4-88.9% across benchmarks.

Query understanding (QU) plays a critical role in production search systems, translating raw user queries into search execution plans that drive downstream retrieval and ranking. While large language models (LLMs) have enabled QU to be framed as a structured multi-task generation problem (e.g., intent classification, query expansion), optimizing such models to produce search-engine-coupled outputs remains challenging: static, label-based supervision fails to capture how each component actually interacts with the underlying search pipeline to affect downstream performance. We present a search-aware reinforcement learning (RL) framework for QU based on a distill-then-RL paradigm. Teacher-student supervised fine-tuning (SFT) first yields a well-formed, schema-compliant policy initialization. The RL stage then optimizes each QU component with rewards derived from live interaction with the search engine, tailored to that component's operational role, rather than a single reward tied to the final search outcome. Experiments on Roblox search show that this component-specific optimization improves both per-component utility and downstream search quality, raising NDCG@20 by 8.9 points over the SFT policy and by 3.5 points over training with a single end-to-end reward.