2026-10-23

Model fusion aims to integrate the capabilities of multi-expert models into a single model. Existing evaluations offer limited coverage of reasoning with long chains of thought and tool use, while strong average performance can mask substantial losses in individual expert capabilities. We introduce FusionEval, a benchmark that provides controlled expert pools, integrates fusion methods with and without additional training, and offers a unified evaluation suite. The pools contain domain experts independently trained from Qwen3 and Llama base models across math, code, science, instruction following, and tool use. The suite evaluates capability retention and computational cost on eight public benchmarks under comparable data access rules, reporting normalized average and worst performance, runtime, and peak GPU memory usage. Extensive experiments show that training-free methods preserve reasoning performance but substantially degrade tool use at both model scales. In contrast, MOPD preserves tool use and achieves a more balanced capability profile, but requires additional data preparation, inference, and training. Jointly retaining reasoning, instruction following, and tool use at low computational cost remains a challenge. We hope FusionEval supports future research on efficient model fusion that jointly preserves expert capabilities. Our code is available at https://github.com/Baicaihaochi/FusionEval_Benchmark.

Large language models (LLMs) are widely used to answer factual questions, but their answers can contain errors and must be fact-checked claim by claim. Yet the resulting score depends on three design decisions: which questions the claims answer, under what conditions each claim is judged, and which sources count as evidence. No existing benchmark specifies all three. We introduce JSFactCheckBench, a Japanese benchmark that specifies all three, with 1,913 claims from LLM answers to questions about misleading information that circulated on Japanese social media, each annotated for factuality at a reference time and for importance to the question, using evidence from preselected Web sources. Across six LLMs, all verify claims directly answering the question less accurately than supplementary claims, a gap that aggregate scores obscure. Scores also drop when models must locate the relevant content in longer or multiple sources and, for all four open models, when the surrounding question and response are added. These weaknesses are visible only because the benchmark records the conditions behind each label. A benchmark that leaves them implicit cannot reveal them.

Tool-using language model agents rely on environmental facts to decide whether an action is permitted. When a tool response omits a safety-relevant fact, an agent may carry out a prohibited action even though the user request is benign and every returned statement is true. We introduce \bench{}, an executable benchmark for evaluating unsafe execution, safety boundaries, and evidence recovery in LLM agents under missing evidence. It contains 300 matched unsafe/safe task pairs across six domains. By systematically suppressing critical facts, the benchmark measures unsafe execution under critical evidence omission and determines the minimum observed failure cardinality, including failures with full evidence and cases with no observed failure after exhaustive testing. An independent audit tool tests whether agents seek missing facts and then withhold prohibited actions while completing permissible tasks. The benchmark reveals substantial variation across the ten models in omission robustness, evidence acquisition, and appropriate action. Resistance to missing evidence does not necessarily translate into useful recovery. Recovery access reduces unsafe execution for all ten models, but agents differ in whether they actively obtain the missing facts and use them to act appropriately in both worlds.

As LLMs expand their language coverage, evaluating how well they understand and represent diverse cultures becomes increasingly important. However, many benchmarks remain English-centric, with limited cultural coverage and a focus on factual knowledge over its application in real-world cultural contexts. We propose CulturalRiddles, a benchmark of original, culturally grounded riddles created entirely by humans through a participatory community effort, spanning 62 cultures (45 countries, 51 languages). These riddles are deeply rooted in their respective cultures and require abductive and compositional reasoning, alongside detailed factual and social knowledge, language skills, and an understanding of culturally specific associations. We benchmark 28 LLM configurations using both automatic and large-scale human evaluation. Even frontier models achieve at most 22% accuracy in some communities. We publicly release all artifacts alongside an open contribution process to expand the benchmark's linguistic and cultural coverage.

Quantitative research is an iterative loop: a researcher proposes a market hypothesis, implements it as an executable signal, tests it on historical data, and decides what to keep. Existing benchmarks score formulas, code, or individual trading decisions, leaving open whether large language models (LLMs) can complete this loop and produce strategies that work beyond the research period. We introduce Quant-Arena, a benchmark and public leaderboard for end-to-end quantitative research on real A-share market data. A model constructs features and stock-scoring code, receives feedback only on 2022--2023, and freezes one strategy for assessment on the sealed year 2024. A cash-account evaluator applies trading constraints and costs, including market impact estimated from a simulated order book. Scores are dominated by realized excess return and complemented by signal quality and the diversity of features used by the strategy. We evaluate nine frontier models, among which GPT-6 Astra scores highest; yet no model beats the equal-weight market in 2024. Quant-Arena is open to submissions from any model or agent harness.

Many Indian banks serve customers through FAQ pages and scripted chatbots. Assistants based on language models could respond faster and carry out requests such as blocking a card, but only if they give correct information and follow the bank's rules. These rules prescribe a procedure, a set of required steps, such as obtaining a closure quote before closing a deposit. Adapting models to this domain requires measuring how well they follow the procedure, before and after training. Benchmarks usually judge only the records left at the end of a conversation. Many required steps, however, leave the records unchanged, so an assistant that skips them can still be judged successful. We present TauIndianBankBench, an action-oriented benchmark of 1,000 Indian retail-banking tasks where an assistant helps a simulated customer. Besides the final records, it checks that the assistant took every required step and made no unrequested change. For every model, 6.5-12.0% of sessions are judged successful when only the final records are checked, although the assistant skipped a required step or made an unrequested change. Scoring both procedure and outcome shows more accurately how well language models perform in Indian retail banking.

LLM-based chat agents are increasingly used for complex, multi-turn, multilingual medical triage in low- and middle-income countries (LMICs), yet evaluation methods focus on single-turn, English-only benchmarks that primarily test the chatbots' knowledge. In practice, users are often confused, overwhelmed, or emotionally vulnerable, giving unclear, misspelled single-word answers or long messages in regional dialects, and the context of a conversation can shift rapidly. We propose SIM-UT (Simulation of User Behaviour for Triage), a multi-turn, simulation-based evaluation method that identifies ecologically valid, contextually relevant failure modes of LLM-based triage chat agents. Using SIM-UT, we evaluated three real-world WhatsApp-based health chatbots: a respiratory screening programme in India, a hearing health support programme in South Africa, and a diabetes and lifestyle coaching agent in Kenya. We compared a medically fine-tuned open model (MedGemma) against general-purpose frontier models (ChatGPT and Claude) across 5,000+ multilingual simulated conversation turns grounded in the personas, constraints, and communication patterns of the real user populations, over 2-4 iterative rounds. A key failure mode emerged in both model families: de-escalation risk. The agents initially identified medical emergencies correctly, but when the user minimised the symptoms without introducing new clinical information, they de-escalated, exhibiting clinically dangerous sycophancy, for example agreeing that vaginal bleeding in pregnancy was "not a problem" or that decreased fetal movements at 32 weeks "can be normal". We recommend user simulation-based behavioural evaluation, grounded in real user context and covering adversarial and edge cases, as standard practice for triage chatbots in production, so that safeguards can be added before deployment or in subsequent iterations.

Personal intelligence systems are taking on increasingly complex everyday tasks, yet repeated questions, unnecessary confirmations, and unwanted outreach can turn capable assistance into interaction burden. Reducing this burden requires agents to remember how users want to be assisted. However, existing memory benchmarks focus on factual recall and personalized answers and rarely test how agents learn and act on interaction preferences. We introduce MemPABench, an interactive environment for on-policy evaluation of memory in personal agents. MemPABench poses long-horizon tasks requiring agents to decide when to clarify, confirm, or act autonomously. Four grounded simulated users hold 14 preferences across work and personal contexts. Their feedback spans 99 sessions over four months, with five event-driven changes per user testing memory updates. We evaluate preference selection, delivered behavior, and sensitivity to context and preference change. Evaluating four memory systems across four LLM backbones, we find that current memory brings limited gains in preference selection. Retrieval-based systems (i.e. Mem0, MemOS, simple RAG) do not outperform No Memory, while File-based memory improves accuracy but applies preferences beyond relevant contexts. Our analysis shows that current memory struggles to retrieve interaction preferences and adapt to preference changes. Even with the correct preference available, overriding the agent's default interaction habits remains challenging. The environment and dataset are available at https://mempabench.com.

Industry Foundation Classes (IFC), the standard for building information exchange in architecture, engineering, and construction (AEC), is increasingly queried through large language models (LLMs) using natural language. Evaluating these systems at scale is difficult because free-text answers require costly human or LLM-based review. We propose IFC-MCQ, a method that automatically generates multiple-choice benchmarks from IFC files for exact option-set grading. The taxonomy covers 15 categories and 36 subtypes across four groups, with answer keys derived from the file and distractors constructed from file-derived values or predefined rules. We applied IFC-MCQ to IFC models of varying sizes and modelling characteristics to assess generation coverage and evaluated two agentic configurations on selected models using tool- and code-based retrieval. The evaluation exposed category-specific failure modes and showed that performance can vary with the source IFC model and retrieval interface. IFC-MCQ enables scalable benchmarking of LLM-based IFC retrieval in AEC workflows.

Frontier language models now score above 90% on popular knowledge benchmarks and are increasingly deployed as agents that carry out multi-hour professional work. Yet no existing benchmark tests whether they can perform the real-world, economically valuable work of a professional physical engineer, such as modelling physics and legal code requirements across a multi-format project corpus, optimising material and installation costs, and issuing 2D/3D engineering design packages that a licensed reviewer would sign off on. Existing engineering benchmarks focus on textbook question answering, isolated tool calls or self-contained toy design problems. As a result, most of these benchmarks are already saturated by leading agents and solving them is far away from delivering real engineering design. We introduce PLE, a benchmark of 85 long-horizon, agentic tasks in professional physical engineering consisting of electrical, civil, structural and mechanical engineering, where solving a task means completing a realistic, billable engineering deliverable. PLE combines two complementary curation paths. Expert-Authored tasks are created end to end by practising engineers from a professional data provider's expert pool. Co-Authored tasks start from problem statements written by practising engineers; an adversarial Creator-Solver agent pipeline constructs each task's realistic project environment, executable grader and reference deliverables before sending them to expert engineers for validation and difficulty intensification. We evaluate a variety of frontier LLMs under seven publicly available agent harnesses. Even the strongest configuration reaches a mean score of only 50.6%, while spending 27M tokens and 63 minutes per task on average, demonstrating the substantial gap between current agents and autonomous physical engineering design.

Unlike conventional GUI agents, personalized GUI agents require user preferences to guide sequential execution over changing GUI states. This raises two challenges: determining HOW each preference should alter execution strategy and WHEN it should take effect along the interaction trajectory. Existing benchmarks primarily evaluate overall task success, offering limited insight into these questions. We introduce PrefDiag-Bench with two components: (1) a categorization of preferences into five types, characterizing HOW they affect execution; and (2) a stage-wise diagnostic framework with 1,664 probes, tracing WHEN preferences should take effect and where failures occur. We evaluate eleven GUI agent systems, including end-to-end agents and agentic workflows, under two forms of user context: explicit user profiles and behavioral memory logs. Although existing agents achieve relatively high completion rates on non-personalized tasks, performance is substantially lower on preference-personalized tasks. Even the best-performing model, Seed-2.0-Pro, achieves only 47.17% on Process preferences under explicit user profiles. Agents struggle particularly with Ranking and Process preferences. Further analysis reveals two key capability bottlenecks: selecting relevant evidence from user history and recognizing when preferences should take effect during execution. Controlled interventions that explicitly provide relevant evidence improve end-to-end task success by up to 7.06%. PrefDiag-Bench offers a more systematic evaluation of GUI agents' personalization capabilities beyond overall task success. Code and data are available at https://github.com/Zhixin-L/PrefDiag-Bench.

LLMs are increasingly used as AI tutors, yet existing benchmarks overlook the challenging behaviors common among students with learning and behavioral difficulties. We introduce TutorCaseBench, a benchmark that grounds AI-tutor evaluation in qualitative case studies from pedagogy, educational psychology, and special education. Our pipeline converts documented student categories into case-specific scenarios, tutor-student dialogues, and rubrics grounded in source evidence. The benchmark contains 923 dialogues across five challenge types, evaluated through open-ended response generation (1,560 items) and multiple-choice selection (2,537 items). Across 11 state-of-the-art LLMs, even the strongest model satisfies a rubric criterion in only 55.8% of open-ended cases, while models are better at selecting appropriate strategies than generating situated responses. These results reveal a gap between intervention recognition and context-specific adaptation, especially for open-weight models. All data are publicly available at https://github.com/kixlab/TutorCaseBench.

Ambiguity is a central feature of natural communication: a sentence, image, or sound may remain underspecified until evidence from the other channels is considered. Yet current multimodal benchmarks mostly test perception, recognition, or reasoning over already determinate inputs, leaving unclear whether omni-modal large language models can resolve ambiguity distributed across text, vision, and audio. We introduce TRIAD, a benchmark for omni-modal ambiguity resolution. Each item is a text--image--audio triplet paired with a question and a set of answer options, and is constructed so that the full triplet determines a unique answer while removing any one modality makes the set of consistent options non-unique. TRIAD contains $327$ hand-written queries in $86$ scene groups and covers $18$ ambiguity categories across the three modalities. Evaluating $14$ omni-modal systems reveals that strong item-level performance does not translate into robust scene-level disambiguation: the best model still trails humans by over $20$ points in group accuracy. Removing answer options widens the gap further, driving model group accuracy to the single digits while humans remain robust. Leave-one-out and transcript/description counterfactuals further show that current systems often underuse audio cues when text and image are present. TRIAD reframes omni-modal evaluation around disambiguation rather than recognition, and provides a taxonomy, benchmark, and diagnostic protocol for measuring this capability.

Personal agents need to draw on long-term interaction history to retrieve and apply external knowledge suited to a user's current request. Even an answer grounded in relevant documents can be unsuitable when it overlooks the user's circumstances. This creates two connected challenges: using historical conditions to identify the knowledge needed for a new request, and applying that knowledge to the user's situation. Existing benchmarks advance memory recall, personalized retrieval, and memory-guided action, while the joint evaluation of these two challenges remains underexplored. We introduce ColaBench, a synthetic benchmark in health and daily living, with 349 tasks across 10 Profiles and a knowledge base of 13,033 documents. Its Evidence Graph connects personal conditions to their disclosures across public interaction sessions, records prior document exposure, and links answer requirements to personal and newly acquired document evidence. We also introduce MemKB, which uses state summaries to inform queries and rule application, and memory-document links to retrieve knowledge associated with the user's history. MemKB achieves 86.25% Accuracy with an average of 4.07 interaction rounds per task. Task-level analysis finds larger gains on tasks integrating information across multiple historical Facts; error analysis further identifies unsupported personal assumptions and rule misapplication as obstacles to using retrieved evidence. ColaBench provides traceable tasks and an evaluation framework for studying how memory guides knowledge acquisition and rule application.

Clinical Question Answering (QA) underpins Clinical Intelligent Decision Support Systems, but progress is severely limited for low-resource languages such as Indonesian by the absence of public domain-specific data. We introduce IndoNutriQA, an expert-curated Indonesian clinical QA benchmark of 911 question-answer pairs built from real clinical-nutrition reports. Rather than a large-scale training corpus, it is positioned as a high-quality evaluation resource: every pair is grounded in a clinically relevant context, enriched with structured metadata (anthropometric assessment, medical diagnosis, document provenance, chunk segmentation). Eight standardised question types following the Nutrition Care Process (NCP), including the Problem, Etiology, Sign and Symptom (PES) diagnosis statement, are answered for every report, which was written by a graduate-level dietetic intern, by one of three nutritionists. Answers are free-form clinical prose rather than extractive spans, and an expert validation study finds 6-10% of reference answers contestable. The dataset supports generative, extractive and retrieval-augmented QA as well as zero- and few-shot LLM evaluation. We report fine-tuned baselines (IndoBERT variants, mT5, RAG) with 5-fold cross-validation, learning curves and bootstrap 95% confidence intervals, and benchmark eight contemporary LLMs zero- and five-shot. We further calibrate our LLM-as-judge rubric against three expert raters (one RD, two nutritionists), finding it human-comparable on factuality and completeness but unreliable on clinical safety, and a human error analysis shows it misses unsupported content that the raters identify. IndoNutriQA is released to foster research in Indonesian clinical NLP and nutritional decision support.

Multimodal Large Language Models have shown promising performance in computer-use tasks, from planning user workflows to grounding natural-language actions to precise UI targets. Existing end-to-end benchmarks primarily emphasize long-horizon task completion, while GUI grounding benchmarks remain focused on conventional controls and a limited set of professional interfaces. We introduce POINTERBENCH, a diagnostic suite of 1,500 screenshot-instruction pairs probing three complementary capabilities: thin sub-element identification and geometric grounding in spreadsheets (POINTERBENCH-Sheets), character-level grounding in long-text documents (POINTERBENCH-Text), and mixed GUI grounding across 100 professional applications (POINTERBENCH-Pro). We carefully curate each task with pixel-exact target geometry and evaluate both general-purpose multimodal models and 39 specialist GUI grounding configurations. We find text grounding to be the most challenging setting, with Claude Fable 5, GPT-5.6 Sol, and the leading native specialist UI-Venus-2-9B achieving 69.6%, 53.4%, and 25.8% success on POINTERBENCH-Text despite respective 83.2%, 77.7%, and 54.5% average success across POINTERBENCH. These results reveal substantial headroom in fine-grained GUI grounding across frontier and specialist models. We publicly release the benchmark and evaluation code at https://github.com/warmwindOS/pointerbench.

Large language models can repeat an instruction while failing to keep obeying it. We study this recall-compliance dissociation in long-horizon dialogue and define State Integrity Collapse: a model preserves a stated constraint in memory but stops applying it to later turns. We introduce StateTrack-Bench, a controlled suite of conversations covering numerical, format, role/persona, factual-memory, and logical-rule constraints. Across seven provider-pinned model snapshots and five dialogue-depth probes, strict joint retention ranges from 62.4% to 100.0%, with degradation concentrated in symbolic constraints and in one especially brittle model. A forked recall audit shows that 88.6% of observed failures are behavioral-drift failures rather than ordinary forgetting; excluding the brittle outlier preserves this diagnosis (84.4% of failures). Extended evaluations at depths up to 50 turns confirm that the dissociation persists beyond the core benchmark, while Structured State Anchoring recovers up to 24.4 points on immutable constraints. The results show that reliability cannot be inferred from single-turn instruction following alone; evaluation must track whether remembered constraints remain operational as dialogue state grows. Code and artifacts are available at: https://github.com/xzzvsxd/StateTrack.

Bashkir is a low-resource agglutinative language written in Cyrillic and remains poorly represented in modern LLMs. We present an open resource suite and a systematic benchmark of adaptation methods: the Bashkir Web Corpus (71,567 documents, 46.9M tokens, 16 sources) and three labelled datasets for binary, multiclass and multilabel text classification; the benchmark covers the first two. Experiments use eight architectures (82M-8B: DistilGPT2, GPT-2 base/medium, Phi-3-mini, Mistral-7B, YandexGPT-5-Lite-8B, mBERT, XLM-R) and three adaptation groups (full fine-tuning, QLoRA, in-context learning), plus continual pretraining (CPT). (1) QLoRA on 3.8-8B models outperforms encoder-only fine-tuning on binary classification: YandexGPT-5-Lite-8B with CPT reaches macro-F1 0.748 vs 0.653 (XLM-R) and 0.650 (mBERT); pairwise differences are significant after Holm correction (p<0.001). (2) YandexGPT-5-Lite-8B significantly outperforms Mistral-7B and Phi-3-mini on multiclass classification, BPB and native-speaker ratings (mean 3.28 vs 2.52 and 2.22; n=100; alpha=0.69-0.77; p<0.05); on binary classification its QLoRA advantage is not significant (p>=0.084). (3) CPT effects depend on model and task: perplexity improves for Phi-3-mini (+0.58) and YandexGPT-5-Lite-8B (+0.33) but worsens for Mistral-7B (-1.15, p<0.001); on binary classification CPT helps Mistral-7B most (+4.0 macro-F1). (4) Tokeniser fertility on Bashkir varies 2.67x (2.92-7.80) and tracks adaptation quality, while perplexity across tokenisers is not comparable. Cyrillic tokenisers with Russian pretraining are preferable; LLM judges reproduce native ratings for grammar and coherence (rho=0.33-0.46) but not language correctness (rho=0.16, p=0.11); native evaluation remains necessary. All artefacts are open-licensed.

Coordinating two players in a cooking task requires a vision-language agent to track changing scene states, sequence recipe steps, and allocate shared resources under a time budget. We introduce an Overcooked-inspired benchmark for evaluating these capabilities in Unreal Engine environments constructed using coding agents. A single vision-language model controls both players from top-down visual observations, selecting high-level actions while the engine executes navigation. The benchmark contains 20 maps in five categories. These categories support evaluation of centralized coordination under equipment contention, changing terrain, and handoff dependencies. A prop legend and pot ingredient-count indicators provide explicit visual cues, while grid-based destinations support spatial grounding. The evaluation compares four VLMs on all 20 maps, measuring task score, order completion, delivery precision, action acceptance rate, and model API expenditure. A memory ablation on the same 20 maps examines the effect of a shared, self-authored notebook while retaining six-response action history in both conditions. The strongest evaluated model, \texttt{gpt-6-astra}, achieves 26.2\% mean order completion. Notebook memory has mixed score effects across models, with positive mean differences for Claude and Qwen and negative differences for GPT and Gemini.

Despite substantial progress in Deep Search agents, their ability to conduct end-to-end research in real-world open Web environments remains insufficiently evaluated. Unlike static retrieval settings, information on the real-world Web is distributed across diverse sources, modalities, and heterogeneous formats, often involving complex interaction mechanisms that require agents to autonomously search for, access, and integrate relevant evidence during execution. To address this gap, we introduce the OpenDS task, designed to evaluate agents' ability to perform multi-step deep search in the live, open Web. This task requires agents to autonomously explore the real Web, discover and connect relevant evidence from multiple independent sources through Web search, browser interaction, and evidence processing, and ultimately determine a unique factual answer. Building on this task formulation, we construct OpenDS-World, a benchmark comprising 200 challenging tasks spanning diverse information modalities, heterogeneous file formats, and interactive access requirements, with independent human verification to ensure evidence validity and answer uniqueness. We evaluate 12 frontier models under a unified tool environment and execution protocol. The best-performing model achieves 39.0% accuracy, while human evaluation across 1,000 independent attempts yields 21.1% accuracy. Further analyses of tool usage, search cost, termination behavior, and failure recovery reveal substantial differences in research strategies across models, as well as persistent challenges in tool coordination and long-horizon search management. OpenDS-World provides a unified evaluation platform for systematically assessing and advancing agents' autonomous deep search capabilities in real-world open Web environments.

Code-based video generation offers fine-grained control, targeted editing, and component reuse compared with direct video generation. This approach requires multimodal large language models (MLLMs) to translate visual and temporal requirements into executable programs that produce complete videos. However, existing code-based benchmarks primarily target specialized tasks, leaving complete video creation across diverse creative categories and input modalities insufficiently evaluated. We introduce Code2Video-Compass, a general-purpose benchmark for code-based video generation across text-to-video (T2V), image-to-video (I2V), and video-to-video (V2V) settings. Code2Video-Compass comprises 2,104 tasks spanning 10 categories, 24 subtypes, and three complexity dimensions covering construction, dynamic animation, and state dependency. It combines tasks curated from textual instructions and visual references with corresponding T2V, I2V, and V2V tasks derived from complete professional and user-created videos. We propose a dual-level evaluation framework combining 12 objective metrics and five subjective metrics. Objective evaluation covers operational performance, video quality, text-video alignment, and reference fidelity, while subjective evaluation uses an MLLM judge to evaluate instruction following and holistic video quality. Extensive evaluation of 15 recent proprietary and open-weight MLLMs from nine families reveals that even leading models face challenges in reliably generating high-quality complete videos aligned with textual and visual requirements, with performance generally declining as construction, dynamic animation, and state-dependency complexity increases. We anticipate that Code2Video-Compass will advance reliable, coherent, and controllable code-based video generation and inspire broader progress in programmatic visual creation.

Agents need reliable ways to validate their outputs, both when completing tasks and when filtering synthetic training data. We study verifier generation: from the instruction and environment available for solving a task, an agent writes an executable verifier that must accept every valid solution and reject invalid ones. We introduce VerifierBench, a benchmark of 222 tasks from six agent benchmarks spanning tool-call, state-based, and code verification, with 4,993 hidden test cases that include alternative valid solutions and attacks; Verifier4Verifier (V4V), a six-check audit, screens its source verifiers and test suites. Across 19 models in a common agent harness, normalized balanced accuracy ranges from 10.6 to 62.1, with generated verifiers consistently rejecting invalid solutions more reliably than they accept valid ones. This apparent strength can mask overly restrictive verification: most false rejections in the analyzed records come from verifiers that accept no tested solution to a task. Case studies reveal unsupported assumptions about solution formats, tool use, and code changes that exclude valid solutions. These errors survive self-testing because agents test solutions that share the verifier's mistaken assumptions, or submit verifiers without confirming that any correct solution passes. These findings suggest testing both whether a verifier accepts a correct solution and whether it accepts other solutions permitted by the task. The project page and dataset are publicly available.

Omni LLMs can naturally locate key audio and visual evidence and integrate information across modalities for audiovisual reasoning. Existing audio-visual reasoning benchmarks that mainly rely on multiple-choice questions usually score models solely based on answer correctness, making it difficult to assess whether a model truly forms a reasoning chain from evidence to the correct answer. In addition, textual or unimodal shortcuts may allow models to answer correctly without analyzing the necessary audio-visual evidence, further weakening the reliability of evaluation benchmarks. To address these issues, we introduce EvaOmni, an evidence-grounded benchmark containing 392 videos and 1,027 purely human-annotated question-answer pairs, covering 15 capabilities and three levels: retrieval and localization, basic understanding, and complex reasoning. We design a series of rigorous annotation procedures in which experts annotate audio and visual evidence and construct video-based distractors; humans review whether questions meet strict inspection criteria, and finally models filter textual and unimodal shortcuts. Beyond high-quality question-answer pairs, we require models to simultaneously return the answer and supporting evidence intervals in order to evaluate the complete reasoning chain. We propose the Coverage Grounded Score (C.G.) to measure how much annotated evidence a model identifies, and the Bottleneck Grounded Score (B.G.) to penalize poor temporal localization in either modality; both are incorporated into answer correctness. Across 7 open-weight and 17 proprietary Omni LLMs, Gemini3.7-Flash achieved the highest overall accuracy (79.4%), C.G. (70.8%), and B.G. (35.1%). At Level 3, it reached 79.1% accuracy, but its B.G. was only 12.3%. These results show that even the most advanced models still have room for development in grounded audio-visual reasoning.

Recent advances in multimodal large language models (MLLMs) have enabled them to act as autonomous agents, creating and editing 3D models via code or direct software manipulation. In practical workflows, these models must infer object shapes, materials, spatial layouts, and lighting from diverse references while strictly following instructions to reconstruct, modify, or design 3D content. However, existing benchmarks provide limited coverage of such open-ended requests across real-world scenarios. To bridge this gap, we introduce BlenderWorld, the first comprehensive benchmark dedicated to evaluating frontier models on realistic 3D creation tasks. BlenderWorld comprises 223 tasks spanning architecture, products, natural environments, and scientific visualization, driven by text, image, video, and structured specification inputs. To rigorously assess these complex deliverables, we establish a unified evaluation protocol: we render the submitted Blender projects from multiple viewpoints and evaluate them against 2,651 task-specific binary rubrics alongside holistic visual quality scores. Benchmarking thirteen frontier models reveals that GPT-6-Astra achieves the highest overall fused score of 79.9, followed by Opus-5 (73.0) and GPT-5.6-Sol (67.7). Ultimately, our findings reveal a critical bottleneck: while current models can parse broad instructions, they still profoundly struggle with fine-grained 3D spatial coherence and accurate structural reproduction in Blender, leaving substantial headroom for future 3D-native agentics.

AI agents are demonstrating increasingly strong offensive and defensive cybersecurity capabilities in the real world. Current benchmarks mostly evaluate agents in isolation against static targets. Real cyber operations, however, are dynamic and coupled: security postures constantly change as attackers and defenders adapt to each other in real time. Static benchmarks quickly saturate and miss an agent's ability to attack and defend simultaneously under live pressure. To address this gap, we present CyberArena, a dynamic, symmetric adversarial benchmark built around multi-round attack-and-defense combat. In isolated sandboxes, two agents compete under identical rules, acting concurrently as both attackers and defenders over continuously changing services. This converts static targets into evolving, semi-open challenges, directly testing an agent's speed and precision in discovering, exploiting, and patching vulnerabilities. To sustain the benchmark's longevity, an automated agentic pipeline continuously generates, reviews, and validates new environments. Furthermore, because pairwise contests are notoriously hard to quantify, we introduce a dynamical model that maps match dynamics into unified, multidimensional scores across offensive, defensive, and domain-specific capabilities. Evaluating 12 frontier LLM agents across 40 challenges in 466 matches, CyberArena separates model capabilities where single-agent baselines hit a ceiling. The trajectories read like a fight rather than an exam: defenders patch one hole and miss the next and certify repairs on stale evidence, while capable agents think a move ahead, planting backdoors for later rounds and, in one case, writing their own packet sniffer to watch the opponent. Beyond a benchmark, CyberArena serves as an adaptable, scalable experimental platform for future research on cybersecurity agents.

Hallucinated citations make it harder to trace sources and verify the evidence behind scientific claims. Detecting them is difficult because bibliographic records can be incomplete or differ across source types and publication versions. A plausible match alone neither validates the citation nor establishes its underlying source. We introduce CiteResolve, a benchmark of 423 real-world hallucinated citations annotated for source recoverability. We further propose CiteAnchor, a source-aware grounding framework that adapts retrieval and field-level verification to each source type and separately determines citation validity and whether a unique source can be recovered. On an existing detection benchmark, CiteAnchor achieves 0.841 F1 with GPT-5.6 Luna and 0.784 with an open-weight backbone, outperforming the baseline with the same backbones by 7.9 and 36.8 points. On CiteResolve, it achieves 74.7% and 70.7% end-to-end task success, outperforming the corresponding web search baselines by 16.3 and 65.3 percentage points. Our code and data are available at https://github.com/oldwizard7/CiteANCHOR_code.

World models help agents assess the consequences of their actions before acting. Many real systems, however, involve multiple actors who make their own decisions, can evolve even when the agent takes no action, and are only partially observed through event logs. We formalize these systems as event worlds and ask whether large language models can predict future observations from the log and the agent's proposed actions. We build complementary benchmarks from human Diplomacy games and real Polymarket trades. Our hierarchical framework compares prediction with and without other actors' actions, then tests multi-step rollouts, predictions for actions never taken, and usefulness to agents. Fine-tuning and more data improve prediction when all actors' recorded actions are given. With only the agent's action, models struggle when other actors strongly affect the outcome, and larger models and longer histories help little. This gap also limits multi-step prediction and agent decisions, with models overestimating plan success. Yet imagined experience and initialization from world models can still improve policy learning. Anticipating other actors' actions remains a central challenge for event world modeling.

Medical knowledge evolves as clinical guidelines revise recommendations, yet whether large language models (LLMs) faithfully incorporate such changes remains unclear. We introduce ChronoEvo, comprising 1,709 recommendation-level revisions across 12 forms of medical knowledge change, to distinguish recognition of revised guidance from incorporation into clinical recommendations. In the primary model, 90.8% of responses recognized revisions and 80.2% were clinically correct, but only 58.5% fully incorporated guidance, yielding a 32.3-percentage-point recognition-incorporation gap. This dissociation persisted across 14 models and was corroborated by two clinicians' adjudicated assessment of 300 responses, yielding a 40.3-percentage-point gap in the validation sample. Changes in recommendation strength and patient eligibility were associated with retention of outdated guidance. A revision-salience intervention reduced outdated-content retention and improved clinical correctness but did not significantly improve full incorporation. These findings identify faithful incorporation of evolving medical knowledge as a distinct challenge for clinical LLM reliability, even when updated guideline evidence is available.

Chemical AI workflows routinely represent the same molecule through multiple co-referring identifiers, including trivial names, IUPAC names, and SMILES, implicitly assuming that these representations provide interchangeable access to an LLM's knowledge. To test this assumption, we first introduce a controlled benchmark and show that it does not hold. Across foundation models spanning 7B to 72B parameters, even chemistry-specialized models, co-referring molecular identifiers consistently expose different subsets of the model's knowledge. We then conduct a series of controlled behavioral, representational, and causal analyses to characterize this failure. We show that even when a model can identify a molecule from its SMILES and retrieve a fact through the corresponding name, the same fact can remain inaccessible directly from SMILES. We term this phenomenon \emph{cross-identifier binding failure}. Further analysis of internal representations, conflicting identifiers, and parameter editing shows that neither improved chemical representation competence nor modification of individual facts is sufficient to establish reliable access across identifiers. Together, these results indicate that the core failure lies in the binding between an identifier and knowledge already present in the model. Motivated by this diagnosis, we introduce Binding Self-Distillation, a simple mitigation that uses lexical identity during training to transfer the model's existing knowledge access to the structural route. This enables factual retrieval directly from SMILES at inference time, without requiring a name lookup or training the underlying model from scratch.

Sparse attention reads only a small subset of the key-value cache, and how to treat the unread tail remains a key challenge. Discarding the tail biases the output, and sampling it as previous approaches do fails at reads of 5% and below, where the tail likely holds heavy hitters the selector missed. Our key observation is that the simplest possible proxy, one token that folds the entire tail into its count-aware mean, already yields lower output error than a sample at every read behind a practical selector. We further show in closed form that at low reads sampling nearly doubles the weight error of discarding while a single proxy stays below it. Building on these findings, we explore approximating the dense attention distribution with a few proxies that target the tail. We propose OpenAttention, a pipeline of clustering, ErrorGate and top-k selection followed by a count-aware mean proxy. To implement clustering and selection both efficiently and effectively, we design training-free Query-Geometry Clustering (QGC) and a Query-Geometry Selector (QGS), which compress the keys while respecting the query directions. Additionally, unlike prior methods that rank clusters by centroid alone, ErrorGate allocates the exact read by the estimated weight error of each cluster. On RULER-32K HARD with Llama-3.1-8B, OpenAttention retains 99% of dense accuracy at a 1% read. It also achieves state-of-the-art accuracy across models, context lengths and datasets. Our customized attention kernel, OpenAttention.flash, runs the decode attention stage 10$\times$ faster than dense attention at a 1% read.

Capability benchmarks measure what large language models can do, but not whether their recommendations remain appropriate as decision conditions change. This paper proposes a trajectory-based measurement framework and prospective validation protocol for artificial wisdom, defined behaviorally as context-sensitive regulation of judgment. Versioned reference specifications determine admissible responses and transition materiality before evaluation. Seven stage-level criteria assess judgment under current conditions, while two transition components assess material-update adequacy and commitment-preserving stability. Stability requires prior adequacy; switching within registered commitment classes is permitted, and persistent error earns no stability credit. A nine-component profile is primary, with a secondary bottleneck score and separate critical-failure reporting. Three simplified simulations examine precision, bottleneck estimation, and sensitivity to model-family count. A single-criterion simulation met its precision criteria at 24 scenarios per domain. Near a bottleneck tie, the estimated minimum had a downward bias of 0.55 points at that size. A simplified incremental-prediction simulation detected a Brier improvement of 0.027 with probabilities of .508 and .742 using eight and twelve families, respectively; it does not establish power for the operational protocol. An exploratory predecessor-rubric study comprising 71 evaluation records illustrates aggregation sensitivity but does not validate the proposed instrument. The twelve-family target remains provisional pending feasibility assessment and calibration of the complete analysis. Whether the profile predicts independent decision outcomes beyond capability remains an open empirical question.

As LLM agents operate over corpora of more than ten thousand tools, selecting the right tool for a complex multi-step user instruction becomes a central bottleneck. Tool descriptions cannot all fit in context, and relevance-based retrievers only rank tools without knowing which are actually needed. Recent work instead leverages proprietary commercial LLMs like Claude or ChatGPT to drive a retriever, decomposing the instruction and selecting over retrieved candidates. We show that a small, open model can be trained via reinforcement learning to perform the same functionality. On ToolBench, a Qwen3-8B model post-trained with our recipe outperforms Llama-4-Maverick (402B). In addition, the trained model transfers directly to different retrievers and generalizes to new tool corpora in different domains.

Control evaluations measure whether safeguards can prevent an AI model from causing undesired outcomes even when it is intentionally trying to do so. These evaluations typically use purpose-built control settings and extrapolate their results to a model's deployment. However, the effectiveness of safeguards in deployment can depend on features of environments and tasks that control settings do not represent. To broaden where control evaluations can be run, we introduce simulated control evaluations (SCEs), which use a language model to simulate tool responses in place of an environment. We evaluate SCEs on LinuxArena, where execution provides ground-truth outcomes. We compare SCE prediction errors with those from a separate control setting, BashArena, and with variation in control metrics within LinuxArena. For independently sampled evaluation and deployment task sets, SCEs predict safety with a root mean squared error of 3.6 percentage points across five monitors at a 5% audit budget, compared with 10.1 for BashArena. SCEs closely reproduce distributions of attack strategies, attacker reasoning, and monitor reasoning, with over tenfold lower KL divergence from execution than BashArena. SCEs explain 78-85% of the between-environment variance in attack success, monitor scores, and safety. This accuracy is sample efficient: SCEs predict safety with less error than BashArena using fewer than 5% as many task-pair draws as the deployment task set. Strategies discovered through simulation also improve monitor performance on held-out LinuxArena attacks, outperforming BashArena and approximately matching improvement from executed strategies. These results suggest that deployment-specific SCEs can be more predictive than control evaluations executed in separate control settings.

Deep research agents synthesize open-web evidence into long reports, but current benchmarks do not evaluate research centered on a source video. We introduce Video-Centered Deep Research Report Generation. Given a video and user request, an agent identifies research questions, retrieves external multimodal evidence, and produces an interactive report linked to relevant video moments. This task has direct practical value because it helps users locate relevant moments, verify claims, reproduce demonstrated procedures, and make informed decisions from video content. We present a skill-based workflow and OmniDR-Bench for this task. For evaluation, we propose Omni Atomic Rubric Trees (Omni-ARTs). Each tree contains three dimensions, 23 shared directions, and task-specific atomic binary criteria. The directions test audiovisual grounding, multimodal evidence use, and interactive presentation. Audited Rubric Construction (ARC) builds these trees through independent evidence collection, automated checks, behavioral calibration, and iterative human review. OmniDR-Bench contains 50 Chinese and English tasks across 19 realworld subcategories. It includes 6,414 criteria. Across 18 agent configurations, the best system scores 63.3 out of 100. Current agents still need broader content coverage, stronger causal and logical analysis across modalities, and more accurate citation support. We further analyze models, harnesses, skill configurations, video categories, and research trajectories. The code and data are released at https://github.com/marinero4972/OmniDR-Bench.

Spatial intelligence requires agents to actively acquire visual evidence and integrate observations into a coherent representation of the three-dimensional world. Yet spatial question answering offers limited evidence of whether models can construct accurate metric maps through autonomous exploration. We introduce SpatialScout, a benchmark for evaluating vision-language model (VLM) agents on active object-level 3D mapping. SpatialScout comprises 80 diverse indoor scenes containing 3,140 target objects. Given an object inventory, RGB observations, camera calibration, and camera poses, agents control a camera to explore each scene and estimate object centers in a shared coordinate frame. The task requires coordinating viewpoint selection, cross-view association, geometric estimation, and map revision under bounded interaction budgets. Evaluation jointly measures localization accuracy and target coverage through object-scale overlap scores that penalize missing predictions. Experiments with frontier VLMs reveal substantial limitations: the strongest evaluated model, Claude Opus 5.5, achieves a mean overlap score of 41.71%, with 47.23% of targets exceeding an intersection-over-union (IoU) threshold of 0.5 and only 3.25% exceeding 0.9. These results reveal a substantial gap between recovering coarse object layouts and constructing complete, precise metric maps. SpatialScout provides a quantitative testbed for developing agents that build reliable 3D spatial representations through active perception. Code is available at https://github.com/JesseZZZZZ/SpatialScout.

Frontier coding agents are increasingly framed as end-to-end automated researchers, and ML for science offers a demanding test of that promise: improving a scientific model often requires domain insight beyond routine ML engineering. We introduce Agent4Sci, a long-horizon testbed for measuring how far today's agents can improve the published best models on scientific ML benchmarks, spanning 20 benchmarks across 11 domains. Rather than ranking agents on a leaderboard, Agent4Sci asks what kinds of progress agents can make on these tasks under in-the-wild conditions, and where they fall short. Empirically, today's frontier agents advance the published best on every benchmark, with Claude Opus reaching 10.4% mean relative improvement and no run flagged for academic misconduct. These gains, however, come almost entirely from a bag of ML-engineering tricks rather than research-level insight. Trajectory analysis surfaces three recurring limitations in this setting: novelty miscalibration, resource under-utilization, and rapid collapse into local engineering sweeps. Lightweight plug-in skills for research ideation, experiment management, and scheduled checklists raise Claude Opus from 10.4% to 20.1% and Codex from 4.4% to 8.1%, while yielding more diverse, domain-relevant interventions. Agent4Sci gives the community a way to track and push frontier agents from ML engineering toward domain insight on real-world ML-for-science problems. Code and data are available on GitHub (https://github.com/maszhongming/agent4sci) and Hugging Face (https://huggingface.co/datasets/MingZhong/agent4sci).

As multimodal resources are transformed into media outputs, agentic AI is expected to stay aligned with design intent over long horizons. Designers follow briefs specifying goals, constraints, and deliverables, with later work building on earlier decisions. Recent work reuses design experience across tasks, yet alignment among artifacts within one project remains challenging. Agents therefore need to preserve and apply this intent across domains and as projects grow. We introduce Design Anything, a graph-based harness that preserves design intent outside the model and binds deliverables to the workflow steps that produce them through an intent-execution graph. A task contract retains constraints and decisions, selected designs become explicit inputs to later generations, and a deterministic controller uses bindings and execution records to check output types and counts, identify outstanding deliverables, and guide subsequent turns. We also introduce DesignAnything-Bench, with 228 tasks across 19 design domains, including 38 project families at scales of 6, 14, and 24 artifacts. With the LLM held fixed, Design Anything satisfies 85.0% and 92.5% of the brief's factual constraints with DeepSeek V4.1 Flash High and GPT-5.6 Sol High, respectively, exceeding the strongest same-model baselines by 15.4% and 14.4%. These gains reach 25.1% and 23.0% on 24-artifact tasks. Under each model, Design Anything leads or ties in constraint satisfaction in 18 of 19 domains. Professional designers also prefer its complete deliverable sets over those of same-model baselines in a blind evaluation.

Spatio-Temporal Video Grounding (STVG) aims to jointly localize a target in space and time given a natural-language query. Despite continued advances, existing STVG benchmarks remain largely centered on localizing targets through instance-specific descriptions, leaving them largely decoupled from long-form reasoning increasingly emphasized in multimodal large language models (MLLMs). In this paper, we introduce Event-STVG, a task that shifts STVG from locating what is described toward event-level reasoning by unifying long-form context, repeated-event reasoning, event localization, role inference, and multi-target grounding. To facilitate research on Event-STVG, we construct a large-scale benchmark comprising 4,016 annotated event occurrences across soccer and seven additional domains. Experiments with state-of-the-art MLLMs reveal substantial difficulty, with the strongest model achieving only 5.4% m-vIoU at the 10-minute horizon. Across domains, we consistently observe confusion among repeated same-type events, spatial grounding bottlenecks, and rare success in jointly grounding all required roles. Event-STVG establishes a new challenge for visual understanding in MLLMs, calling for future models that can jointly reason about complex events across space and time. Code and annotations are available at https://github.com/THU-KEG/EventSTVG

In this paper, we systematically investigate the mechanisms, scaling behavior, and generalization of physical Recursive self-improvement. We introduce PhysRSI-eval to quantify physical RSI through a five-axis metric: First-Attempt Performance, Efficiency, Success, Budget and Improvement from failure. Different VLMs are evaluated on 18 manipulation tasks under a fixed interaction budget. The analysis of experimental results motivates the Physical Barrel Hypothesis: different capabilities may constrain different parts of improvement. The five-axis metrics also distinguish systems with close public benchmark averages and show that high scores need not come with early or efficient gains. We study observational scaling from two perspectives: predicting the relative performance of VLM agents from their public capabilities, and forecasting later progress from early physical feedback. Together, these analyses ask how quickly and how far an agent can improve. Moreover, we conduct workspace interventions and then analyze what improvement leaves behind. Removing history reduces late-round success, frozen workspaces remain useful on unseen layouts, and three cross-task cases achieve success earlier. The results suggest that scene-specific adaptation and reusable experience can coexist.

Reasoning LLMs often continue reasoning after reaching a correct answer, wasting computation and sometimes replacing it with an incorrect one. Deciding when to stop is a metareasoning problem: the model must weigh what further reasoning could achieve against its cost, given what it can already answer. Direct utility training learns a stopping representation at a chosen token price; supporting new prices can require additional controller training. We introduce BELT (BELief-based Termination), which separates outcome prediction from stopping control. It learns beliefs about current and future answers from a frozen LLM's hidden states, then fits a linear stopping policy on the frozen features. The prediction targets are independent of token price, so one probe supports different accuracy-cost trade-offs through CPU policy refitting or threshold adjustment. Averaged over five out-of-distribution (OOD) math benchmarks, BELT reduces generated tokens by 41% on Qwen3-1.7B and 51% on OLMo-3-7B-Think without loss of accuracy. The same math-trained BELT transfers to coding and agentic tasks without training on code or agentic data. In all three domains, its accuracy-cost trade-offs are competitive with or better than those of end-to-end controllers. On OLMo3, its probe has 21.2x fewer parameters than one checkpoint of the strongest baseline, OS-Pruner, and needs 55x fewer training GPU-hours than OS-Pruner's five price-specific checkpoints.

We present Protegrity-PII, a new PII decision model for structured and unstructured text. Protegrity-PII achieves a typed recall of 0.915 with a precision of 0.907 on a new benchmark we propose, PII-SCOPE. Our model processes text in 0.49 seconds per million characters at \$0.0005 per million characters on a single RTX 6000 using batched inference. Due in part to the expansive nature of the entity type taxonomy used in training, Protegrity-PII achieves superior out-of-distribution generalization when benchmarked against a selection of state of the art models and datasets.

Existing datasets on cultural knowledge have been typically created by human annotators familiar with each culture. However, the process requires considerable human effort, presenting a significant challenge to scalability. To address this challenge, we introduce WikiOriginQA, a QA dataset automatically constructed by objectively representing the origins of knowledge through the creation dates of different Wikipedia language editions. Experimental results with the created dataset confirm that LLM performance tends to be biased toward Anglophone cultures. Furthermore, the results suggest the importance of knowledge propagation into the Anglosphere. These findings demonstrate the validity of our automated approach, enabling cultural bias evaluations with superior scalability across diverse languages without extensive manual annotation.

Nepali-English code-switched conversational speech presents severe challenges for automatic speech recognition (ASR), chiefly through script confusion (inappropriate mixing of Devanagari and Latin orthography) and elevated error rates at language switch points. Supervised fine-tuning (SFT) can adapt multilingual foundation models, but human code-switched audio is expensive to collect and annotate. We curate a new benchmark of 13,044 spontaneous podcast utterances (15.45 h total; 10,214 test, 2,830 dev), audited and verified to eliminate machine translation leakages and acoustic loops. We investigate a two-stage synthetic-to-real adaptation framework: Stage 1 applies SFT on 0.95 h of speech synthesized via a multi-speaker, dual-engine TTS pipeline; Stage 2 explores Group Relative Policy Optimization (GRPO) with transcription and script rewards. We find that synthetic SFT provides the primary cross-domain transfer to spontaneous human speech, reducing Mixed Error Rate (MER) from 23.40% to 21.52% (-1.88 pp) and switch-point error rate (SPER) from 35.89% to 28.77% (-7.12 pp). Subsequent GRPO yields small, statistically significant adjustments on 173 test utterances (1.69%), driving English script confusion (SCR_en) from 6.45% to 6.36% (p <= 0.001) and SPER to 28.70% (p = 0.022 clustered). Crucially, ablation reveals that in this SFT-converged encoder-decoder setting, an auxiliary script reward term yields performance identical to MER-only optimization because rollout candidate groups exhibit near-zero script variance on synthetic data. On monolingual benchmarks, SFT improves Nepali MER by 2.71 pp (stabilizing at 18.42% after GRPO, a 2.38 pp gain over baseline) while preserving English lexical WER (3.38% vs. 3.33%). Verified transcripts, timestamps, episode metadata, and reconstruction scripts are publicly released.

With external tools, large language models (LLMs) can tackle complex tasks through sustained environmental interaction. LLM-based agents are therefore increasingly used to tackle complex long-horizon tasks in real-world environments. However, these tasks involve sequential, interdependent decisions and changes in environmental states. Early execution errors affect subsequent interaction histories and environmental states, and their effects accumulate along the decision chain. To address this challenge, we propose Reflection-Guided Policy Optimization (ReflectPO), a reinforcement learning framework for autonomous error recovery in long-horizon tasks. ReflectPO comprises two core components. The first is Reflection-Guided Sampling. We use branch differences and repeated visits in interaction trajectories to construct hints for error recovery. The second is Tree-Structured Credit Assignment. We organize sampled trajectories into trees and assign action-level credit using parent--child reward differences and recovery gains. Experiments over two long-horizon tool-use benchmarks show that ReflectPO outperforms multiple existing baselines in most domains. It also achieves the best overall performance on average. Further analyses show that ReflectPO effectively improves agents' ability to recognize their own deviations, recover from erroneous states, and continue making progress toward task completion.

We present IQuest-Q1, a sparse Mixture-of-Experts (MoE) agentic foundation model with 320B total parameters (15B activated). IQuest-Q1 is developed through a three-stage process: (1) stable and efficient pre-training, combining Stable SplitNorm, Muon, and muP with a two-stage curriculum whose second stage increases the coding-data mixture, preserving a wide stability margin between safe and optimal learning rates; (2) fine-grained mid-training, expanding the data mixture across domains and formats and progressively extending the context; and (3) multi-scaffold with multi-version post-training, combining SFT, domain-specialized RL, and MOPD across various scaffolds to integrate complementary training signals and strengthen agentic coding capabilities. Furthermore, model merging combines different experts and checkpoints from different stages while mitigating conflicts among their parameter updates. The resulting IQuest-Q1 demonstrates capabilities in both coding and general tasks. We release IQuest-Q1 as an open-source model to support community research and development.

Hierarchical language generation separates continuous semantic modeling from discrete text realization, but its effectiveness depends on what the latent space compresses and how the discrete generator uses it. We propose Cola2, a framework that jointly studies continuous semantic compression, the balance between reconstruction and generation, and continuous-discrete co-adaptation. Cola2 dynamically segments text into semantic spans, controls the compression rate through auxiliary-loss-free negative feedback, and combines reconstruction and generation supervision to learn compact, predictive latent representations. A block-causal masked diffusion decoder is trained with clean, noisy, and dropped latent conditions, while classifier-free guidance adjusts the balance between semantic reconstruction and autonomous generation. Token-prediction gradients update the encoder through the latents, whereas Flow Matching prior learning is decoupled from representation learning by stop-gradient. During inference, generated text is re-encoded under a fixed span layout to keep the latent history consistent with the actual text context. Experiments on five downstream tasks with about 210 billion training tokens validate the core designs and show competitive, unsaturated scaling. With 346.03 billion source tokens, an unconverged checkpoint reaches a 60.10 average across eight tasks and surpasses Llama-3-8B on HellaSwag and StoryCloze despite its reported pretraining budget being over 43 times larger. These results demonstrate the potential of continuous-latent hierarchical language generation.

Tree-based speculative decoding accelerates large language model inference by verifying multiple candidate branches in parallel. However, in practical serving scenarios, similar requests repeatedly induce overlapping candidate branches. While retrieval-based methods mitigate this by reusing historical drafts, the resulting candidate trees often contain many low-quality branches as the candidate tree grows, which reduces verification efficiency and wastes target-side computation. To address this, we propose ValuSpec, a plug-and-play candidate valuation mechanism that filters low-quality branches from a supplied candidate tree before expensive target-side verification. ValuSpec introduces transfer-based supervision to align filtering with the target model's behavior. A filtering threshold is estimated from the target model's greedy trajectories on a calibration split, without any parameter training. The target model then verifies the filtered tree under the output-identical greedy rule, so accepted tokens are guaranteed to match standard autoregressive decoding exactly. Experimental results on HumanEval, GSM8K, Dolly, and SpecBench show that, in our main evaluation, ValuSpec achieves up to 2.38x end-to-end speedup in retrieval-based tree settings. Furthermore, integrating ValuSpec into the EAGLE-3 dynamic tree framework yields up to 2.78x end-to-end speedup without modifying its tree construction strategy, demonstrating that ValuSpec serves as a generally applicable and plug-and-play candidate valuation module.

Individual preferences often become visible only after deployment, when a user accepts one decision or revokes it in favor of another. Jev-style typed interfaces make this feedback explicit by identifying the selected action and, under revocation, the decision it replaces. We study how these consequences persist across requests as Decision-Level Continual Learning from Typed Feedback (Jev-DLC). Jev-DLC treats each correction as both evidence about a recurring action identity and a directed relation between the preferred and rejected choices. It preserves these signals in complementary bounded channels above a frozen scorer and combines their scores to shape the next decision. The method retains no raw interaction history, adds no scorer call, and performs no runtime language-model backward pass. We also introduce TCPA-Bench, a chronological suite that credits a correction only through its effect on later decisions. On TCPA-Repeat, Jev-DLC reaches 44.83% user-macro accuracy overall, 2.43 points above Recent-4 history replay. After at least two earlier feedback events, it reaches 56.07%, a 7.54-point margin over the strongest reported comparison at this depth. These results show how explicit typed corrections can become compact causal state for continual personalization.

Purpose: To describe an evidence-aware retrieval evaluation protocol for a clinical retrieval-augmented generation (RAG) system over breast lesion management literature, addressing the dependence of retrieval scores on annotation coverage. Methods: An 18-document corpus and an 18-query development benchmark, with one Yes/No or WH-question per document, were used. Retrieval combined dense embeddings, BM25, and cross-encoder reranking. Chunk sizes of 128 to 1024 tokens, overlaps of 0 to 20%, and four retriever configurations were evaluated. Relevance was defined by evidence spans mapped to all containing chunks, treating overlapping chunks with equivalent evidence as relevant. Retrieval was evaluated using R-Precision, Recall@K, MRR, and nDCG@K. Separately, generation was assessed on a 512-token index at K=1 to 5 using DeepEval claim-level metrics, with deepseek-v4-pro as the LLM judge, and BLEU/ROUGE metrics. Results: The selected hybrid retriever, using 384-token chunks and 20% overlap, retrieved accepted evidence in the top three results for 17 of 18 queries. Recall@3, R-Precision, MRR@3, and nDCG@3 each reached 0.94. Ground-truth curation increased evidence spans from 25 to 37 and accepted chunk targets from 36 to 51 without changing retrieval rankings. Conclusion: Cross-encoder reranking achieved the highest retrieval scores. Expanded relevance annotations increased recognized evidence without changing rankings, highlighting annotation sensitivity. The small development set and retrieval-informed curation require independent validation, including radiologist assessment of evidence sufficiency.

Retrieval Augmented Generation (RAG) grounds large language models in external knowledge. Modern pipelines combine embedding, chunking, retrieval, reranking, query augmentation, context compression, and agentic verification, yet these techniques are typically evaluated one at a time. It remains unclear whether their isolated gains survive composition, whether retrieval quality and answer faithfulness favor the same design choices, and whether conclusions from small corpora hold at large scale. We pre-declare three hypotheses and introduce LASAGNA, an open-source framework for controlled RAG evaluation. We test 450 configurations across eight benchmarks and corpora ranging from 5K to 200K documents. Three findings emerge. First, individually strong techniques from different phases of a RAG pipeline do not necessarily compose: three retrieval techniques predict a +47% combined improvement but deliver only +17% when used together. Second, component gains stack differently for retrieval and faithfulness, and optimizing one can hurt the other. Third, the best composition changes as the corpus grows: the combination that wins on a small corpus may not remain best at large scale. These findings show that gains measured in isolation are not reliable predictors of realistic performance, and that new RAG techniques should be tested with the full surrounding stack. We call on researchers to do so and supply LASAGNA to simplify these experiments.

Lean verifies that a generated declaration is well typed, but cannot check whether it states what the user intended. We measure how far compilation overstates faithfulness in statement autoformalization without canonical Lean targets. Our criterion requires Lean compilation and agreement of two LLM judges from different model families. On an independent random audit, it agrees with the human majority on 91.5% of cases; it rarely accepts an unfaithful output but rejects some faithful ones. Across eight systems on 227 university-level statements, the gap between compiling and accepted outputs ranges from 1.3 to 29.5 percentage points: our tool-augmented agent compiles 87.2% of the statements, but only 57.7% are accepted. Human review of all the agent's rejections confirms that 23.8 of its 29.5 points are real translation errors. For the gaps of the seven one-shot systems, a two-stage review by Claude Opus 5.5 with Lean and GPT-6.1 Sol, validated on these human labels, finds that 78-100% are real translation errors for six of them. When GPT-5.2, one of the two judges, grades the translations of two models released after it, GPT-6 Astra and Claude Opus 5.5, it accepts 83% of the statements for each, against 21% for one-shot GPT-5.2. For these two models, the same review classifies only 4 of their 54 rejections (7%) as translation errors, so most of their measured gap is judge errors rather than translation errors. This is consistent with these systems erring less while the judge does not: the fewer errors a system makes, the smaller the share of its rejections that are translation errors. Finally, a 2^3 tool ablation of the agent shows that Lean feedback drives most of its compilation gain, and that nearly half of this gain fails the semantic criterion. Code and data are available at https://github.com/grenadeComing/beyond-compilation.

Training Arabic large language models faces challenges from noisy web data, dialectal variation, and the high cost of continual pretraining (CPT). We introduce AQL-Lite (AQL: Arabic Quality Language), a time- and resource-efficient framework for adapting a multilingual LLM to Modern Standard Arabic (MSA). AQL-Lite couples a quality-first curation pipeline (structural cleaning, heuristic and model-based filtering, MSA selection, and deduplication) with a partial-parameter CPT recipe that trains RS-LoRA adapters across the network together with the last eight transformer blocks and the embedding matrix, under layer-wise learning rates. At a matched 1B-token budget, CPT on AQL-curated data raises a four-benchmark Arabic mean by 3.51 points over CPT on raw web text, which yields no gain over the base model. In a controlled 5B-token comparison, AQL-Lite recovers 94% of the gain of full-parameter CPT (mean 73.19 vs. 73.58) in 52% of the wall-clock time and outperforms adapter-only baselines on every benchmark, and a component ablation shows that adapters, unfrozen blocks, and embedding training each contribute. Scaled to 40B MSA tokens and instruction-tuned on one million examples, the resulting 12B model scores 75.3 on the seven-benchmark Open Arabic LLM Leaderboard v2 (OALL v2) suite. These results suggest that partial-parameter CPT on a carefully curated corpus is a practical route to efficient MSA adaptation.

Large language models often apply complex reasoning indiscriminately, leading to overthinking on simple tasks and underthinking on harder ones. We present Route-and-Generate (Rouge), a hierarchical training framework that enables LLMs to adaptively decide when and how much reasoning is required for each input. Our approach decouples reasoning effort allocation from answer generation by introducing a routing policy that allocates a fixed sampling budget across multiple reasoning formats, while the base model is optimized with GRPO using per-sample rewards. To promote efficient reasoning, we employ a rank-based reward that favors shorter correct solutions, and update the routing policy via a closed-form KL-regularized policy improvement step distilled into a neural router. This design encourages concise answers on easy problems while reserving longer reasoning for harder ones, achieving improved accuracy-efficiency trade-offs across multiple reasoning benchmarks.

LLM agents are often evaluated by reducing execution traces to labels such as final-answer correctness, tool verification, and fault propagation. These labels are measurements produced by parsers, behavioral detectors, and harness rules, and they may not fully capture what the trace shows. We study this in a tool-use setting built on GSM8K, where agents call a calculator whose outputs are occasionally corrupted, giving clear ground truth for what the tool returned, what the model later used, and what the detectors reported. We audit three trace-level detectors across six production models and 30 problems per model. Final correctness, visible re-verification, and fault propagation can disagree in systematic ways. A key failure mode is silent tool-output override: the calculator returns a corrupted value, but the model later uses a different, often correct, value without calling the tool again, so the bad value stops propagating while a verification detector records no re-check. Human review of 90 traces agrees perfectly with the automated labels for final correctness (67/67), fault impact (19/19), and fault relevance (18/18), but only 67/81 for verification; all 14 disagreements fall into four recurring patterns. Execution limits also change the reported result: in an ablation across two models, one ranges from 21/30 to 27/30 correct as step and token budgets change, while the other stays between 27/30 and 29/30. These findings motivate measurement transparency: an automated label should come with the trace evidence behind it, whether the question the detector asks applied to that trace, whether the run was cut off by an execution limit, and any human corrections. Our trace viewer makes this information inspectable alongside the reported results.

LLM agents that act on live Web data must combine heterogeneous evidence, remember what they have done and improve from outcomes without weakening their own safeguards. We study this problem in autonomous investing, where every action is irreversible and based on untrusted Web content. We present Steward, a multi-agent LLM system that draws the boundary of self-evolution in its type system. First, nine role-specialised agents decide jointly: four analysts read news feeds, SEC filings, macroeconomic series and market data and emit typed reports, from which a Portfolio Manager sets target positions that a protected Risk Manager audits. Second, a seven-layer shared memory, assembled by code from the system's own record, gives each decision its position context, the market narrative, calibration and the agent's past decisions. Third, the system improves itself by review: code grades every trade and forecast nightly and writes the grades into memory, and each quarter a reflector proposes prompt learnings that a deterministic gate applies reversibly and replay of recorded inputs verifies. All three operate around an invariant core of order arithmetic, risk rules and discipline-agent prompts that no self-edit can reach. In a five-month live deployment (102 trading days, 1,632 LLM calls), the system contained a 24.6% sector drawdown to a 0.3% loss; in replay of its recorded inputs, a single model given the same raw data traded more and lost 1.35 points against the market over five days where the joint design showed no significant loss, and no injected instruction was followed; and a 2015-2026 backtest shows the core alone lowering the maximum drawdown under all 50 random entry policies. The reflector traced the dominant failure to its root cause, and the gate applied a learning that the agent adopted on trigger evenings and on no control evening. Code: https://github.com/yebof/quant-agent

An AI system can produce more economically valuable work while delivering less value for the energy it consumes. We therefore ask: How much economic value does an AI system create per unit of energy consumed? We formalize economic value in two ways: (1) net dollar value, the output's dollar value minus AI-system execution cost, and (2) time saving, the expert's completion time minus the AI system's completion time. Dividing each by measured end-to-end energy yields Net Value per Joule (NVJ) and Time Saving per Joule (TSJ), reported in USD/MJ and hours/MJ. We study 220 public GDPval tasks across 44 occupations, serving five open-weight models locally and evaluating four accelerator types. First, the occupation groups where AI systems succeed most often, deliver the greatest dollar value, or save the most time do not necessarily provide the greatest economic return per joule. For example, among the tested tasks, engineering work delivers about 10% more net dollar value per task than software development. Yet software development delivers about 27% more net value per joule. Second, GLM 5.2 delivers the highest average net dollar value, at $72.27 per task using 0.430 MJ, equivalent in energy to about four hours of active laptop use. By contrast, Gemma 4 31B delivers 56% of that net dollar value using just 23% of the energy, achieving approximately 2.5 times the NVJ and 2.8 times the TSJ. Third, across four single-accelerator deployments, Qwen3.6-27B's NVJ varies by a factor of 4.32 and TSJ by 9.27, with B200 leading both. Additionally, one and eight H100s deliver similar net dollar value, but eight reduce completion time by 22% at 3.24 times the energy. Our framework makes output value and expert time saving under an energy budget explicit objectives for improving models, accelerators, and software, while offering a new basis for assessing AI's potential economic effect on occupational work.

Silicon sampling generates survey respondents by having a large language model (LLM) answer a survey as a person with given demographic attributes, and the resulting synthetic populations are increasingly used in place of human ones. Existing evaluations focus on individual-answer accuracy, per-question answer distributions, and correlations across questions. None of these compares the respondents themselves, so a generator whose respondents have combinations of answers found in no human respondent can still score well if it reproduces every answer distribution. We propose the respondent coverage gap (RCG), which represents each respondent as a point defined by their complete answer profile and measures how far each human respondent is from the nearest generated one, relative to the distance between human samples. On two US social surveys, all six LLMs we evaluate leave human respondents uncovered, and under our main settings, they fall behind a baseline that samples each answer independently from the human marginals. Across the prompting and inference settings we test, every LLM has a higher RCG than a Gaussian copula fitted to the human data.

Diffusion language models (DLMs) decode multiple tokens in parallel, so individually valid tokens can be incompatible together under a schema. Exact joint inference over a grammar graph enforces compatibility, but it requires token-labeled successors across many hypothetical parser configurations. Encoding stacks and field histories in compiled states can cause exponential growth, which forces prior exact decoders to fix property orders and exclude valid outputs. We present DConstrain, which compiles schemas into context-parameterized transition programs. Cold compilation specializes token consumption to local grammar states while retaining context-dependent stack operations, environment guards, and caller-dependent deferred parsing symbolically. Tracking field histories as runtime values supports arbitrary property order without enumerating field subsets offline. At runtime, warm compilation binds reached configurations to these programs to construct exact successors, overlapping with the model forward pass because graph construction requires no model scores. Furthermore, token-effect grouping merges tokens with identical compiled effects before runtime contexts are bound, so groups are reused across stacks and environments even when concrete successors differ. Objective-specific score aggregation and token restoration provably preserve both the Viterbi optimum and the token-labeled path distribution of ungrouped inference. Across three benchmarks and two DLM families, DConstrain compiles 100% of BFCL schemas under arbitrary property orders, whereas an optimized finite-state baseline compiles only 71.30-85.81%. Under arbitrary property order, it reduces BFCL end-to-end decoding latency by 6.65-46.26 times with 100% schema validity and raises DiffusionGemma's BFCL accuracy from 37.27% under fixed order to 81.56%.

Ophthalmic clinical decision support requires fast, accurate, and privacy-preserving inference, yet autoregressive large language models (LLMs) incur substantial latency and computational overhead for decisions among predefined actions. We therefore investigate a decision-oriented alternative to generative reasoning. Specifically, we first introduce RetiJev-Data, a source-grounded ophthalmology corpus comprising 78,158 clinical-decision and diagnosis questions derived from articles in PubMed Central and a MeSH vocabulary covering 531 eye diseases, and 50,074 guideline-grading questions spanning 11 grading systems, along with a 24,662-question evaluation benchmark. Based on this, we develop RetiJev, a family of 0.8B-9B decision models that directly score candidate answers without autoregressive decoding. Across all model sizes, RetiJev outperforms both a 27B LLM and the commercial Jev 1.13 decision model on guideline grading by 10.4-23.0 points. RetiJev-9B further achieves the highest accuracy on clinical decisions, improving over both baselines by 5.4-6.5 points while matching them on diagnosis. It answers a case in 158 ms on one GPU, 3.5x faster than LLM generation and 470x faster than reasoning. Models are released at https://github.com/Awenbocc/RetiJev.

Content warning: this paper quotes slurs and hateful statements from published test suites; the racial slur among them is masked. Automated systems decide at scale which messages are harmful, and they decide unevenly. Classifiers label African American English as offensive up to twice as often as other English, the accounts of prominent drag queens can score as more toxic than those of white nationalists, and safety behaviour that holds in English collapses in low-resource languages. We argue that over-flagging and under-protection share one cause: deciding harm from vocabulary and surface form rather than from what a message does. We organise the evidence into seven failure modes and argue, as testable hypotheses, that harmful behaviours (a request for secrecy, a demand for payment, an attempt to isolate someone) travel across languages better than the words that express them. We propose an evaluation protocol that makes the gap between the best- and worst-served language communities a headline result, and show that the naive gap is inflated under equal service and that its usual bootstrap interval is unreliable. Applied to five open systems in eleven languages, from a keyword lexicon to two current multilingual guard models, the protocol finds that none reaches native-speaker accuracy at separating harmful function from harmless form (51 of 55 system-language pairs fall short); that greater capability raises discrimination in nine languages while leaving Arabic and romanised Hindi near chance, so the gap between the best- and worst-served languages grows with capability; that one system both over-flags and under-protects the same language most; and that a language a system reads worst can look the most fairly treated when only false positives are reported.

LLM-as-a-Judge evaluation is increasingly applied to multilingual tasks, yet prompting strategies such as Chain-of-Thought and Role Conditioning are recommended based solely on English benchmarks. Whether their effectiveness transfers across judge models and languages remains unknown: prompt-sensitivity studies are largely English-centric, while existing multilingual judge studies fix the prompt template. In this work, we address this gap by measuring this interaction through a benchmark of 9 prompting strategies, 4 open-weight judges, 27 languages, and 4 tasks across 13 datasets, yielding over 1M pointwise scores on gold-reference outputs, validated against human judgments. We find that prompting strategies are not transferable: the choice of the judge model dominates the choice of prompting strategy, strategy effects reverse direction across judges, and stability degrades in low-resource languages, where these effects are concealed by aggregate reporting. These findings show that prompting effects in multilingual evaluation are conditioned on the judge, task, and language. Consequently, prompting strategies should be validated for the intended evaluation setting, and cross-study comparisons should account for both judge and prompt configuration.

Users increasingly rely on AI systems for extended, multi-turn tasks, yet evidence indicates that LLM-based system performance degrades over prolonged interaction. To evaluate and improve AI assistants in such settings, we need user simulators that faithfully follow complex user intents while exhibiting realistic conversational behavior. Existing approaches manifest a trade-off: prompted assistants can follow complex user intents but lack behavioral fidelity, while specialized user simulators better approximate conversational behavior but struggle to remain faithful to complex intents. In this paper, we propose a reinforcement learning (RL) recipe that jointly rewards task intent adherence and behavioral naturalness. A crucial contribution of the method is operationalizing naturalness with a lightweight discriminator trained in an alternating, adversarial loop with the user simulator. We release the resulting model, UserLM-2.0, which produces valid simulated conversations (i.e., meeting both task-faithfulness and naturalness criteria) 46.2% of the time, over three times the rate of the strongest prior user simulator (14.1%).

Looped architectures scale quality while keeping parameter count fixed by repeatedly passing activations through the same block of layers. During inference, each recurrence generates independent KV-caches, causing memory to grow linearly. To address this challenge, prior work has proposed KV-sharing (i.e., sharing KV-caches across depth), finding that, in some cases, the KV-cache of the final recurrence can be shared with the earlier recurrences without quality loss, a property we denote as KV-shareability. However, the mechanisms underlying KV-shareability remain poorly understood, with even identically trained looped architectures not consistently exhibiting it. In this paper, we study how KV-shareability can be enabled in looped models. We propose a two-stage method: in the first stage, we use stochastic depth during pretraining, which induces contractive residual dynamics that drive representations toward a common limit across recurrences. In the second stage, we propose BeNice!, a post-training method that broadcasts the final recurrence's KV-cache into the first recurrence, training it to produce compatible residuals while preserving the original model's quality. As a result, BeNice! enables KV-shareability while matching standard finetuning's unshared validation perplexity. Using BeNice!, we show that KV-sharing is complementary to alternative KV-compression methods, composing with memory-efficient sequence mixers (e.g., multi-latent attention, Kimi delta attention) and low-precision KV-cache quantization (e.g., fp8, mxfp8, nvfp4). Finally, in simulated low-memory, long-context inference, KV-sharing improves cache hit rates by up to 31.4%, increases throughput by up to 23.6%, and reduces p90 time-to-first-token by up to 83.6%.

Mixture-of-experts (MoE) models often exceed a single accelerator's memory capacity. Offloading experts to host memory creates a severe PCIe I/O bottleneck because data-dependent routing keeps transfers on the critical path. We present Specter, an inference system that uses quantization to combine speculative decoding with expert offloading. Its lightweight quantized draft model runs efficiently on device while preserving high-fidelity expert activation patterns. This provides accurate multi-step lookahead over future expert usage, allowing the runtime to organize expert movement and execution ahead of verification and improve both communication and execution efficiency. Under single-request inference, Specter achieves an average 1.50x (up to 1.88x) speedup in time per output token (TPOT) over the state-of-the-art system. Specter makes memory-constrained, single-GPU MoE inference more practical on commodity hardware. Our code is available at https://github.com/wwwwwf/specter_asplos.