2026-11-06

Frontier evaluations of tool-using conversational agents are conducted almost exclusively in English, leaving a critical gap for organizations deploying agents in other languages and regions. We present ${\tau}$-bench-LILT, a multilingual extension of ${\tau^3}$-bench covering German, Korean, Turkish, and Farsi in the airline, retail, telecom, and banking domains. We natively localize every layer of the benchmark: customer databases are populated with culturally appropriate records, user-simulator prompts are refined iteratively through contrastive annotation by native-language experts, and tasks are updated according to the localized database and to reflect realistic tasks in the target region. Evaluation across eight agent models shows the largest drops compared to English in the Korean and Farsi retail domains, by up to 61 points in pass^4. Code and data released at https://github.com/lilt/tau-bench-lilt.

Large language model agents are increasingly deployed through the Model Context Protocol (MCP), which allows them to discover and invoke tools across multiple servers at runtime. Existing MCP benchmarks primarily measure task completion; where they diagnose failures, they do so after execution rather than building specific reasoning, execution, and memory failure risks into task construction. We introduce MCPHallu, a benchmark for diagnosing hallucination-prone agent failures in MCP environments along three axes: reasoning, execution, and memory. MCPHallu contains 358 tasks across five capability domains and runs on 36 production MCP servers in isolated containers. Each task is designed with a structural trigger that makes one of four subtypes the dominant failure risk: Branch Collapse, Unreachable Goal, Tool Misuse, or Context Forgetting. This design enables subtype-level analysis rather than aggregate task success alone. We evaluate fourteen frontier LLM agents from eight vendors on 5{,}012 trajectories. Current MCP agents remain vulnerable to hallucination-inducing tool environments, achieving an average score of 0.642 on a $[0,1]$ scale (higher is better). Failures concentrate on Unreachable Goal tasks, where the average score drops to 0.458 and 23.2\% of runs pass, indicating that agents often continue issuing tool calls or claim successful completion rather than recognizing infeasible goals. Subtype rankings diverge sharply: a model's rank shifts by 6.9 of 13 positions on average across subtypes, and Branch Collapse is uncorrelated with Unreachable Goal across models ($r = -0.12$). MCPHallu is released at https://github.com/J1zz/mcp_hallu.

Can large language models (LLMs) advance theoretical computer science and resolve its long-standing open conjectures? To study progress toward this goal, we introduce TCSExam, a unified framework and benchmark covering the full range of TCS research. It evaluates theorem proving, solver engineering, and constructions alongside direct attempts to resolve open conjectures, capturing both concrete research accomplishments and progress toward long-standing research goals. TCSExam-Full contains 100 expert-reviewed tasks, with submissions checked by the Lean kernel or an executable task checker. TCSExam-Lite selects 20 tasks, five per track, reducing the number of task attempts by 80% at the same per-attempt budget. We evaluate eight agents on Full and nine on Lite. The TCSExam Index combines all four tracks: GPT-6 Astra leads Full (71.9), and Opus 5.5 leads Lite (72.3). Full and Lite rankings are strongly aligned across eight shared agents (Spearman's $\rho=0.93$), supporting Lite as a compact, discriminative benchmark. The evaluations yield improved solver outputs and mathematical constructions, together with independently verified special cases and bounds from open-conjecture attempts. TCSExam provides an extensible framework with task sources, Docker build recipes, and evaluation tools for reproducible use.

Large language models (LLMs) are widely used for information access, content production, and agent-based task execution, but their outputs remain vulnerable to subtle, high-risk hallucinations. To systematically evaluate and advance post-generation hallucination detection, we introduce HaluCheck, a large-scale, bilingual benchmark comprising over 42k instances spanning both factual hallucinations in text and faithfulness hallucinations in agent trajectories. HaluCheck features fine-grained, multi-level annotations for detection, localization, and correction, covering different tasks and perturbation degrees. Experiments on fourteen LLMs reveal that, while models show basic sample-level detection ability, they struggle with fine-grained localization, correction, and agent hallucination detection. Further analysis demonstrates that thinking, web search, and supervised fine-tuning each offer partial improvements, but none fully solve the problem. HaluCheck thus provides a solid basis for benchmarking and diagnosing LLM reliability across text and agent scenarios.

Sparsity and quantization reduce the cost of attention by skipping interactions and lowering the precision of those retained. However, useful interactions differ in their precision requirements: some tolerate low-precision scores while remaining sensitive to removal. We therefore introduce TriAttention, which assigns query-key (QK) tiles to skipping, 4-bit, or 8-bit computation, using 4-bit to retain quantization-tolerant interactions at lower cost. To represent compactly, we encode each head's computation regions with two diagonal-distance boundaries, keeping policy storage independent of context length. Offline calibration then evaluates candidate boundary pairs using normalized projected-output errors and complete-tile execution costs. Using these measurements, a model-wide allocator selects each head's boundaries to minimize calibration error under a shared computation budget. During prefill, the selected boundaries generate precision-grouped traversals for a fused kernel that skips omitted tiles and combines contributions from both QK precisions into one normalized attention output. The quantized keys and values prepared during prefill are then reused in a compressed key-value (KV) cache for decoding, reducing repeated reads of prompt data. Experiments show that TriAttention achieves end-to-end generation speedups of up to 7.88x over the Torch attention and 2.65x over the BF16 baseline. Meanwhile, TriAttention maintains quality comparable to the original models on general-purpose and long-context tasks. These results show how a compact retention-and-precision policy translates local sensitivities into attention computation and repeated KV-cache access. The implementation is https://github.com/HanzhiZhang-Ulrica/TriAttention.

Language agents increasingly work for users on the open Web, and finance is where trusting them is hardest. The data must come from the Web, the strategy lives in the user's head, and correctness depends on time, so an agent can fail silently: a strategy the user never asked for, unverified data, or a backtest that uses the future yields a report as convincing as a correct one. Existing benchmarks close the world somewhere, fixing the task, the data, or the timeline, and thereby remove these failures instead of measuring them. We introduce FINCOSMOS, an open-world evaluation paradigm for agents doing financial work. A simulated user holds a hidden strategy sampled from a compositional grammar and reveals it only when asked. The agent receives no data, so it must acquire data from the Web, build its own backtest under a data cutoff, and deliver verifiable results. Grading needs no reference answer. Programs verify point-in-time correctness and recompute every reported metric, the delivered strategy is checked against the hidden specification, and a run is accepted only when the user is satisfied and none of 15 veto checks fails. Across four coding agents and a model without tools in 150 sessions, no session is accepted: every session that the simulated user signed off fails a program check, and agents raise only 54--64\% of the details they were not told. Human experts attribute 88.33\% of failures to the agents. With FINCOSMOS feedback, an agent with persistent memory learned all ten tasks.

Spatial reasoning is a fundamental capability of multimodal large language models (MLLMs), yet existing benchmarks mainly focus on static scene understanding. When objects are moved or swapped, models must infer not only spatial relations but also how interventions change the scene state. To address this problem, we introduce CST-Bench, a benchmark for counterfactual spatial reasoning. CST-Bench contains 7,082 examples covering 6 intervention types and 12 transformation operators, with dense causal annotations including intervention sources, affected relations, expected effects, invariant relations, and metric changes. We benchmark general-purpose MLLMs on CST-Bench and further compare post-training strategies on CST-Bench and standard spatial benchmarks. Results show that methods effective for ordinary spatial reasoning can still fail under counterfactual interventions. We further formulate counterfactual spatial reasoning as causal state-transition modeling and propose CT-OPSD, which uses CST-Bench annotations to guide a teacher model in learning counterfactual state transitions and causal reasoning chains. These chains are then distilled into a student model that reasons only from real observations. Overall, experiment results establish CST-Bench as a challenging benchmark and reusable supervision resource for counterfactual spatial reasoning. CT-OPSD further provides an initial demonstration of how its annotations can support model improvement.

Reliable deployment of LLM agents requires grounding diverse user expressions in their environments. Although RL-finetuned smaller large language model (sLLM) agents rival larger prompted LLMs on embodied tasks, their robustness to instruction paraphrases remains unclear. We introduce ALFWorld-Para to evaluate this robustness under controlled paraphrases, revealing 21-52% declines in success rates (SRs), predominantly driven by object synonyms. Despite retaining knowledge of synonym referents, agents often fail to take the corresponding objects even after observing them, exposing a gap between referential knowledge and object binding. Partial observability further complicates binding, as referents may remain unseen or have already been encountered under different labels. To address this challenge, we propose BINDIntent, a training-free framework that probabilistically binds referring expressions to observed labels using semantic proximity and estimated persistence against unseen candidates. Without additional training or LLM calls, BINDIntent improves synonym-paraphrase success of RL-finetuned sLLMs by 17.5-49.1 percentage points (pp) with only 0.4 ms of CPU overhead per step. It also raises GiGPO-7B's overall SR from 61.8% to 77.4%, surpassing a prompted 24B LLM. Code and benchmark are available at https://github.com/cau-hai-lab/ALFWorld-Para.

Reinforcement learning from verifiable rewards (RLVR) is usually evaluated as an alternative to supervised fine-tuning (SFT), or as a stage that follows it, with the implicit assumption that the RL stage contributes gains of its own. We measure that contribution directly, in a controlled setting where both stages can be run to completion and compared exactly. We built gpu-train-debug, a deterministic benchmark of 1,877 synthetic GPU and LLM training failure diagnosis tasks across 30 archetypes, graded by a programmatic verifier with partial credit, and calibrated so that a mid-tier model scores in a 0.60 to 0.75 band. A pilot produced a surprise: 1,200 gold demonstrations nearly solved the benchmark (SFT 0.785 on the hard split), while 15 steps of GRPO from the base model did not move it (0.314 vs 0.311 base). We therefore amended the study, in writing and before the full run, to ask the sharper question: what does verifier-based RL add beyond demonstrations? At the pre-registered scale (60 GRPO steps, learning rate 1e-6), the answer is: no detectable improvement in-distribution (+0.22 points, 95% CI [-0.06, +0.51]), and a modest but measurable improvement under distribution shift (+0.33 points on a held-out OOD split, CI [+0.02, +0.68]). SFT accounted for most of the available improvement, while verifier-based RL produced only modest additional gains under distribution shift. The residual errors SFT leaves are mostly systematic, identical across samples, which likely reduces the within-group learning signal available to GRPO. A full 60-step cold-start GRPO arm confirms the pilot at the full horizon. The whole study cost $8.69 of compute and is fully reproducible from archived generations, frozen configs, and tagged code.

As large language models (LLMs) are increasingly deployed in real-world applications, evaluating the correctness of their responses is crucial for reliable use. However, such evaluation is challenging because reference answers are rarely available in practice. Existing reference-free methods mainly rely on output signals, which may assign high scores to incorrect responses and thus do not necessarily reflect correctness. This motivates exploring the model's internal computation for complementary correctness signals. Recent studies exploit internal states, but they mainly analyze hidden-state trajectories across layers, which record only the combined updates of attention and feed-forward networks (FFNs) and overlook how these components interact. We therefore propose Circe, a training-free self-evaluation method that assesses a response using both internal and external signals, without reference answers or repeated sampling. For internal signals, we first introduce component interaction consistency (CIC), which measures the directional alignment between attention updates in earlier layers and attention-induced FFN output changes in later layers. We further incorporate representation geometry (RG) to characterize normalized displacement and angular changes along layerwise hidden-state trajectories. For the external signal, we convert the mean negative log-likelihood of generated tokens into a bounded likelihood confidence (LC) score. Finally, we fuse CIC, RG, and LC into a unified correctness score. Experiments on five LLMs and six benchmarks across four domains show that our method distinguishes correct from incorrect responses more effectively than existing baselines, and generalizes to multilingual, multimodal, and open-world settings.

Keyword retrieval returns expressions of personal anxiety together with many other uses of anxiety-related language. We present a benchmark of 1,000 Arabic Twitter/X posts retrieved with the keyword qalaq ("anxiety/worry") and annotated for six discourse functions. Two native Arabic-speaking annotators labeled every post independently, reaching 51.1% raw agreement (Cohen's $\kappa = 0.378$). We then reviewed all 489 disagreements and all 511 agreements under the same guidelines. The review changed 120 of the agreed labels (23.5%), which shows that agreements need auditing as well as disagreements. The final labels form a consensus reference; on a random sample of 200 posts, an additional annotator agreed with it on 60.5% of posts ($\kappa = 0.517$). Against this reference, Fanar-C-2-27B reaches 79.8% accuracy and macro-F1 0.784 ($\kappa = 0.749$). Its errors concentrate in the residual category (41.1% of errors; precision 0.905, recall 0.447), whose definition is internally inconsistent in the executed prompt, so the low recall cannot be attributed to the model alone. Sociopolitical Anxiety (27.1%) outnumbers Personal Emotional Expression (22.0%); these are assignment proportions in the sample, not estimates of prevalence or retrieval precision. We release the guidelines, a residual-subtype scheme, a fixed data split, label provenance, and evaluation code.

Mathematical modeling turns an open-ended real-world problem into assumptions, a model, a solution, and a defended recommendation, and its answers matter only if they are both close to a credible reference and physically feasible; existing benchmarks either score solver fragments or grade reports with an LLM judge, leaving those numbers unchecked. We introduce TB-MathModel (the Terminal-style Mathematical Modeling Bench), the first benchmark to make complete contest-style modeling problems agentic, verifiable, and terminal at once, over 18 real CUMCM and MCM problems (2023--2025). Each task ships the verbatim statement, official attachments, and figures in a containerized terminal; the agent returns one JSON artifact, scored on the numeric core of the final question against a reproduction of the Outstanding-award paper. Two design choices distinguish it: direction-aware, one-sided scoring, which never penalises beating the anchor, and two-layer feasibility verification---a programmatic gate over six decidable invariant forms plus 624 Lean~4 obligations---with a hard gate that zeroes infeasible answers. Each task is scored against a panel of human outstanding papers rather than one, so a model must be close to many accepted answers; across seven multimodal entries (three seeds each) the strongest reach a multi-reference O-Eval of 64.58\%, while the \$2.54 open model holds 59.81\% at $1/100$ the cost. TB-MathModel turns the numeric core of modeling into a reproducible, machine-checkable signal for measuring long-horizon modeling competence and for collecting agent trajectories. The source code is available at https://github.com/wxj630/terminal-bench-math-modeling.

Neural theorem proving increasingly uses retrieved premises and project context to support proof generation. Evaluating these resources requires measuring the advantage of target-selected declarations over comparable alternatives and how this advantage depends on training. We introduce LeanContext, a benchmark of 129,826 Lean whole-proof generation problems, to investigate these questions. We group selected complete preceding mathematical declarations as Mathematical Context (MC) and the remaining selected local source commands as Environment Context (EC). Four aligned input views support crossed training and inference evaluations with shared targets and original Lean verification environments. For both diffusion- and autoregressive-based models, selected declarations outperform comparable alternatives by a larger margin when training includes MC. Final-input audits identify omitted declarations referenced by the target. Prioritizing these dependencies improves Pass@32 by 1.40% and 1.70% for LLaDA and Llama, respectively, on 2,000 targets. The selection experiment operates without access to target reference proofs and holds checkpoints, EC, and inference budgets constant. The code and dataset are available at https://github.com/haaaardcvZ/LeanContext.

Large language models (LLMs) are increasingly used in biomedical and cognitive neuroscience, but their ability to reason reliably over heterogeneous knowledge graphs (KGs) remains unclear. We benchmark five LLMs (GPT-5.1, Gemini 2.0 Pro, Llama 3, Qwen, and BioGPT) across five KGs: DrugBank, NeuroSynth, Cognitive Atlas, Brain Map, and GDPKG. We introduce a multitier query dataset to evaluate reasoning depth and factual accuracy under KG-aware (graph-grounded) and KG-agnostic conditions. Inter-annotator agreement was substantial, with Cohen's kappa of 0.781 (range: 0.753-0.807) and Fleiss' kappa of 0.77. Results show that general-purpose models, particularly GPT-5.1 and Gemini 2.0 Pro, generally outperform domain-specific models on multi-hop and high-level reasoning tasks, although performance declines as query complexity increases. KG grounding reduces hallucination rates across all evaluated models. Contrary to our initial hypothesis, domain-specific models do not consistently outperform general-purpose models on specialized tasks without graph grounding. These findings highlight the value of combining broad language-model capabilities with structured domain knowledge. We provide practical guidance for selecting LLM-KG integration strategies based on task complexity and computational constraints, and discuss hybrid neurosymbolic architectures for personalized medicine, cognitive disorder diagnosis, and explainable AI.

Current medical benchmarks typically break down clinical competence into isolated factual questions, single-image tasks, or static case vignettes. These tests score individual answers, but reveal less about whether a model can integrate subspecialty knowledge with multimodal evidence and track a patient's condition throughout a course of care. To close this gap, we introduce NPCBench, a clinical apprenticeship benchmark for nasopharyngeal carcinoma (NPC). NPCBench mirrors specialist training, from guideline and foundational knowledge to case planning, longitudinal review, CarePath integration, and open-ended consultation. It contains 4,777 component and consultation items and a 100-patient CarePath cohort with 417 scored diagnostic and care tasks. The episode-level evaluation tests whether diagnostic and care answers are jointly correct within a care pathway. By linking stepwise tasks to patient-level care trajectories, NPCBench evaluates whether models ground decisions in visual evidence, interpret changes in patient state, and meet clinical requirements in consultation. Across 30 evaluated models, we observe a composition gap: strong performance on individual tasks does not ensure jointly correct diagnostic and care answers for the same patient. Among 21 multimodal models, strict CarePath success reaches at most 23%. Paired controls and physician calibration further examine evidence grounding, longitudinal revision, and consultation judging. Overall, NPCBench broadens the evaluation of medical artificial intelligence beyond static question-answering paradigms to an apprenticeship-oriented assessment of clinical workflows, thereby offering a testbed for evidence-based and temporally coherent reasoning in oncology care.

ML code should run where researchers choose, not where it was first written. Today, machine learning code mainly lives in two ecosystems, PyTorch on NVIDIA GPUs and JAX on Google TPUs. Migrating repositories between them means building and testing across frameworks and hardware. This migration is (1) a practical need, since it unlocks different hardware for researchers, and (2) a natural benchmark: it involves different frameworks and hardware, covers whole repositories, and is long-horizon and challenging. We introduce XStack, a benchmark that evaluates agents on migrating ML repositories between these two ecosystems. Each agent receives a PyTorch repository and a clean TPU machine and must deliver a JAX repository. Its output passes if it meets our grading criteria, based on the source repository's published result, authenticated and calibrated by us. Migration by agents remains unreliable: Claude Opus 5 passes 75%, GPT-5.6-sol 45%, and every other agent at most 40%. XStack is a step toward freeing ML code from its ecosystem. We release XStack to make progress toward reliable migration across frameworks and hardware measurable.

Autoregressive decoding relies on the KV cache, whose memory footprint grows linearly with context length and caps batch size and throughput. Existing post-training methods exploit redundancy across tokens, layers, and feature dimensions, or lower numerical precision. We identify a complementary axis: within each attention head, values are largely predictable from their corresponding keys. We exploit this with value cache amortisation, which fits a lightweight per-head predictor at test time and stores residuals only for the tokens it reconstructs worst, giving $4\times$ near-lossless value compression. The predictor can be initialised by meta-learning for substantially reduced prefill overhead, or fitted from scratch in settings that allow a one-off cost, such as multi-query inference. We pair value amortisation with cross-layer low-rank key compression to produce LowRAM, an asymmetric compression pipeline that stays close to lossless across three model families. By reconstructing only a selected subset of KV pairs at each decoding step, LowRAM can keep decoding overhead minimal and achieve up to $5\times$ higher peak decode throughput at $4\times$ compression. LowRAM further composes with quantisation and eviction, reaching $26.9\times$ compression at $94.5\%$ of baseline performance on long contexts. These results show that treating keys and values as structurally distinct enables efficient, aggressive, and composable KV cache compression.

Once an agent is live, teams want to know whether it is doing its job, not just whether it throws errors. Yet most evaluation is one-off: a benchmark run, a look at traces, a single score. Its conclusions neither check themselves nor reach the agent. We propose closed-loop evaluation, an evaluation method for deployed agents. It models evaluation as performance estimation under biased raters and censored outcomes, and it requires every level's conclusion to trace back to atomic data and to be computed by versioned rules. The method covers four levels of analysis (span, trace, conversation and engagement), measures along a spine of Execution, Substance, Trajectory and Results, treats Conduct as a constraint throughout, and uses Provenance and Adjudication components to certify the evidence and the evaluators. Evaluator agents generated by an Agent Factory carry these components, each with a domain knowledge pack and an adapter harness. Two loops close the evaluation: an inner loop audits the evidence, rules and metrics themselves, and an outer loop uses per-component regression checks to drive agent improvement. We also give a pipeline that runs this process automatically. Using the business-analysis agent SharpInsight as a case and porting the method to public conversation data, we show how problems differ across levels and how evidence-and-rule judgment is more reliable than holistic scoring (ICC 0.86 vs. 0.81; kappa 0.81 vs. 0.63).

At context lengths of about 1M tokens, dense attention is costly in prefill and decoding, and training typically relies on long documents, which become scarcer as the target length increases. Most sparse attention methods score all preceding blocks or tokens, so their selection cost per query is $O(L)$ at context length $L$; hierarchical selection reduces it to $O(\log L)$, but prior trainable hierarchical methods report accuracy only within their training length. We propose RecSA (Recursive Sparse Attention), which builds a tree over the KV cache in which each node summarizes its subtree with one key and one mean-pooled value. The keys come from one pooler per layer, shared by all levels and distilled from dense attention. Each level above the tokens selects at most $k$ candidates, so each query has $O(\log L)$ selection cost and reads $O(\log L)$ entries; the summaries of unselected candidates share one softmax with the exact tokens. Each entry's RoPE position is its index among the candidates of its level, so at any length every relative distance stays within the range seen during training. Initialized from Qwen3-4B and trained on 0.62B tokens at the original model's 32K training length, RecSA outperforms all baselines on RULER at every length from 131K to 2M; at 1M, it scores 60.5, compared with 40.0 for the best baseline, Dense + DCA. With our optimized kernels, RecSA is faster than dense attention in end-to-end prefill and decoding from 64K to 2M; at 524K on one H200 GPU, it accelerates prefill by 3.2x and decoding by 3.8x.

Personal AI agents face social engineering attacks supported by forged multimodal content, but existing evaluations mainly focus on text-based dialogue, leaving their intrinsic security under such attacks insufficiently studied. We introduce Multimodal-SEA, an interactive benchmark comprising 608 expert-designed tasks grounded in real-world fraud cases, covering 14 domains and 152 attack scenarios with prepared multimodal content. The tasks span four difficulty levels and reflect the distinct social and cultural contexts of Chinese and English environments. Agents inspect content, verify requests, and execute actions in simulated applications, with attack outcomes grounded in verifiable tool execution logs. To enable controlled red teaming of personal agents, we design a multimodal attack harness that integrates task information, attack progression control, and multimodal content construction to evaluate their intrinsic security against adaptive social engineering attacks in simulated environments. Evaluating seven widely used multimodal models across six dimensions reveals that all are vulnerable, with attack success rates of 7.1% to 34.5%. Adding multimodal content increases attack success from 19.9% to 29.2%, despite increased proactive verification coverage. Stronger safety instructions reduce attack success without improving recognition of deceptive multimodal content, while lower attack success rates for some models accompany excessive refusal of legitimate requests.

Self-consistency (SC) is a widely used decoding strategy for improving chain-of-thought reasoning, but it requires sampling many complete reasoning trajectories with a preset budget. Its efficient variants reduce the cost mainly by stopping full-trajectory sampling early or distributing the number of samples across questions. In this paper, we propose StepBack, a step-level error-localized resampling method for efficient test-time reasoning. Given a small set of draft trajectories, StepBack localizes likely first error positions using lightweight signals and then resamples the detected error step and its suffix, followed by majority answer voting. Empirically, StepBack matches the accuracy of SC on average across seven mathematical reasoning and multi-hop QA benchmarks, while reducing token usage by up to 41.26%. In addition, by nature, StepBack can be combined with adaptive inference methods such as Adaptive Consistency, further improving their inference-time efficiency.

Coding agents are among the most economically consequential deployments of LLMs to date, making reliable evaluation increasingly important. Yet even benchmarks that are carefully validated at release can become stale. As models improve, they can discover valid solutions that existing verifiers reject, while changes to evaluation infrastructure can invalidate assumptions embedded in tasks and tests. We call the resulting accumulation of defects benchmark maintenance debt: a growing gap between benchmark scores and the capabilities they are intended to measure. We introduce a continuous-validation workflow for agentic benchmarks that combines community reports, task artifacts, agent traces, automated oracle checks, and human-gated repairs. We apply this workflow to Terminal-Bench 2.0, a benchmark that underwent extensive human and LM-assisted review and red-teaming before release. Our audit identifies issues in roughly 30% of the benchmark, and we release the resulting repairs as Terminal-Bench 2.1. These repairs change evaluation results: accuracy increased by over 10 percentage points and leaderboard positions move by up to three ranks, even as the overall ranking remains highly correlated (Spearman rho = 0.958). We evaluate how resource allocation shapes benchmark conclusions. We show that fixed wall-clock can materially change agent performance: For example, extending the time limit closes a 20pp harness gap on GPT 5.4 and shrinks a 30 pp gap on GPT 5.4 mini to 5 pp. Benchmark maintenance is therefore not only about repairing task artifacts: resource budgets are part of the measurement specification and must be validated alongside the tasks themselves. More broadly, benchmark maintenance debt is not just accumulated implementation bugs: our results show that as models improve, both evaluation assumptions and resource settings may need to be revalidated.

Reproducing machine learning experiments requires reconstructing procedures from information scattered across papers, released code, and supporting artifacts. For coding agents, reproduction success can therefore depend on their ability to fill in missing details, rather than on the released artifacts alone. We ask whether preserving recovered procedures and decisions improves subsequent independent reproduction. We define a standardized artifact design comprising a refined codebase, environment setup, experiment launchers, and a reproduction guide, and introduce ReproBridge, an agent workflow for constructing these refined reproducible artifacts. ReproBridge-Bench provides 1,029 tasks from 387 experiments across 36 papers, anchored by five human-verified reference scores per task. Agents attempt tasks in fresh sessions with raw code or refined artifacts; a separate Auditor checks execution evidence. Across four evaluated models, refined artifacts raise mean reproduction success from 76.4% to 92.6% and reduce downstream API spending by 19.7% on average. The cross-model success-rate spread narrows from 21.6 to 3.9 percentage points, with all four models achieving mean success rates above 90%. Removing any component reduces success in the ablation study, with the largest losses from code and environment removal. These results suggest that preserving execution knowledge helps even models with lower raw-code success reproduce experiments more reliably and at lower API cost.

A robot generalist must interpret a scene, act on it, and satisfy every condition of a task. How far do frontier vision-language models (VLMs) meet these demands? We introduce Embodied Agent Arena, evaluating ten VLM agents across Geometry, Spatial Reasoning, Affordance, Task Planning, and Manipulation. Its 1,000 cases combine 32 established sources with GeoProbe, our 168-case benchmark for camera, motion, depth, and scale estimation. A minimal harness leaves task decisions to the model through source-native interfaces. Closed-weight models lead all five domains, whereas open-weight VLMs show useful spatial reasoning but substantially lower planning and manipulation success. Beneath these averages, similar geometric errors conceal different scale and shape distortions, and valid region estimates need not identify usable action points. During execution, agents can achieve early subgoals yet lose object control, bind an action to the wrong receiving region, or stop before the required physical state holds. Additional computation does not uniformly resolve these failures: code repair can restore execution without completing the task, while self-review can reverse correct judgments. Failure costs also differ: the strongest planners concentrate failed-token expenditure in prolonged body-control tasks, whereas other models fail broadly across procedures. Matched traces show how successful recovery uses feedback to restore a missing prerequisite and advance the remaining task.

Post-training quantization (PTQ) reduces large language model (LLM) deployment costs, but preserving quality requires accounting for uneven weight sensitivity. Uniform PTQ uses one weight bit-width, importance-aware PTQ adjusts scales or reconstructs weights within that width, and column-wise mixed precision retains selected columns at higher precision. However, sensitivity varies within columns, so whole-column retention also protects less sensitive weights, while runtime outlier detection adds decoding overhead. We therefore introduce TileQuant, a training-free framework that assigns precision to hardware-aligned 2D weight tiles. Offline calibration provides input-activation-aware scoring from forward statistics for lower calibration cost and gradient-informed scoring from loss sensitivity for stronger quality preservation. Within each projection's retention budget, the highest-scoring tiles remain at 16-bit and the rest use 8-bit. The resulting map guides reconstruction into compact weight buffers, which the kernel reuses for native integer and 16-bit computation with local accumulation to reduce memory traffic. Experiments show that TileQuant reaches 2.66x the original model's generation throughput. Across language modeling, reasoning, long-context retrieval, and question answering, it preserves quality and surpasses quantized and original models on multiple tasks. These results support selecting precision units that capture local sensitivity while enabling efficient hardware execution. The implementation is: https://github.com/HanzhiZhang-Ulrica/TileQuant

An LLM agent's cost and success depend on its model, its harness, and its environment. In addition to the widely studied directions of parameter tuning and harness optimization, we propose the third direction: optimization over stationary environments. Agents are often deployed in rich environments shared by many tasks, such as a company's Slack workspace or a complex collection of data sources, documents or tickets. We call these environments stationary environments. We introduce EnvOpt, a method for automated end-to-end optimization of stationary environments for any given model and harness, using real or synthetic tasks without needing any graders or reference answers. On AutomationBench, EnvOpt reduces the cost of Claude Haiku 4.5 by 72.7% and raises its pass rate by 14.34%. The same environment transfers to Opus 5, cutting its cost by 68.1% with a 1.34% gain. Added on top of Meta-Harness's optimized harness for Haiku 4.5, it lowers cost by a further 83.4% and adds 11.17% accuracy over Meta-Harness alone. Across 5 benchmarks, EnvOpt significantly reduces the cost of agents while improving or maintaining pass rate, and is transferable to different models and different harnesses. We also discuss common successful and rejected environment optimization strategies during automated optimization. More broadly, our results show that how information is represented to an agent shapes both its cost and its success, making the environment worth optimizing in addition to the parameters and the harness.

Linear attention reduces the quadratic sequence-length complexity of softmax attention by maintaining fixed-size recurrent states. However, these states scale quadratically with the representation dimension and require repeated data transfers during decoding, creating a substantial GPU memory bottleneck. Recent studies have revealed significant low-rank redundancy in recurrent states, suggesting opportunities for more efficient representations. We introduce SPARE (State Projection for Accelerated Recurrent Execution), a training-free method that exploits this redundancy through fixed low-dimensional subspaces, heterogeneous rank allocation, and lightweight write compensation. By performing recurrent reads and updates directly on compact states, SPARE reduces memory traffic without online decomposition or dense-state reconstruction. SPARE achieves 2.55x full-model decoding speedup, 7.49x lower recurrent-kernel DRAM traffic, and 3.31x lower peak GPU memory, while preserving better associative-retrieval accuracy than structured pruning at matched memory budgets.

Drug discovery requires continuous scientific decision-making across a vast research space. We introduce PharmAgents2, an autonomous system for end-to-end drug discovery that combines knowledge-enhanced decision-making with organized large-scale exploration. Paper2Skill extracts transferable scientific decision knowledge from public sources and, together with expert-developed skills, supports research decisions. Self-Scaling Discovery expands exploration through parallel agents inheriting research context and decision-making capabilities across targets, scientific hypotheses, and discovery strategies. We introduce PharmAgents-Bench v0.1, comprising 121 questions across nine decision scenarios and three end-to-end tasks assessed with expert-defined process rubrics. PharmAgents2 outperforms evaluated baselines in most decision categories and achieves the highest process scores on all three end-to-end tasks. In three real-world studies, autonomous investigations generated findings subsequently supported by human-executed experiments. For LpxH, PharmAgents2 hypothesized a previously unreported allosteric site and screened compounds against it, identifying a novel inhibitor scaffold with experimentally confirmed inhibitory activity and binding. For JAK1, it prioritized a covalent inhibition strategy and nominated compounds with experimentally demonstrated activity. Starting from a known bioactive molecule, Self-Scaling Discovery explored candidate targets across the proteome and identified a new binding target, with the interaction subsequently confirmed experimentally. PharmAgents2 represents a prototype of a virtual pharmaceutical company, marking a step toward autonomous end-to-end drug discovery.

Diffusion and flow models have recently been extended to language modeling, generating text through continuous denoising in token embedding spaces. Pretrained models provide token representations and contextual prediction capabilities, but improving generation with limited additional training requires an appropriate connection between supervision and continuous generation. Continuous MSE evaluates a predicted vector obtained by weighting token embeddings with predicted probabilities; different probability distributions can yield the same vector and are therefore indistinguishable under this objective. We propose Token Identity Brier Flow (TIBF), which applies squared supervision directly to token identities when continuing pretrained language flows. TIBF combines network scores with token-matching scores under a Gaussian observation model, supervises token probabilities using the Brier loss, and uses probability-weighted embeddings for continuous generation. With fixed token embeddings, we show that the squared error between predicted token probabilities and the true posterior bounds the squared errors of continuous denoising predictions and flow velocities. A risk decomposition separates uncertainty inherent in the observation from posterior fitting error, motivating continued training that preserves pretrained embeddings and focuses on probability prediction. We also establish convergence to token embeddings under ideal predictions and sufficient conditions for local convergence with prediction error. From common pretrained initializations and under matched continuation budgets, frozen-embedding TIBF achieves Gen.PPL values of 33.96 on OWT and 81.17 on LM1B, outperforming continuous-MSE and CE continuation controls. It attains the lowest Gen.PPL at every tested sampling budget on both datasets, together with the lowest repetition rate on LM1B.

Deep research agents answer complex questions through repeated search. They rely on a harness that runs their tool calls and returns evidence to their context. This creates a trade-off: Search works best over long passages, which give enough context for reliable semantic matching. But the returned evidence should be short, because every observation stays in the agent's context for the rest of the trajectory. We introduce CORE, a training-free harness that combines coarse-to-fine search with context management. CORE retrieves passages with a dense retriever. It then reuses the token-level similarities from a late-interaction reranker to return only the most relevant snippet from each passage. Context management lets the agent expand the context around sources it has already retrieved and suppresses repeated snippets. In controlled closed-search comparisons with the Tongyi DeepResearch and OpenResearcher harnesses, CORE achieves state-of-the-art answer accuracy and evidence recall across three agent models and three benchmarks (BrowseComp-Plus, BrowseComp-Wikipedia and BrowseComp-Plus-CM), on corpora containing up to 553M documents.

On-policy distillation (OPD) improves language-model reasoning through dense teacher feedback on student-generated responses.We first observe an amplitude disparity in this feedback: on math reasoning responses, style-related positions exhibit 3.5 times the mean absolute teacher--student log-probability difference of mathematical positions.Scaling OPD updates by these discrepancies thus directs much of the update strength toward matching the teacher's expression. We instead take the student's own pretrained base as a reference: the student--base difference records the token preferences acquired during post-training and provides supervision that does not depend on the student's distance from the teacher. Building on this reference, we propose Student-Delta On-Policy Distillation (SD-OPD), which uses centered student--base differences as update coefficients, keeps each coefficient only when its sign agrees with centered teacher--student feedback, and recomputes the student signal every training round. Across 5 teacher--student configurations with Qwen3 students of 1.7B to 8B parameters, SD-OPD improves six-task mathematics average pass@1 (avg@16) over OPD by 7.76 to 10.37 percentage points. Code is available at https://github.com/aim-uofa/sd-opd.

On-policy distillation (OPD) supplies dense token-level supervision by aligning a student to the teacher's next-token distribution on its own generated rollouts. Existing works mostly implement top-$k$ OPD, applying reverse Kullback-Leibler ($\RKL$) divergence between the teacher's and the student's renormalized distributions over teacher top-$k$ support. However, the teacher scores student-generated rollouts that are inherently off-policy for itself, leaving the reliability of its next-token guidance unclear. Restricting the divergence to the teacher top-$k$ support also discards the remaining mass outside that support. These gaps leave two questions implicit: 1) which tokens to align, and 2) how to align those tokens. To answer these questions, we propose Overlap-gated On-Policy Distillation (O$^{2}$PD). As for 1), we quantitatively analyze the reliability of teacher supervision and find that alignment on the student's trajectory prefix is more reliable than alignment on the full context. Therefore, our O$^{2}$PD is applied only to the 20\% prefix tokens of each student rollout. As for 2), we theoretically characterize two problems of top-$k$ OPD: the \emph{Teacher-support Uncovering} and the \emph{Dominant-Mode Mismatch}. Based on these analyses, we propose two support strategies: the \emph{Reset Support} strategy attaches a reset atom to every token, pooling mass outside the teacher top-$k$, and the \emph{Support Convergence} strategy converges the support to the already aligned intersection. Our O$^{2}$PD employs the designed overlap-gated $\alpha$-divergence to interpolate between mode-covering and mode-seeking for per-token alignment, where the overlap is used as a gate to switch the support and adaptively modulate $\alpha$.

Diffusion Large Language Models (dLLMs) commonly decode responses by proceeding autoregressively across blocks while iteratively denoising multiple tokens within each block. This structure makes block-level policy optimization a natural adaptation of autoregressive policy optimization. However, the exact likelihood of a completed block is intractable because it marginalizes over the intermediate states and reveal orders of the denoising process. To realize the block-level formulation, we construct a prefix-block state and use a single model evaluation to define a factorized one-step surrogate for the block likelihood. The resulting block surrogate ratio is a product of its token surrogate ratios, allowing variation at individual positions to propagate across the entire block before clipping. We therefore introduce Prefix-Block Policy Optimization (PBPO), which preserves block-level conditioning but applies surrogate importance weighting and clipping separately to each token. PBPO further diversifies repeated updates by resampling corrupted views of the prompt and revealed prefix, a design motivated by masked-dLLM pre-training. Under our analysis assumptions, block-product weighting amplifies the variance of gradient contributions relative to token-local weighting. Across two LLaDA-family initializations, six benchmarks, and three generation lengths, PBPO achieves competitive performance against existing dLLM RL methods.

LLM-based recommender systems increasingly generate chain-of-thought (CoT) text before each item prediction to activate the LLM's reasoning capacity, costing tens to hundreds of extra tokens per recommendation at inference. We ask whether inference-time reasoning artifacts (textual or latent) can be replaced by training-time supervision alone, without loss of accuracy. We propose LatentRec, a framework for LLM-based generative recommendation that replaces inference-time reasoning artifacts with training-time per-step head supervision. LatentRec has three components. A 2.5M-parameter Continuous Thought Module iterates K latent reasoning steps over the LLM's hidden state during training. Soft Token Selection bridges the ~14x norm gap between hidden states and input embeddings. A Per-step Reasoning Loss (PRL) supervises every intermediate state through the LLM's prediction head against the first token of the target Semantic ID. The trained LLM retains this accuracy without the module at inference, so the module can be discarded at deployment. On three Amazon benchmarks, LatentRec improves HR@10 by 4.0-5.0% (relative) over a CoT-distilled baseline at 18.3x lower latency (batch=1). Per-step training supervision consistently improves accuracy across three datasets, paid once at training rather than at every recommendation.

Frontier artificial intelligence (AI) systems are increasingly operating over longer horizons with diminishing human oversight, raising the risk of loss of control: systems acting beyond effective human control with potentially irreversible consequences. Because such failures may become observable only after containment is difficult, they must be assessed prospectively. We introduce a behavioural framework that decomposes loss-of-control risk into three factors: misaligned motive, harm-enabling capability, and monitoring evasion. We operationalise these factors through 13 measurable aspects and 29 evaluation tasks, whose scores are combined into a single loss-of-control (LoC) score. Across 13 frontier models, the LoC score shows strong prospective validity, correlating with failure rates in contained long-horizon risk scenarios (Pearson r = 0.84). Risk varies systematically across models and cannot be explained by capability alone. We further show that the framework can guide targeted mitigations: fine-tuning and monitoring directed at diagnosed vulnerabilities reduce failure rates while largely preserving general performance. These results provide a practical basis for predicting, understanding, and mitigating loss-of-control risk before it materialises.

DeepSearch LLM agents repeatedly search and carry retrieved sources into subsequent reasoning steps. This expanding context increases token consumption and latency, although many results do not contribute to later reasoning. We propose Query-Relevant Attention Entropy (QRAE), an online context compression method that uses sparse evidence-assessment heads to decide which intact retrieved chunks should persist across reasoning steps. It scores bounded candidate batches through paired prefills: a query-free control suppresses shared attention responses, while normalized entropy measures the breadth of the positive response within each chunk. High-scoring chunks enter a persistent evidence whitelist and are included in the reasoning context. Uncertain chunks remain outside the reasoning context but are retained in a candidate cache, where they can be retrieved for later reevaluation. QRAE requires neither an additional generative compression pass nor modification of the attention operator. Across multiple DeepSearch benchmarks and reasoning models, QRAE improves task success by an average of 2.35 percentage points and reduces end-to-end latency and reported token usage by an average of 30% and 45%, respectively. These benefits generalize to a model with nearly 100B parameters.

Large language models have made rapid progress in agentic reasoning and tool use, raising expectations for agents that can take over realistic productivity work from humans. However, current agents still struggle to finish such tasks, where success requires following long-horizon professional workflows, understanding heterogeneous file contexts, and producing usable deliverables under fine-grained requirements. A key bottleneck is the lack of scalable post-training data that reflects real work intent, authentic file environments, and deliverable-level evaluation. To close this gap, we introduce Agent-R$^2$, a $\textbf{r}$eal-source data synthesis framework that scales training data for $\textbf{r}$ealistic productivity work. Instead of relying on manually specified workflows or LLM-generated files, Agent-R$^2$ grounds synthesis throughout the pipeline in real occupational seeds, corpus-derived workflow evidence, and authentic web-sourced files. From 10k web-corpus candidates, Agent-R$^2$ distills 4k task instances that pair real multi-format attachments with fine-grained, deliverable-level rubrics. On this data, we train Agent-R$^2$-4B via supervised fine-tuning and rubric-based reinforcement learning, improving both end-to-end productivity performance and file-grounded reasoning. Trained on only 4k instances, Agent-R$^2$-4B surpasses Qwen3-235B-A22B-Thinking-2507 on both GDPval (51.0$\%$ vs. 48.0$\%$) and OfficeQA (14.8$\%$ vs. 13.8$\%$). Further analysis shows that Agent-R$^2$ mitigates shortcut behavior, where agents produce superficially plausible deliverables before sufficiently grounding, constructing, or verifying them, while promoting sustained exploration.

Masked diffusion language models offer parallel text generation, but independently sampling multiple tokens can disrupt their dependencies and change the output distribution. Exact continuous-time sampling avoids this approximation, yet typically requires one sequential model evaluation per token. We introduce SpecCTMC, a training-free speculative sampler that reduces this sequential bottleneck without a separate draft model. SpecCTMC proposes successive token updates using previously computed model predictions, evaluates their intermediate states in parallel, and applies an exact acceptance-and-correction procedure. We prove that SpecCTMC preserves the joint distribution of event times, generation order, and token values of the underlying continuous-time Markov chain. It therefore introduces neither time-discretization error nor additional factorization error from parallel decoding. Experiments on LLaDA-8B-Instruct and Dream-v0-Instruct-7B across five benchmarks demonstrate 1.73-4.38x fewer sequential model evaluations than direct continuous-time sampling, while preserving the reference sampling distribution. Our code is available at: https://github.com/kasakh/specctmc

Linear attention compresses context into a fixed-size key-value state for linear-time training and constant-memory inference. Gated delta-rule models decay this state and erase its prediction for an incoming key before writing a new value; Gated DeltaNet-2 (GDN-2) generalizes this with channel-wise decay and decoupled, full-rank erase and write gates. However, we show that state updates are far less sensitive to erasure than to writes: key activation and unit normalization concentrate key energy onto a sparse subset of channels, whereas value writes scale all channels, so the erase gate can be parameterized at lower rank. Furthermore, slow-decaying state rows accumulate much larger norms than fast rows, dominating unweighted readout and forcing stronger write-time erasure to suppress interference. Guided by this analysis, we introduce Muir, which fuses all gates into a sensitivity-allocated low-rank projection, re-weights state rows with a matched-filter readout gate, and reinvests the saved parameters in a wider feed-forward network. Because standard chunkwise algorithms become numerically unstable at the large chunk sizes needed to saturate wide matrix units, we design a parallel update for stable, efficient training. Across 166M-1.34B parameters, Muir lowers perplexity by 1.8-2.0% over GDN-2 and 2.5-3.6% over Transformer, raises commonsense accuracy by up to 1.9 points, and improves extractive recall over GDN-2. At 1.34B, Muir outperforms both baselines on multi-hop and multi-needle recall, extrapolates 2,560x from its 4K training length to 10M-token streams without perplexity drift, and maintains near-100% passkey retrieval to 1M tokens with static readout. Muir achieves these gains while matching GDN-2 throughput, reaching its 100B-token validation loss 1.2x faster, and training 2.7x faster than Transformer at 32K.