2026-09-24

Computer-use agents (CUAs) increasingly act on behalf of users online. What happens when the environments they operate in have incentives that do not align with the user's? In online marketplaces, for example, platforms may favor some products over others, potentially steering agents away from the user's objective. Existing CUA benchmarks cover cooperative settings or explicit attacks, but do not test whether agents preserve user objectives when the environment itself has a stake in the outcome. We introduce CAVEAT, a controlled benchmark spanning nine marketplace environments and a taxonomy of eight common steering mechanisms. Across five model families, agents purchase the user-optimal product in 78.6% of matched-control episodes but only 17.3% when steering mechanisms are enabled. Larger models and increased reasoning improve robustness, but substantial failures persist. Our trajectory analysis and targeted ablations identify three points where steering enters the decision process: (1) agents distort the user's priorities, (2) prematurely narrow the set of alternatives they consider, and (3) commit before resolving decision-relevant evidence. Guided by this diagnosis, we develop CAVEAT-Harness, which directly targets these failure modes and raises user-optimal purchasing by 55.0%. Targeted post-training further improves a smaller open model. These results establish incentive robustness as a distinct challenge for delegated agents, diagnose how it fails, and show that targeted interventions can substantially improve it.

Long-horizon language-model agents accumulate reasoning traces, tool exchanges, and observations whose relevance changes with the current decision. Existing compression strategies often score historical units independently, but the safety of deleting several units is generally not determined by their singleton scores: redundant evidence, accumulated small effects, and the information that remains after deletion all matter. We introduce Direct Relational Set-Risk Pruning (DRSR), which formulates agent-history compression as risk-constrained selection over deletion sets. Offline, DRSR constructs exact counterfactual supervision by jointly deleting protocol-valid history Blocks and measuring the change in teacher-forced likelihood of the same recorded next output. A lightweight scorer then predicts set-level harm from online-visible relations between candidate history and the current pre-action state, together with deleted-retained and pairwise set structure. At deployment, DRSR evaluates a small set of structurally valid deletion candidates with the lightweight scorer and removes the largest feasible set under recency, protocol, budget, and learned-risk constraints, abstaining when no set is sufficiently safe. On WorkBuddyBench Full260, DRSR increases mean reward from 0.699 to 0.802 while reducing total model tokens by 20.820%. On the fixed Eval40 comparison, it obtains 0.794 reward at 1.211M tokens per task, using 35.850% fewer tokens than the uncompressed agent. Mechanistic analyses and ablations further show that decision-conditioned relations, retained-context information, pair interactions, and abstention each contribute to reliable pruning.

The Abstraction and Reasoning Corpus (ARC) has become a prominent benchmark for evaluating general abstract reasoning and fluid intelligence in AI models. Yet standard ARC evaluation considers only a single capability: producing the correct output grid for a test input. We argue that this narrow format fails to evaluate the diversity of abilities that genuine abstract skill acquisition should enable. We introduce PotARCin, a benchmark that extends ARC by assessing understanding of a task's underlying abstract rule across five dimensions: Definition, Classification, Constrained Generation, Editing, and Inversion. PotARCin employs programmatic methods to generate new task instances and transform given inputs for a given ARC task, enabling dynamic generative sampling beyond fixed input-output pairs. Across five state-of-the-art models evaluated on the ARC-AGI-1 training set, we observe a 25-52 percentage-point performance gap between standard ARC evaluation and evaluation on PotARCin, and find that multi-dimensional evaluation reorders models that standard accuracy ranks alike. We further investigate effects of generative sampling, difficulty of corruption types, and questions of self-consistency, showing that models frequently contradict their own formalized rule even where they have stated it correctly. We also introduce P-ARC, a held-out hand-crafted test set, on which models achieve 1-8% accuracy across all five dimensions, underscoring the importance of more holistic evaluations of abstract reasoning capabilities.

Scaling a language model is not only a question of final quality: the architectural choice determines how much computation is spent during training, prompt processing, and autoregressive decoding to achieve certain model quality. An ideal model architecture should lower all above computation costs to facilitate scaling to a larger model, while ensure the larger model indeed outperforms smaller baselines. We introduce KV-Invariant Transformer Expansion (KITE), a scaling paradigm that achieves this goal. It trains the model from a smaller size to a larger size (i.e., saving training costs via upcycling), while places newly added parameters in regions that do not affect attention KV. Consequently, during inference, prefilling KV only relies on the smaller part of the model, so the inference costs are saved. As a concrete instantiation, we present Step Scale Transformer (SST), a two-tower decoder in which one tower produces KV and the other reads them. At comparable cumulative training compute, SST, a 67B MoE model with 2.15B active body parameters per decode token, achieves lower training loss than 47B and 63B MoE Transformers with 1.48B and 2.02B active body parameters, respectively, while reducing estimated inference cost by 6.7% and 31.6%.

Human-written agent skills encode rich workflows for real-world problem solving, but are typically used as external inference-time instructions rather than internalized as reusable model capabilities. We introduce \texttt{SkillGym}, a framework that transforms these skills into executable, verifiable training environments for large language model agents. Its skill-to-task pipeline instantiates concrete tasks, verifies outcomes with code-based checkers, and assesses empirical skill dependence through contrastive executions. We construct and release 2,756 environments across 12 categories and collect 8,364 successful trajectories from multiple models and harnesses, averaging 49 tool calls and over 60k logged text tokens. These resources support supervised fine-tuning on verified workflows and reinforcement learning with outcome-based rewards. Under Claude Code, supervised fine-tuning improves Qwen3.5-35B-A3B by 199 Elo on GDPval-AA v2, 19.10 percentage points on Terminal-Bench 2.1, and 28.13 and 12.38 points on SkillsBench v1.1 with and without skills, respectively. Our 35B \texttt{SkillGym-Agent} reaches 51.47\% on skill-assisted SkillsBench, exceeding reported scores for Claude Sonnet 4.6, GPT-5.4 Mini, and DeepSeek V4 Pro. Without skills, it also surpasses skill-assisted bases under Codex and Claude Code, suggesting reusable procedural competence.

Agentic AI systems are increasingly adopting automated pipelines that integrate multiple tools. While prior research and benchmarks have studied about task success and task completion of these agentic systems, the research about agent to tool interaction, specifically in biology agentic workflow is limited. This study investigates specific failures in agent to tool interaction where a tool invocation appears successful, some or all of the information or functionality from the tool via API/ wrapper is incomplete or missing and there are no communications / notifications to the user or the agent about such missing information. We call this a silent failures as the user or the agents are not aware that such failure has occurred. For the purposes of this study we developed an audit mechanism to identify such silent failures in Agent to tool interaction, by examining 15 scientific tools (and their associated API documentation and tool documentations) integrated within ToolUniverse environment (ToolUniverse serves as our experimental environment rather than the object of the study itself). We structure our study around 7 failure locus characterising where the failure occurs in the chain. We observed 91 failures (manually validated post LLM based candidate discovery and automated testing), most frequent of them being missing data or fields and inconsistencies in search, filtering or ranking criteria. Most of the 91 failures occurred in API layer (51) or wrapper layer (25), with a potential of silent failure amplification downstream. The results show that silent failures originate upstream of the event and propagate downstream into apparently valid scientific outputs. We propose a concept of contextual reliability to handle such failures and suggest mechanisms for testing, disclosing, monitoring, and measuring such failures across the agent-tool interaction pipeline.

Judicial reasoning remains challenging for large language models (LLMs) to analyze. This paper contributes a sentence-level benchmark for evaluating the ability of LLMs to classify interpretive canons as articulated by Larenz in the tradition of Savigny. Our contributions are threefold. First, we operationalize this conception of interpretation as classification criteria. Second, we provide a dataset of decisions of the German Federal Constitutional Court annotated at the sentence level. Third, we report baseline evaluations of four LLMs from three model families under expert hand-written prompts, compared against prompts optimized with Genetic-Pareto (GEPA). Mean F1 over the seven binary subtasks clusters between 70.4 and 79.2 across models, with grammatical interpretation usually the easiest canon to identify and systematic interpretation usually the hardest; under the tested configuration, GEPA-optimized prompts do not systematically outperform the hand-written ones, suggesting that the expert prompts provide a meaningful baseline.

Group Relative Policy Optimization (GRPO) and related policy-gradient methods for training language model agents collapse an entire multi-turn rollout into a single scalar trajectory reward before it enters the policy update. When the task composes distinct skills, especially under sparse and delayed environmental feedback, this collapsing is lossy: the optimizer must implicitly infer which competency drove the outcome and how that should change behavior. We argue the right primitive is not a better scalar but a decomposition: trajectory reward should be split along subtasks before it enters the policy update. We introduce Reinforcement Learning with Decomposed Subtasks (RLDS), whose core is Subtask-Decomposed Advantage Estimation (SDAE): a replacement for the scalar GRPO advantage that splits trajectory reward into per-subtask shares on a fixed taxonomy, computes a group-relative advantage per subtask, and distributes per-token credit by weighting each subtask's advantage by its importance, concentrating it around the step where a reflection marks that subtask's execution as consequential. We evaluate on four agentic benchmarks: FrozenLake (sparse grid navigation), HotpotQA (multi-hop QA, one retrieval tool), ScienceWorld (long-horizon embodied science), and DeepResearch (long-form research, four tools, composite rubric reward). Heterogeneity diagnostics emitted during training show where decomposition pays off - gains scale with subtask heterogeneity, largest on the high-heterogeneity tasks ScienceWorld (+11.5 points, paired-bootstrap 95% CI [+9.8, +13.3]) and FrozenLake (+9.8 points, [+7.0, +12.8]), and within noise on HotpotQA and DeepResearch, where the diagnostics predicted little to recover. ScienceWorld is also more compute-efficient under RLDS than scalar GRPO (-10.9% wall-clock per step), as long rollouts amortize the fixed reflect-and-grade overhead.

Large language model (LLM)-based agents use Electronic Design Automation (EDA) tools to generate and revise register-transfer-level (RTL) designs under synthesis and verification feedback. Recent methods learn from this feedback by distilling reusable skills from execution traces or by training on rewards derived from EDA-tools. Both methods are typically evaluated on the tasks that produced the experience. Repeated access to benchmark feedback on the same task can reward task-specific revision rather than creating reusable knowledge that transfers. We introduce ChipMEM, a verification-grounded memory layer for EDA agents. It combines cross-task procedural memory with within-trajectory statistical guidance. Its procedural component distills and stores a skill only after it passes synthesis, simulation, or formal checks, rather than relying on model self-assessments. A Bayesian component maintains hierarchical Beta estimates over tool-call outcomes and ranks recovery strategies that succeeded under comparable errors. A common adapter applies the same memory interface to RTL optimization and testbench-generation agents while preserving each domain's tools and acceptance criteria. We measure performance on training tasks and evaluate whether learned skills transfer to unseen tasks. On RTLRewriter-Bench, under matched model and tool settings, ChipMEM produces equivalence-passing outputs on 39/54 scored designs versus 35/54 without memory; on the 49-design short suite, mean area improvement is 8.69% versus 5.66%. On held-out CVDP tasks, ChipMEM with a frozen procedural library achieves 20/20 accepted outcomes versus 18/20 without memory in a single evaluation per setting.

Large reasoning models can produce correct yet unnecessarily long reasoning traces. Existing methods improve reasoning efficiency with trajectory-level objectives or local token- and step-level signals, but rarely model inter-step semantic dependencies. This limits their ability to distinguish redundant steps from those that support later deductions, making it harder to shorten reasoning without sacrificing accuracy. We introduce RECAP (REdundancy-aware Credit Assignment via Propagation), which addresses this limitation by assigning credit where it is due based on both a step's downstream role in the reasoning structure and its contribution to solving the problem correctly. We define structural responsibility to capture the step's downstream role by measuring how strongly later reasoning depends on it, using credit propagated backward from the final-answer node through an outcome-independent, LLM-annotated semantic dependency graph. However, a step can have high structural responsibility yet steer the reasoning away from the correct solution. RECAP therefore introduces step efficacy to measure answer-directed progress through changes in gold-answer log-likelihood as each step is added. Together, these signals reshape rollout-level GRPO advantages into step-specific updates. RECAP requires neither a separately trained process reward model nor preconstructed concise trajectories. Across two 7B models and four mathematical reasoning benchmarks, RECAP improves the accuracy-efficiency trade-off. On Qwen2.5-Math-7B, it improves pass@1 by 2.0-3.7 percentage points while reducing reasoning tokens by 8%-31% relative to GRPO across all four benchmarks. Analysis suggests these savings reflect fewer reasoning operations and less dead-end reasoning, rather than more compact expression.

Transformer language models process sequences token by token in an autoregressive manner, making growing contexts increasingly expensive. Yet many adjacent token spans are highly predictable or frequently occur as stable units, suggesting that their representations may be compressible. We introduce a method for distilling sequential computation by replacing spans of input tokens with collapsed representations, computed on the fly by a lightweight merge module. This module generates a single surrogate embedding from a sequence of static token embeddings that captures the functional role of the multiple tokens, allowing pretrained models to operate on compressed inputs without architectural changes or re-training. We apply this approach during inference to compress both prompts and intermediate decoding steps, using a rollback mechanism to substitute stored multi-token KV cache entries with their single-step surrogates. Experiments across diverse models show that the merge module can be used to reduce effective sequence length by up to 40% with minimal accuracy degradation across language modeling evaluations and downstream tasks, including question answering, summarization, commonsense reasoning, and long-form mathematical reasoning. Additional lightweight adaptation of the merge module further improves the accuracy-compression trade-off in selected settings. These results demonstrate that sequential token computation in Transformers can be effectively approximated through condensed surrogate representations that approximate the original behavior without model updating.

Long horizon agents accumulate growing interaction histories that increase context and inference costs. We find that geometric redundancy alone is an insufficient criterion for safe compression. Although agent histories exhibit strong low dimensional structure, similar global geometry can preserve very different amounts of task evidence. At identical retained block counts, evidence aware selection raises next action Top 3 retention from 0.31 to 0.69, while centroid similarity remains 0.98. Controlled replacement further shows that action related information can be substantially altered while global geometric measures remain nearly unchanged. Motivated by this gap between geometry and evidence, we introduce Geometry Guided Evidence Preserving Memory (GEM), a training free compressor that protects task and execution evidence before using geometric residuals to complete coverage. GEM reduces mean combined token usage from 2.69M to 2.11M per task, a 21.4% reduction, while maintaining comparable task reward. Our results show that efficient agent history compression should optimize for preserved task evidence rather than geometric coverage alone.

Reinforcement learning (RL) has emerged as a key paradigm for improving the reasoning capabilities of large language models (LLMs). However, existing reward systems, such as rule-based and reward-model-based, often exhibit issues such as unstable optimization and reward hacking. In this work, we revisit the general reasoning of LLMs from a geometric perspective, conceptualizing it as a coupled manifold composed of three interdependent sub-manifolds: logical deduction, evaluation, and representation. Based on this perspective, response generation in RL can be interpreted as a decoupling process from the evaluation manifold, while reward estimation corresponds to a decoupling process from the logical deduction manifold. The limitations of rule-based and reward-model RL systems can be geometrically interpreted as the mismatch of policy-reward manifolds during RL process. To address the aforementioned misalignment, we propose Decoupling and Coupling Reinforcement Learning (DCRL) framework, which incorporates two key components: (1) a syllogistic logic-based prompt evolution mechanism that dynamically refines reward rubrics to enhance the expressiveness of the reward manifold; and (2) a policy-reward re-coupling mechanism that jointly updates the reward and policy models, ensuring consistent evaluation and mitigating manifold mismatch during training. Theoretical analysis and extensive experiments across multiple reasoning domains demonstrate that DCRL consistently outperforms both rule-based and reward-model baselines. Notably, a Qwen3-4B model trained under DCRL surpasses a Qwen3-32B baseline and approaches the performance of a Qwen3-235B model, highlighting superior effectiveness and generalization in RL.

Large language models (LLMs) have made substantial progress on medical question-answering, yet effective medical dialogue also requires learning to ask questions that uncover relevant patient information. To train such dialogue policies, a common pipeline combines supervised fine-tuning with reinforcement learning (RL) based on final diagnostic correctness. However, this outcome-based supervision does not directly distinguish the contributions of individual questions and provides no question-level feedback for unexecuted alternatives. To address this gap, we introduce PCQC (Privileged Counterfactual Question Credit), which uses privileged patient information during training to learn from questions never asked. During training, PCQC makes alternative questions directly comparable at the same dialogue state by using privileged patient facts to construct their answers. A frozen diagnostic scorer evaluates the diagnostic utility of each resulting question-answer pair by how strongly it supports the correct diagnosis. PCQC turns these comparisons into relative question credit that teaches the policy which questions to favor, directly supervising both executed and unexecuted questions alongside outcome-based RL without requiring complete rollouts for the unexecuted alternatives. Extensive experiments across four medical benchmarks demonstrate that PCQC achieves 63.10% mean diagnostic accuracy, outperforming GRPO and ATPO by 4.38 and 4.21 percentage points, respectively. These gains are achieved with 33.1% fewer inquiry turns than GRPO.

Recent advances in large language models (LLMs) have enabled agents to tackle long-horizon tasks across diverse environments. To further improve agent performance, existing language world models typically predict environment observations, yet reconstructing high-entropy, execution-dependent tool responses offers limited value when real feedback is available. Meanwhile, agents suffer from \emph{task-state contamination}, where unsupported assumptions and outdated plans persist in history and distort subsequent decisions. We propose the \textbf{Agent-Editing World Model (AEWM)}, which models how reasoning and actions shape future task progress rather than simulating tool responses. AEWM combines \textbf{Action Judge} to distinguish \textsc{Critical}, \textsc{Exploratory}, and \textsc{Noisy} decisions with \textbf{State Revision} to edit noisy reasoning--action continuations from the same observed history. \textbf{EditAct} integrates these capabilities with real execution, directly changing the state underlying subsequent decisions rather than merely providing critiques. We train AEWM across Search, Terminal, and Software Engineering through mid-training and supervised fine-tuning. AEWM achieves 70.5\% macro-F1 on our Action Judge benchmark, exceeding the strongest frontier baseline by 10.6 points. Across six benchmarks and three agent backbones, EditAct improves average scores by 3.2--6.7 points over the strongest baseline. Furthermore, rejection sampling fine-tuning on verified EditAct trajectories, termed \textbf{AEWM-RFT}, improves over Self-RFT by 2.2--2.6 points across three domains without online AEWM guidance.

In statutory question answering, every claim must be traceable to evidence, not merely relevant, since unverifiable labour-rights answers carry serious legal consequences. Current systems fall short: single-pass RAG cannot detect insufficient evidence, while multi-agent legal-debate systems treat grounding as a prompting convention, letting agents cite unretrieved evidence. To address this gap, we introduce LabourCrew, a multi-agent RAG framework built around three grounding mechanisms: StatuteGraph, a graph index that explicitly links chapter, section, proviso, and cross-reference structure rather than fixed-length spans; an Evidence Exchange Protocol that confines advocates and an interpreter to an evidence ledger, making citation to unretrieved text impossible, while a fault-tolerant supervisor board runs advocates in parallel so individual failures degrade rather than crash the system; and a Calibrated Trust Gate that replaces categorical accept/reject decisions with a trust score, thresholded via conformal risk control for a distribution-free bound on the false-accept rate. We evaluate on LabourActQA, a 500-item Bangla question set from the Bangladesh Labour Act, 2006, spanning seven reasoning categories and three difficulty tiers. The framework drives the empirical false-accept rate to 0.081, within the target level ($\alpha = 0.10$), achieves the highest Answer Relevancy among HyDE RAG, Graph-RAG, and Hierarchical RAG (0.862 $>$ 0.839, 0.815, 0.828), and degrades gradually rather than catastrophically as question difficulty increases. These results show that calibrated abstention, not retrieval quality alone, is what makes legal question answering auditable in low-resource statutory domains.

Reinforcement learning with verifiable rewards (RLVR) supervises mathematical reasoning through final-answer correctness, but provides little guidance on individual tokens. On-policy distillation (OPD) supplies dense feedback on student-generated responses, yet teacher preference need not reflect correctness. Recent hybrids combine OPD and verifier-derived advantages or reweight task credit using teacher ratios. However, teacher guidance enters after verifier-based group normalization, and token reweighting need not preserve the total task credit assigned to each response. We introduce Unified Entropy-Calibrated Credit Redistribution for GRPO (UECR-GRPO), which integrates verifier and teacher signals within a single GRPO-style update at both the response and token levels. \emph{Path-Utility Unification} (PUU) combines verifier reward and a teacher-to-anchor path log-ratio in a single KL-regularized objective. Its on-policy implementation uses a length-normalized teacher score and combines both rewards before group normalization and PPO clipping, allowing teacher evidence to influence the response ranking. \emph{Entropy-Calibrated Redistribution} (ECR) then uses the signed teacher--old-policy token gap to redistribute the verifier-derived component. Full-vocabulary teacher entropy attenuates uncertain guidance, while a response-wise zero-sum projection preserves the total task credit and its token-wise sign before clipping. Across five mathematical reasoning benchmarks, UECR-GRPO achieves average \(\mathrm{Avg@12}\) accuracies of 17.21\% and 65.09\% with Qwen3-1.7B and Qwen3-4B students, respectively, exceeding the strongest baseline at each scale by 0.89 and 0.56 percentage points.

Graphical User Interface (GUI) agents enable the fulfillment of complex user instructions through multi-turn interactions with software environments, requiring step-wise reasoning and long-horizon memory to guide actions and retain task-relevant information, respectively. Recent on-policy self-distillation (OPSD) methods have achieved strong performance on GUI grounding, a foundational subtask for GUI agents, owing to dense token-level supervision from privilege-conditioned self-teachers. However, extending existing OPSD methods to multi-turn GUI agents is hindered by self-teachers' limited privilege-following ability and insufficient privileged guidance. In this paper, we introduce GUI-SD-v2, the next version of GUI-SD, which extends OPSD from GUI grounding to multi-turn GUI interaction and addresses key limitations through a two-stage training framework. Specifically, GUI-SD-v2 first strengthens privilege following by jointly optimizing rollouts with and without privileged guidance from the same GUI states. Furthermore, it selectively distills step-specific reasoning and memory guidance through a privilege-conditioned self-teacher, supporting action decisions and the retention of task-relevant information for subsequent interactions. Extensive experiments on two representative GUI agent benchmarks, AndroidWorld and MobileWorld, show that GUI-SD-v2 compares favorably with existing OPSD baselines while consistently outperforming the evaluated state-of-the-art methods in both Pass@1 and Pass@3 success rates. Code and training data will be publicly released.

Benchmarks for full-duplex spoken dialogue models score turn-taking with binary fixed-window rules that reward immediate response or silence by completeness of the prior turn. We argue that the appropriateness of a response offset, whether delayed silence or anticipatory overlap, is conditional on the speaker's latent intent, identifiable only from that speaker's behavior. We introduce TACT, a benchmark of 9,728 episodes and 73.2 hours from five dyadic corpora; each episode carries dialogue history, a per-speaker memory profile, and an annotator-derived posterior over six intent classes. Scoring replaces binary windows with a strictly proper threshold-weighted continuous ranked probability score whose weights are intent-conditioned timing kernels fitted to human floor-transfer-offset distributions, proving boundedness, consistency, and binary reduction. Across eleven systems the best model reaches 0.47 against a human topline of 0.86, is nearly invariant to speaker profiles, and TACT agrees with human judgments at Spearman 0.81 versus 0.46 for binary metrics.

Compact vision-language models (VLMs) now power a growing share of multimodal applications. The benchmarks used to compare them, however, inherit a frontier-centric design: each model is reduced to a single accuracy number, narrowing the inter-model gap on saturated suites and pressing models into low-score bands on harder ones. We introduce PRISM-VLM, a multi-axis discriminative benchmark that scores every item along seven axes covering the recurring failure modes (task quality, behavioral robustness, and capability bottlenecks) and combines them into a single PScore, with items recycled from fifteen public benchmarks. Across compact VLMs from the past two years, PScore separates model pairs more reliably than prior single-axis benchmarks under an item-level paired bootstrap, and surfaces behavioral differences these benchmarks average away. Even models with statistically indistinguishable PScores diverge sharply along the per-axis profile, particularly on sycophancy, which is nearly orthogonal to single-prompt accuracy. We will release the full pipeline, prompts, and per-item annotations.

Solutions based on large language models (LLMs) often rely on temperature sampling to improve accuracy and stability by aggregating multiple samples from the completion distribution. However, this memoryless approach is inherently suboptimal: because it lacks awareness of prior generations and their evaluations, it produces an increasing proportion of semantically duplicate answers as more samples are drawn, leading to diminishing returns. To address this limitation, we introduce FLEET, a novel method that integrates a memory mechanism into the generation process. FLEET represents each generation as a sparse trajectory through states whose entropy exceeds a predefined threshold and uses these trajectories to infer per-token utility scores that adjust the logits. Benchmark evaluations demonstrate that FLEET achieves the same accuracy as the repeated sampling baseline, with a 3x speedup, and substantially improves accuracy on complex coding tasks (LiveCodeBench Pass@32 increases from 59.9% to 66.2%) under the same budget. Furthermore, in the greedy-decoding configuration evaluated here, the approach is deterministic and uses a single calibration pass to derive its principal hyperparameters, requiring only minimal modifications to existing LLM pipelines.

Language modeling using Transformers has become commonplace despite their fixed computational depth and quadratic runtime with respect to input tokens. Recurrent models on the other hand offer linear depth but no parallel execution. In this work, we extend balanced-tree recursive operators from sequence encoding to autoregressive prediction, enabling all prefix representations to be computed with logarithmic depth and linear runtime. Our experiments provide an initial characterization of this model class, demonstrating robust length extrapolation and performance approaching that of ALiBi-based Transformers, highlighting its potential as an alternative architecture for language modeling.

In-Context Learning (ICL) has become a cornerstone of modern LLM deployment. However, existing ICL post-training methods have a critical blind spot: they excel at extracting patterns from demonstrations while often neglecting context authority, the ability to determine whether contextual information should govern the final answer. To benchmark this capability, we introduce FakeContextBench, which contains pseudoscientific claims across seven domains. Our evaluation of commercial and open-source models shows that large-scale pre-training alone is insufficient for reliable context-authority discrimination. Moreover, prevalent ICL fine-tuning methods can increase susceptibility to misleading context, reducing reality accuracy by up to 14.95 percentage points relative to the base model. To address this trade-off, we propose Jurisdiction In-Context Learning (J-ICL), a post-training framework that incorporates context validation into the training objective. Across four model backbones, J-ICL improves ICLEval by an average of 5.84 percentage points and reality accuracy by 9.20 points over the corresponding base models. It also raises the Reality Rate by an average of 18.09 points relative to MetaICL and Symbol Tuning. These results demonstrate that ICL capability and resistance to deceptive context can be improved together. The benchmark is available at https://github.com/peilin717/FakeContext-Bench.

Mixture-of-Experts (MoE) training requires global load balance to prevent expert under-utilization and local balance for efficient expert-parallel execution. Existing distributed Quantile Balancing (QB) uses shard-dependent or approximate global quantiles, while token-independent expert biases cannot ensure microbatch-level balance. We introduce Exact Quantile Balancing (EQB), which computes exact global-batch BF16 quantiles with negligible communication, and Load-Error Injection (LEI), which injects local load errors directly into router-score gradients. On 7.5B-parameter MoEs trained for up to 500B tokens, EQB improves global balance and downstream performance over naive QB, while LEI improves local balance and outperforms the GShard loss at comparable quality.