Daily Papers

LLM judges are widely used to evaluate agentic tool-calling systems, yet their reliability on structured, dependency-driven workflows remains largely unexamined. We present AgentJudgeBench, the first benchmark to systematically study LLM-as-a-judge reliability for agentic tool-calling over workflow DAGs, as distinct from the broader LLM-as-a-judge task of open-ended text or preference evaluation. The benchmark comprises 3,808 instances spanning six DAG topologies and three difficulty tiers, evaluated with five generators (3B-70B open-weight models and GPT-5.4) and six judges (20B to frontier scale) under paired with- and without-ground-truth conditions. Judge alignment degrades monotonically with task difficulty, 1.5x faster without ground truth, and on hard queries without ground truth all six judges converge to a narrow 77-82% band regardless of scale, revealing a structural ceiling driven primarily by task difficulty, though its height is partly prompt-dependent for weaker generators, that model capacity alone cannot overcome. Ground-truth exposure is not uniformly beneficial: it reduces alignment for GPT-5.4 (1.5 pp) and Gemini-2.5-Pro (3.9 pp), consistent with over-anchoring. Among mitigation strategies, chain-of-thought reasoning and judge temperature both have negligible effect, while structured evaluation rubrics improve alignment by up to 6.5 pp but do not generalize uniformly across judge-generator pairs. With ground truth, QwQ-32B best matches the programmatic reference, while a human validation study identifies GPT-OSS-120B as the most human-aligned judge; without it, frontier judges lead only marginally within the shared ceiling. These results expose fundamental limitations of current LLM judges and yield practical guidelines for reliable evaluation in agentic systems.

Misinformation and disinformation demand fact checking that goes beyond simple evidence-based reasoning. Existing benchmarks fall short: they are largely single modality (text-only), span short time horizons, use shallow evidence, cover domains unevenly, and often omit full articles -- obscuring models' real-world capability. We present MMM-Fact, a large-scale benchmark of 125,449 fact-checked statements (1995--2025) across multiple domains, each paired with the full fact-check article and multimodal evidence (text, images, videos, tables) from four fact-checking sites and one news outlet. To reflect verification effort, each statement is tagged with a retrieval-difficulty tier -- Basic (1--5 sources), Intermediate (6--10), and Advanced (>10) -- supporting fairness-aware evaluation for multi-step, cross-modal reasoning. The dataset adopts a three-class veracity scheme (true/false/not enough information) and enables tasks in veracity prediction, explainable fact-checking, complex evidence aggregation, and longitudinal analysis. Baselines with mainstream LLMs show MMM-Fact is markedly harder than prior resources, with performance degrading as evidence complexity rises. MMM-Fact offers a realistic, scalable benchmark for transparent, reliable, multimodal fact-checking.

Current Vision-Language Models (VLMs) fail at quantitative spatial reasoning because their architectures destroy pixel-level information required for counting and measurements. Vision encoders compress images through patch embeddings, reducing spatial indexing and losing the precise pixel-level tracking required for accurate counting. We present two contributions to address this fundamental limitation. First, we introduce SQuID (Satellite Quantitative Intelligence Dataset), a benchmark of 2,000 satellite image Question-Answer pairs with both numerical range and categorical answers, designed to evaluate quantitative spatial reasoning. The dataset spans three difficulty tiers with annotations automatically generated from human labels and their learned variability. Second, we propose QVLM (Quantitative Vision-Language Model), a code-generation architecture that maintains pixel precision by decoupling language understanding from visual analysis. Instead of encoding images into embeddings, QVLM generates executable code that first calls a segmentation model to obtain pixel-level masks, then operates directly on these masks, preserving spatial indexing throughout the reasoning process. Our experiments show that QVLM using GPT-5 as coder achieves 42.0% accuracy on SQuID compared to 28.1% for a VLM prompted with image-question pairs. Our work reveals that, for quantitative spatial reasoning, architectural decoupling enables better accuracy on quantitative tasks.

In real-world data science and enterprise decision-making, critical information is often fragmented across directly queryable structured sources (e.g., SQL, CSV) and "zombie data" locked in unstructured visual documents (e.g., scanned reports, invoice images). Existing data analytics agents are predominantly limited to processing structured data, failing to activate and correlate this high-value visual information, thus creating a significant gap with industrial needs. To bridge this gap, we introduce DataCross, a novel benchmark and collaborative agent framework for unified, insight-driven analysis across heterogeneous data modalities. DataCrossBench comprises 200 end-to-end analysis tasks across finance, healthcare, and other domains. It is constructed via a human-in-the-loop reverse-synthesis pipeline, ensuring realistic complexity, cross-source dependency, and verifiable ground truth. The benchmark categorizes tasks into three difficulty tiers to evaluate agents' capabilities in visual table extraction, cross-modal alignment, and multi-step joint reasoning. We also propose the DataCrossAgent framework, inspired by the "divide-and-conquer" workflow of human analysts. It employs specialized sub-agents, each an expert on a specific data source, which are coordinated via a structured workflow of Intra-source Deep Exploration, Key Source Identification, and Contextual Cross-pollination. A novel reReAct mechanism enables robust code generation and debugging for factual verification. Experimental results show that DataCrossAgent achieves a 29.7% improvement in factuality over GPT-4o and exhibits superior robustness on high-difficulty tasks, effectively activating fragmented "zombie data" for insightful, cross-modal analysis.

The rise of autonomous AI agents suggests that dynamic benchmark environments with built-in feedback on scientifically grounded tasks are needed to evaluate the capabilities of these agents in research work. We introduce Stargazer, a scalable environment for evaluating AI agents on dynamic, iterative physics-grounded model-fitting tasks using inference on radial-velocity (RV) time series data. Stargazer comprises 120 tasks across three difficulty tiers, including 20 real archival cases, covering diverse scenarios ranging from high-SNR single-planet systems to complex multi-planetary configurations requiring involved low-SNR analysis. Our evaluation of eight frontier agents reveals a gap between numerical optimization and adherence to physical constraints: although agents often achieve a good statistical fit, they frequently fail to recover correct physical system parameters, a limitation that persists even when agents are equipped with vanilla skills. Furthermore, increasing test-time compute yields only marginal gains, with excessive token usage often reflecting recursive failure loops rather than meaningful exploration. Stargazer presents an opportunity to train, evaluate, scaffold, and scale strategies on a model-fitting problem of practical research relevance today. Our methodology to design a simulation-driven environment for AI agents presumably generalizes to many other model-fitting problems across scientific domains. Source code and the project website are available at https://github.com/Gudmorning2025/Stargazer and https://gudmorning2025.github.io/Stargazer, respectively.

Cancer screening is a reasoning task. A radiologist observes findings, compares them to prior scans, integrates clinical context, and reaches a diagnostic conclusion confirmed by pathology. We present RadThinking, a Visual Question Answering (VQA) dataset that makes this reasoning explicit and trainable. RadThinking releases VQA pairs at three difficulty tiers. Foundation VQAs are atomic perception questions. Single-step reasoning VQAs apply one clinical rule. Compositional VQAs require multi-step chain-of-thought to reach a guideline category such as LI-RADS-5. For every compositional VQA, we release the chain of foundation VQAs that solves it. The chain follows the rules of the governing clinical reporting standard. The dataset spans 20,362 CT scans from 9,131 patients across 43 cancer groups, plus 2,077 verified healthy controls with >1-year follow-up. To our knowledge, RadThinking is the first cancer-screening VQA corpus that stratifies questions by reasoning depth and grounds compositions in clinical reporting standards. The foundation tier supplies atomic perception supervision. The compositional tier supplies chain-of-thought data and verifiable rewards for reinforcement-learning recipes such as DeepSeek-R1 and OpenAI o1. RadThinking enables systematic training and evaluation of whether AI systems can reason about cancer, not merely detect it.

Autonomous agents are increasingly expected to support end-to-end medical-AI research workflows, moving beyond isolated prediction tasks or short-form clinical question answering. However, existing medical agent benchmarks primarily evaluate final outputs, providing limited visibility into agent behavior within the research process. To address this gap, we present AutoMedBench, a workflow-aware benchmark for autonomous medical-AI research across diverse medical imaging and multimodal inference tasks, organizing agent execution into a unified five-stage workflow (S1-S5): Plan, Setup, Validate, Inference, and Submit. It comprises long-horizon tasks with each run averaging 33 agent turns, spanning five research tracks: segmentation, image enhancement, visual question answering (VQA), report generation, and lesion detection. Each task is evaluated under two difficulty tiers, Lite and Standard, which use the same data and metrics but differ in the amount of task-brief scaffolding, and each run is scored using both final task performance and S1-S5 stage scores, enabling stage-level analysis from the initial task brief to the final submitted artifact. Across thousands of recorded runs, stage-level scoring reveals that Validate is the weakest workflow stage on average, whereas Setup is the strongest, suggesting that current agents are better at making pipelines executable than at verifying their reliability. Post-run error analysis further shows that verification and submission failures dominate tagged errors, accounting for 37.7% and 38.1% of fired codes respectively, whereas task-understanding errors are rare at 0.9%, and runs with one fired error code have a 48% lower overall score than runs with no error code on average.

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

In recent years, the rapid development of large reasoning models has resulted in the saturation of existing benchmarks for evaluating mathematical reasoning, highlighting the urgent need for more challenging and rigorous evaluation frameworks. To address this gap, we introduce OlymMATH, a novel Olympiad-level mathematical benchmark, designed to rigorously test the complex reasoning capabilities of LLMs. OlymMATH features 200 meticulously curated problems, each manually verified and available in parallel English and Chinese versions. The problems are systematically organized into two distinct difficulty tiers: (1) AIME-level problems (easy) that establish a baseline for mathematical reasoning assessment, and (2) significantly more challenging problems (hard) designed to push the boundaries of current state-of-the-art models. In our benchmark, these problems span four core mathematical fields, each including a verifiable numerical solution to enable objective, rule-based evaluation. Empirical results underscore the significant challenge presented by OlymMATH, with state-of-the-art models including DeepSeek-R1 and OpenAI's o3-mini demonstrating notably limited accuracy on the hard subset. Furthermore, the benchmark facilitates comprehensive bilingual assessment of mathematical reasoning abilities-a critical dimension that remains largely unaddressed in mainstream mathematical reasoning benchmarks. We release the OlymMATH benchmark at the STILL project: https://github.com/RUCAIBox/Slow_Thinking_with_LLMs.

In this paper, we present a holistic multimodal benchmark that evaluates the reasoning capabilities of MLLMs with an explicit focus on reasoning width, a complementary dimension to the more commonly studied reasoning depth. Specifically, reasoning depth measures the model's ability to carry out long-chain, sequential reasoning in which each step is tightly and rigorously linked to the next. Reasoning width tends to focus more on the model's capacity for broad trial-and-error search or multi-constrained optimization: it must systematically traverse many possible and parallelized reasoning paths, apply diverse constraints to prune unpromising branches, and identify valid solution routes for efficient iteration or backtracking. To achieve it, we carefully curate 1200+ high-quality multimodal cases spanning heterogeneous domains, and propose a fine-grained tree-of-thought evaluation protocol that jointly quantifies reasoning width and depth. We evaluate 12 major model families (over 30 advanced MLLMs) across difficulty tiers, question types, and required skills. Results show that while current models exhibit strong performance on general or common-sense VQA tasks, they still struggle to combine deep sequential thought chains with wide exploratory search to perform genuine insight-based reasoning. Finally, we analyze characteristic failure modes to provide possible directions for building MLLMs that reason not only deeper but also wider.

Existing medical-agent benchmarks deliver imaging as pre-selected samples, never as an environment the agent must navigate. We introduce ABRA, a radiology-agent benchmark in which the agent operates an OHIF viewer and an Orthanc DICOM server through twenty-one function-calling tools that span slice navigation, windowing, series selection, pixel-coordinate annotation, and structured reporting. ABRA contains 655 programmatically generated tasks across three difficulty tiers and eight types (viewer control, metadata QA, vision probe, annotation, longitudinal comparison, BI-RADS reporting, and oracle variants of annotation and BI-RADS reporting), drawn from LIDC-IDRI, Duke Breast Cancer MRI, and NLST New-Lesion LongCT. Each episode is scored along Planning, Execution, and Outcome (Bluethgen et al., 2025) by task-type-specific automatic scorers. Ten current models, five closed-weight and five open-weight, reach at least 89% Execution on real annotation but only 0-25% Outcome; on the paired oracle variant where a simulated detector supplies the finding, Outcome on the same task reaches 69-100% across the models evaluated, localising the bottleneck to perception rather than tool orchestration. Code, task generators, and scorers are released at https://github.com/Luab/ABRA

Web-agent benchmarks overwhelmingly measure depth -- pinning one obscure answer behind a chain of constraints -- while breadth, exhaustively enumerating a closed set and filling each item's attributes, is barely evaluated, especially outside English. Breadth is also hard to build: certifying that a gold set is complete and every cell correct is far costlier than checking a single answer. I introduce Ko-WideSearch, a Korean breadth-search benchmark built by an automated synthesize-and-verify pipeline. Each task names a set-parent entity -- a TV season, a dynasty, a league, an administrative region, an election -- and asks for its full membership plus a per-item attribute table, graded by Item-, Column-, and Row-F1. It spans 228 tables over 190 entities and sixteen categories across three difficulty tiers, set by two structural knobs I dial independently -- table width and a 2-D composite key -- so cross-product membership climbs from 0\% to 100\% across the tiers. A single normalization-aware comparator is shared between gold construction and grading, so stable date and count columns are not over-dropped on formatting alone. Across twenty web agents, the failure is consistent: agents recover the set but not the rows (e.g.\ Item-F1 92.8 against Row-F1 53.7), accuracy falls steadily as the knobs harden, and neither more search nor more spend closes the gap. Broken down by cell, the hard part is finding the right value, not formatting it: open-ended free-text cells fail most, while cells with a standard answer such as a date or a name usually come out right.

Large language model (LLM) agents extend generative models with reasoning, tool use, and persistent memory, thereby enabling the automation of complex tasks. In healthcare, such systems could support documentation, care coordination, and clinical decision making. Their reliable deployment in hospitals, however, remains constrained by safety risks, limited transparency, and inadequate mechanisms for handling longitudinal clinical context. Here we propose an architecture that adapts LLM agents to hospital environments. The design comprises four components: a restricted execution environment inspired by multi-user operating systems, a document-centric interaction model linking patient and clinician agents, a page-indexed memory architecture for longitudinal context management, and a curated library of composable medical skills. Implemented on top of OpenClaw, an open-source agent orchestration framework, this design provides the basis for an Agentic Operating System for Hospitals: a computing layer for coordinating clinical workflows while preserving safety, transparency, and auditability. To evaluate the memory component, we introduce manifest-guided retrieval for hierarchical navigation of longitudinal patient records. In a benchmark derived from the MIMIC-IV dataset (v2.2) comprising 100 de-identified patient records and 300 clinical queries stratified across three difficulty tiers (100 per tier), manifest-guided retrieval matched a metadata-filtered RAG baseline on overall recall (0.877 versus 0.876) while achieving 2.2x higher precision (0.779 versus 0.352) and retrieving fewer documents; on tier-3 longitudinal queries, manifest recall was 21% higher (0.846 versus 0.701), confirming that LLM-guided hierarchical navigation is most valuable when queries span multiple care episodes. These results outline a practical path toward hospital-scale agentic infrastructure.

While current Computer Use Agent (CUA) benchmarks measure task completion effectively, they provide limited assessment of enterprise deployment readiness, emphasizing functional correctness over the operational reliability required for production systems. We present UI-CUBE (UiPath Computer Use BEnchmark), a systematic benchmark comprising 226 tasks across two difficulty tiers designed to expose fundamental architectural limitations in current CUAs. Our evaluation covers simple UI interactions (136 tasks) and complex workflows including copy-paste tasks (50 tasks) and enterprise application scenarios (40 tasks), with systematic interface variation coverage, multi-resolution testing and automated validation of task success through the application state. Evaluation of five state-of-the-art models reveals a sharp capability cliff rather than gradual performance degradation. Simple UI interactions achieve 67-85% success rates (compared to 97.9% human performance), but complex workflows drop precipitously to 9-19%. Human evaluators with no prior application experience achieve only 61.2% on complex tasks despite near-perfect performance on simple tasks, establishing realistic performance ceilings. This discontinuous performance pattern -- where agents achieve 68-87% of human performance on simple tasks but only 15-32% on complex workflows -- indicates fundamental architectural limitations in memory management, hierarchical planning, and state coordination rather than incremental capability gaps addressable through better training or prompting. UI-CUBE functions as an enterprise-readiness diagnostic, revealing that while current CUAs can manipulate individual interface elements, they cannot yet function as reliable workflow automation tools. These findings provide architectural insights essential for developing production-ready CUAs capable of managing complex, multi-step enterprise processes.

Engineering construction automation aims to transform natural language specifications into physically viable structures, requiring complex integrated reasoning under strict physical constraints. While modern LLMs possess broad knowledge and strong reasoning capabilities that make them promising candidates for this domain, their construction competencies remain largely unevaluated. To address this gap, we introduce BuildArena, the first physics-aligned interactive benchmark designed for language-driven engineering construction. It contributes to the community in four aspects: (1) a highly customizable benchmarking framework for in-depth comparison and analysis of LLMs; (2) an extendable task design strategy spanning static and dynamic mechanics across multiple difficulty tiers; (3) a 3D Spatial Geometric Computation Library for supporting construction based on language instructions; (4) a baseline LLM agentic workflow that effectively evaluates diverse model capabilities. On eight frontier LLMs, BuildArena comprehensively evaluates their capabilities for language-driven and physics-grounded construction automation. The project page is at https://build-arena.github.io/.

Evaluating the quality of tables generated by large language models (LLMs) remains an open challenge: existing metrics either flatten tables into text, ignoring structure, or rely on fixed references that limit generalization. We present TabReX, a reference-less, property-driven framework for evaluating tabular generation via graph-based reasoning. TabReX converts both source text and generated tables into canonical knowledge graphs, aligns them through an LLM-guided matching process, and computes interpretable, rubric-aware scores that quantify structural and factual fidelity. The resulting metric provides controllable trade-offs between sensitivity and specificity, yielding human-aligned judgments and cell-level error traces. To systematically asses metric robustness, we introduce TabReX-Bench, a large-scale benchmark spanning six domains and twelve planner-driven perturbation types across three difficulty tiers. Empirical results show that TabReX achieves the highest correlation with expert rankings, remains stable under harder perturbations, and enables fine-grained model-vs-prompt analysis establishing a new paradigm for trustworthy, explainable evaluation of structured generation systems.

Vision-language models (VLMs) have made strong progress on high-level image-text alignment, yet their ability to perceive subtle visual differences remains limited. We study this problem in rendered web interfaces, where localized visual changes are both a diagnostic test of fine-grained perception and a practical requirement for GUI agents and design tools. We introduce DiffSpot, a code-driven benchmark for open-ended spot-the-difference on web interfaces. DiffSpot constructs controlled image pairs by mutating a single CSS property of a target element in self-contained HTML, re-rendering the page, and recording the changed property, element, and mutation magnitude. A grounding gate retains only pairs whose rendered pixel difference is confined to the target element. The benchmark contains 4{,}400 pairs, including 3{,}900 has-diff pairs balanced across 13 CSS-property operators and three difficulty tiers, plus 500 no-diff pairs for hallucination control. Evaluating 13 frontier VLMs zero-shot, we find that even the best model identifies only 40.7% of true changes, with Hard-tier Recall below 23% for every model. DiffSpot further shows that difficulty is strongly property-dependent: across CSS operators, neither pixel magnitude nor CLIP distance reliably predicts Recall.

Fitting quantitative models to data is a central step in scientific workflows, yet it remains one of the least automated. Recent agent-based systems leverage language and vision-language models (VLMs) to iteratively propose and refine statistical models, but these systems struggle on more challenging modeling tasks. To address these limitations, we introduce VESTA: Visual Exploration with Statistical Tool Agents, a framework that equips VLMs with a dynamically growing exploration toolkit to guide model refinement through data transformations, hypothesis-driven visualizations, and robust statistical tests. Unlike prior systems that rely on iterative critique alone, VESTA actively explores data before and during refinement by selecting or creating diagnostic tools, which accumulate in the model's context and can be reused later. We evaluate VESTA against established baselines in three toolkit configurations: no tools, static expert-written tools, and dynamic model-written tools. To support this evaluation, we introduce DAWN (Dataset for Automated Workflows and Numerical Modeling), a benchmark targeting distribution fitting and time series modeling with varying difficulty tiers, and culminating in real-world astronomy tasks including modeling initial mass functions and gravitational-wave chirp signals. We find that VESTA's dynamic tool creation outperforms prior agentic pipelines, with the largest gains on complex and domain-specific tasks. We further show that dynamically generated tools are substantially more sophisticated than those produced by existing visual tool-creation systems, covering more diagnostic categories per function and strongly preferring visual outputs that the VLM critic can reason over directly.

Front-end web code has become a core product surface for every frontier LLM release, yet evaluating these interactive applications at development speed remains costly because human-judged leaderboards like Arena do not scale. Existing automated proxies typically lean on reference implementations, test suites, or rigid checklists, and tend to miss the reasoned synthesis a human reviewer performs over a live session. We articulate a new evaluation regime that is simultaneously reference-free, autonomously driven, and holistically reasoned, and instantiate it through two artifacts. \dataname is an 11-domain, 54-leaf, 1,000-query WebDev benchmark spanning both static-presentation and interactive-application tasks, balanced across three difficulty tiers and three target-language groups, with briefs rewritten to resist recall from circulated prompts. \framename, grounded in Flavell's metacognitive monitoring, separates evidence accumulation from judgment across three stages: Static Perception forms a first impression from passive observation; Agent-Driven Interaction explores the application autonomously while capturing continuous screen video, audio, and per-step screenshots; Dynamic Scoring issues holistic functionality and aesthetics verdicts with structured failure attribution only after the evidence chain is complete. On \dataname, \framename aligns closely with expert human ratings while surfacing substantial headroom across 13 frontier LLMs on interactive web generation. \noindenthttps://anonymous.4open.science/r/Cookie-3CE/

Multilingual fluency often invites a stronger assumption: a model that can speak a user's language must also understand the culture encoded by that language. We call this the Illusion of Cultural Alignment. To test this assumption directly, we introduce MSQA, a benchmark of 1,064 natively sourced questions across 11 language groups, five cultural dimensions, and three difficulty tiers. Unlike translated benchmarks, MSQA targets locally grounded knowledge and reduces shortcuts from English-centric cross-lingual transfer. Evaluating 18 LLMs, we find substantial cultural degradation and a pronounced Locality Effect: cultural competence tracks pre-training exposure more closely than general reasoning ability. We further show that common inference-time remedies do not dissolve the illusion. Models remain overconfident on unfamiliar cultural questions, repeated sampling yields unstable rather than reliable correctness, and retrieval augmentation helps unevenly on long-tail facts. These findings indicate that cultural alignment cannot be inferred from multilingual ability alone and requires deeper intervention than calibration, sampling, or retrieval at inference time

We introduce the Self-Evolving Agentic Operating System (SE-AOS): a new class of AI agent that treats exploit capability as a mutable, versioned kernel it extends at runtime, observing its own failures, synthesising new capabilities, proving them against a live target, and hot-loading them back into itself. Mako is the first SE-AOS instance for security research and the autonomous web exploitation engine developed within LaunchSafe. LaunchSafe builds autonomous security agents for continuous offensive testing and agent-driven security research; Mako is the core engine behind that platform. On the public XBOW validation-benchmarks, 104 containerised, CTF-style web applications spanning 26 vulnerability classes across three difficulty tiers, Mako achieves full-suite coverage: it drives every one of the 104 targets to emit a cryptographically fresh, per-build flag, under a verification regime that makes fabricated or memorised results impossible. Our central result is a law of autonomous exploitation: once a capability exists and is discoverable, difficulty collapses; capability, not reasoning, is what is scarce, together with an architecture and formalism that turn that law into a self-improving system. Mako further runs a gated self-evolution loop that proposes, sandboxes, and commits improvements to its own agents and rules when fitness does not regress. We deliberately withhold the operational results, payloads, exploit chains, and tool source, because a system that reduces full-spectrum web exploitation to a repeatable, machine-speed pipeline is dual-use research of concern. We publish the science; we withhold the weapon.

LLM-based coding agents have advanced rapidly on single-process SWE tasks, with frontier models now clustering in the high-70s on SWE-bench Verified. Distributed-system debugging, however, remains an under-explored regime: bugs span processes, nodes, and protocol interactions, with root causes rarely recoverable from source alone and brute-force exploration intractable across non-deterministic interleavings. This leaves two gaps in LLM and agent evaluation: no code-repair benchmark targets distributed-system bugs, and no controlled study isolates how much externally provided debugging context changes agent success on them. We introduce DDBench, a code-repair benchmark of 60 historical bugs mined from 13 open-source distributed systems, partitioned into three difficulty tiers. DDBench evaluates every case under two matched conditions: a symptom-only condition where the agent receives only the bug symptom and repository, and a context-augmented condition where it additionally receives a bounded debugging context (logs, traces, runtime state, and targeted code-investigation notes), isolating the effect of debugging context from model capability. The evaluation of ten LLMs on DDBench reveals several findings. First, distributed debugging exercises a reasoning dimension that single-process benchmarks do not surface: models' pass rates span 61 pp, and pairwise bootstrap separates 9 of 15 top-tier model pairs at p < 0.05 on DDBench's hardest case-set. Second, bounded debugging context lifts aggregate pass rate by +18.1 pp, and the lift is asymmetric: weaker models gain pass rate, while stronger models gain efficiency. Third, debugging context requires careful curation, as even faithful debugging context can sometimes mislead LLMs.

Graphical user interfaces have emerged as an important environment for evaluating autonomous AI agents on multimodal interactive tasks. Existing benchmarks such as AndroidWorld and MobileWorld provide strong foundations for mobile agent evaluation, but their application coverage and task design do not yet fully capture the diversity and complexity of realistic mobile use. We present GMA, a benchmark for evaluating general mobile assistants in challenging real-world scenarios. GMA introduces seven applications based on open-source projects, spanning domains such as lifestyle sharing and travel planning, and 300 tasks across four difficulty tiers, from atomic actions to complex multi-step workflows. We evaluate eight frontier models and find that performance declines substantially as task complexity increases, with current agents remaining far from reliably handling realistic user requirements. We further conduct controlled ablation studies of agent harness choices, including context retention and explicit state tracking, under a shared environment, model setting, and task taxonomy. Results show that appropriate harness design can meaningfully improve performance, particularly on demanding workflows, while the effectiveness of specific designs can vary across foundation models. Overall, GMA complements existing benchmarks by expanding application coverage and task complexity, providing a challenging testbed for evaluating mobile agents and studying how harness design supports reliable execution in complex mobile workflows.

We present SPL (Structured Prompt Language), a declarative language that composes deterministic and probabilistic computation modes in a single specification. While existing frameworks separate these -- orchestration systems (AutoGen, CrewAI, LangGraph) for LLM calls, symbolic tools (SymPy, SageMath, Lean) for computation -- SPL unifies them. It provides GENERATE/EVALUATE for probabilistic computation and SOLVE/ASSERT for deterministic computation, sharing syntax, variable bindings, and runtime routing. A .spl specification runs unchanged across local nodes (Ollama), cloud APIs (OpenRouter, Anthropic), and distributed grids (Momagrid), with model and verifier selection deferred to invocation time. We validate SPL through an extensive 78-recipe cookbook and a controlled 1,200-run experiment (10 models x 20 problems x 2 arms x 3 repetitions; the 20 problems span 6 difficulty tiers). The solver arm achieves 82-93% machine-verified correctness (sonnet-4-6: 85%, gemma4:e2b: 93%) while the LLM-only arm measures output production without mathematical verification, making the comparison one of verified correctness against unverified fluency. A backend difficulty gradient emerges (SymPy 78%, Sage 54%), and the dominant failure mode is solver_error (kernel-rejected expressions), not format non-compliance.

Interactive video generation models such as Genie, YUME, HY-World, and Matrix-Game are advancing rapidly, yet every model is evaluated on its own benchmark with private scenes and trajectories, making fair cross-model comparison impossible. Existing public benchmarks offer useful metrics such as trajectory error, aesthetic scores, and VLM-based judgments, but none supplies the standardized test conditions -- identical scenes, identical action sequences, and a unified control interface -- needed to make those metrics comparable across models with heterogeneous inputs. We introduce WorldMark, the first benchmark that provides such a common playing field for interactive Image-to-Video world models. WorldMark contributes: (1) a unified action-mapping layer that translates a shared WASD-style action vocabulary into each model's native control format, enabling apples-to-apples comparison across six major models on identical scenes and trajectories; (2) a hierarchical test suite of 500 evaluation cases covering first- and third-person viewpoints, photorealistic and stylized scenes, and three difficulty tiers from Easy to Hard spanning 20-60s; and (3) a modular evaluation toolkit for Visual Quality, Control Alignment, and World Consistency, designed so that researchers can reuse our standardized inputs while plugging in their own metrics as the field evolves. We will release all data, evaluation code, and model outputs to facilitate future research. Beyond offline metrics, we launch World Model Arena (warena.ai), an online platform where anyone can pit leading world models against each other in side-by-side battles and watch the live leaderboard.

In industrial procurement, an LLM answer is useful only if it survives a standards check: recommended material must match operating condition, every parameter must respect a regulated threshold, and no procedure may contradict a safety clause. Partial correctness can mask safety-critical contradictions that aggregate LLM benchmarks rarely capture. We introduce IndustryBench, a 2,049-item benchmark for industrial procurement QA in Chinese, grounded in Chinese national standards (GB/T) and structured industrial product records, organized by seven capability dimensions, ten industry categories, and panel-derived difficulty tiers, with item-aligned English, Russian, and Vietnamese renderings. Our construction pipeline rejects 70.3% of LLM-generated candidates at a search-based external-verification stage, calibrating how unreliable industrial QA remains after LLM-only filtering.Our evaluation decouples raw correctness, scored by a Qwen3-Max judge validated at κ_w = 0.798 against a domain expert, from a separate safety-violation (SV) check against source texts. Across 17 models in Chinese and an 8-model intersection over four languages, we find: (i) the best system reaches only 2.083 on the 0--3 rubric, leaving substantial headroom; (ii) Standards & Terminology is the most persistent capability weakness and survives item-aligned translation; (iii) extended reasoning lowers safety-adjusted scores for 12 of 13 models, primarily by introducing unsupported safety-critical details into longer final answers; and (iv) safety-violation rates reshuffle the leaderboard -- GPT-5.4 climbs from rank 6 to rank 3 after SV adjustment, while Kimi-k2.5-1T-A32B drops seven positions.Industrial LLM evaluation therefore requires source-grounded, safety-aware diagnosis rather than aggregate accuracy. We release IndustryBench with all prompts, scoring scripts, and dataset documentation.

We introduce OenoBench, a wine-domain knowledge benchmark of 3,266 multiple-choice questions across six pillars (regions, grape varieties, viticulture, winemaking, producers, business) and four difficulty tiers. The corpus is built from 38,104 atomic, source-anchored facts extracted by 35 provenance-verified scrapers from government registries (INAO, TTB, OIV), peer-reviewed journals, and Wikipedia/Wikidata. Our methodological contribution is an LLM-driven pipeline in which language models reformat verified facts and audit the result, but never serve as the source of truth: every claim traces to a URL, every question is generated by one of five strategies across five generator families, and every question is scored by a nine-agent audit calibrated against a human gold sheet via Cohen's κ. Evaluating sixteen frontier configurations, we find: (i) overall accuracy spans 53%-84%, led by o3 at 83.6%; (ii) reasoning-mode lift concentrates in DeepSeek R1 (+6.8pp) and is absent in Claude Opus and Gemini Pro; (iii) Anthropic shows +9pp self preference on its own questions while Google shows -8pp inverse preference; (iv) frontier open-weight models share the cost-vs-accuracy Pareto frontier with proprietary reasoning models; and (v) every config gains around 33pp on closed-book solvable items, revealing a parametric-recall ceiling that only the contextual slice avoids. We release corpus, audit findings, human-review app, and construction code under CC-BY-SA-4.0.

Large Language Models (LLMs) have demonstrated remarkable abilities in scientific reasoning, yet their reasoning capabilities in materials science remain underexplored. To fill this gap, we introduce MatSciBench, a comprehensive college-level benchmark comprising 1,340 problems that span the essential subdisciplines of materials science. MatSciBench features a structured and fine-grained taxonomy that categorizes materials science questions into 6 primary fields and 31 sub-fields, and includes a three-tier difficulty classification based on the reasoning length required to solve each question. MatSciBench provides detailed reference solutions enabling precise error analysis and incorporates multimodal reasoning through visual contexts in numerous questions. Evaluations of leading models reveal that even the highest-performing model, Gemini-2.5-Pro, achieves under 80% accuracy on college-level materials science questions, highlighting the complexity of MatSciBench. Our systematic analysis of different reasoning strategie--basic chain-of-thought, tool augmentation, and self-correction--demonstrates that no single method consistently excels across all scenarios. We further analyze performance by difficulty level, examine trade-offs between efficiency and accuracy, highlight the challenges inherent in multimodal reasoning tasks, analyze failure modes across LLMs and reasoning methods, and evaluate the influence of retrieval-augmented generation. MatSciBench thus establishes a comprehensive and solid benchmark for assessing and driving improvements in the scientific reasoning capabilities of LLMs within the materials science domain.

Large Language Models (LLMs) have made notable progress in automated theorem proving, yet existing formal benchmarks remain limited in both mathematical coverage and difficulty. Most are concentrated in areas that are easier to formalize, such as algebra and elementary number theory, and provide limited coverage of subfields that require deeper reasoning, including mathematical analysis. To address this gap, we introduce MA-ProofBench, to the best of our knowledge, the first formal theorem-proving benchmark dedicated to Mathematical Analysis. The benchmark contains 200 formalized theorems covering 6 core topics and 27 subcategories, including measure and integration theory, complex analysis, and functional analysis. The problems are divided into two difficulty levels, an undergraduate level (Level I, 100 problems) and a Ph.D. qualifying level (Level II, 100 problems), to evaluate how well LLMs perform formal reasoning at different mathematical depths. Each problem is constructed through a human-led, LLM-assisted formalization pipeline followed by independent expert review, ensuring that the formal statements remain faithful to the original mathematics. We evaluate a range of recent general-purpose reasoning models and formal theorem provers on MA-ProofBench. However, most models perform poorly: even the best-performing model, GPT-5.5, achieves only 16% Pass@8 on Level I and 5% on Level II, while most models stay close to 0% on Level II. Further analysis identifies Mathlib hallucinations and incomplete proofs as the two dominant failure modes, while an evaluation on the natural-language version of the benchmark exposes a clear gap between informal and formal reasoning. MA-ProofBench is intended to serve as a reliable reference for tracking progress in formal mathematical reasoning in advanced domains.

Vision-Language Models (VLMs) have advanced rapidly in multimodal perception and language understanding, yet it remains unclear whether they can reliably ground language into spatially coherent, plausibly executable actions in 3D digital environments. We introduce SleepWalk, a benchmark for evaluating instruction-grounded trajectory prediction in single-scene 3D worlds generated from textual scene descriptions and filtered for navigability. Unlike prior navigation benchmarks centered on long-range exploration across rooms, SleepWalk targets localized, interaction-centric embodied reasoning: given rendered visual observations and a natural-language instruction, a model must predict a trajectory that respects scene geometry, avoids collisions, and terminates at an action-compatible location. The benchmark covers diverse indoor and outdoor environments and organizes tasks into three tiers of spatial and temporal difficulty, enabling fine-grained analysis of grounding under increasing compositional complexity. Using a standardized pointwise judge-based evaluation protocol, we evaluate three frontier VLMs on 2,472 curated 3D environments with nine instructions per scene. Results reveal systematic failures in grounded spatial reasoning, especially under occlusion, interaction constraints, and multi-step instructions: performance drops as the difficulty level of the tasks increase. In general, current VLMs can somewhat produce trajectories that are simultaneously spatially coherent, plausibly executable, and aligned with intended actions. By exposing failures in a controlled yet scalable setting, SleepWalk provides a critical benchmark for advancing grounded multimodal reasoning, embodied planning, vision-language navigation, and action-capable agents in 3D environments.

Solving mathematical problems requires advanced reasoning abilities and presents notable challenges for large language models. Previous works usually synthesize data from proprietary models to augment existing datasets, followed by instruction tuning to achieve top-tier results. However, our analysis of these datasets reveals severe biases towards easy queries, with frequent failures to generate any correct response for the most challenging queries. Hypothesizing that difficult queries are crucial to learn complex reasoning, we propose Difficulty-Aware Rejection Tuning (DART), a method that allocates difficult queries more trials during the synthesis phase, enabling more extensive training on difficult samples. Utilizing DART, we have created new datasets for mathematical problem-solving that focus more on difficult queries and are substantially smaller than previous ones. Remarkably, our synthesis process solely relies on a 7B-sized open-weight model, without reliance on the commonly used proprietary GPT-4. We fine-tune various base models on our datasets ranging from 7B to 70B in size, resulting in a series of strong models called DART-MATH. In comprehensive in-domain and out-of-domain evaluation on 6 mathematical benchmarks, DART-MATH outperforms vanilla rejection tuning significantly, being superior or comparable to previous arts, despite using much smaller datasets and no proprietary models. Furthermore, our results position our synthetic datasets as the most effective and cost-efficient publicly available resources for advancing mathematical problem-solving.

Motion blur restoration on consumer mobile devices is typically evaluated using aggregate metrics that obscure performance variation across blur difficulty, masking model behavior under real deployment conditions. This work introduces iPhoneBlur, a difficulty-stratified benchmark of 7,400 image pairs synthesized from high-framerate iPhone 17 Pro videos captured in diverse real-world scenarios. Samples are partitioned into Easy, Medium, and Hard categories through PSNR-guided adaptive temporal windowing, with stratification validated by monotonic 2.2x increase in optical flow magnitude across tiers. Each sample includes comprehensive metadata enabling investigation of ISP-aware and difficulty-adaptive restoration strategies. Spectral analysis confirms synthesized blur exhibits high-frequency suppression patterns consistent with authentic motion degradation. Evaluation of six architectures reveals consistent 7-9 dB performance degradation from Easy to Hard subsets, a substantial gap entirely hidden by aggregate reporting. The benchmark further exposes a domain gap between professional and consumer cameras which targeted fine-tuning substantially recovers. By coupling difficulty stratification with deployment-critical metadata, iPhoneBlur enables systematic assessment of model reliability and failure modes for resource-constrained edge systems.

We present CycliST, a novel benchmark dataset designed to evaluate Video Language Models (VLM) on their ability for textual reasoning over cyclical state transitions. CycliST captures fundamental aspects of real-world processes by generating synthetic, richly structured video sequences featuring periodic patterns in object motion and visual attributes. CycliST employs a tiered evaluation system that progressively increases difficulty through variations in the number of cyclic objects, scene clutter, and lighting conditions, challenging state-of-the-art models on their spatio-temporal cognition. We conduct extensive experiments with current state-of-the-art VLMs, both open-source and proprietary, and reveal their limitations in generalizing to cyclical dynamics such as linear and orbital motion, as well as time-dependent changes in visual attributes like color and scale. Our results demonstrate that present-day VLMs struggle to reliably detect and exploit cyclic patterns, lack a notion of temporal understanding, and are unable to extract quantitative insights from scenes, such as the number of objects in motion, highlighting a significant technical gap that needs to be addressed. More specifically, we find no single model consistently leads in performance: neither size nor architecture correlates strongly with outcomes, and no model succeeds equally well across all tasks. By providing a targeted challenge and a comprehensive evaluation framework, CycliST paves the way for visual reasoning models that surpass the state-of-the-art in understanding periodic patterns.

We introduce InterChart, a diagnostic benchmark that evaluates how well vision-language models (VLMs) reason across multiple related charts, a task central to real-world applications such as scientific reporting, financial analysis, and public policy dashboards. Unlike prior benchmarks focusing on isolated, visually uniform charts, InterChart challenges models with diverse question types ranging from entity inference and trend correlation to numerical estimation and abstract multi-step reasoning grounded in 2-3 thematically or structurally related charts. We organize the benchmark into three tiers of increasing difficulty: (1) factual reasoning over individual charts, (2) integrative analysis across synthetically aligned chart sets, and (3) semantic inference over visually complex, real-world chart pairs. Our evaluation of state-of-the-art open and closed-source VLMs reveals consistent and steep accuracy declines as chart complexity increases. We find that models perform better when we decompose multi-entity charts into simpler visual units, underscoring their struggles with cross-chart integration. By exposing these systematic limitations, InterChart provides a rigorous framework for advancing multimodal reasoning in complex, multi-visual environments.

Multi-modal large language models (MM-LLMs) have shown strong performance in medical image understanding and clinical reasoning. Recent medical agent systems extend them with tool use and multi-agent collaboration, enabling complex decision-making. However, these systems rely almost entirely on frontier models (e.g., GPT), whose API-based deployment incurs high cost, high latency, and privacy risks that conflict with on-premise clinical requirements. We present Meissa, a lightweight 4B-parameter medical MM-LLM that brings agentic capability offline. Instead of imitating static answers, Meissa learns both when to engage external interaction (strategy selection) and how to execute multi-step interaction (strategy execution) by distilling structured trajectories from frontier models. Specifically, we propose: (1) Unified trajectory modeling: trajectories (reasoning and action traces) are represented within a single state-action-observation formalism, allowing one model to generalize across heterogeneous medical environments. (2) Three-tier stratified supervision: the model's own errors trigger progressive escalation from direct reasoning to tool-augmented and multi-agent interaction, explicitly learning difficulty-aware strategy selection. (3) Prospective-retrospective supervision: pairing exploratory forward traces with hindsight-rationalized execution traces enables stable learning of effective interaction policies. Trained on 40K curated trajectories, Meissa matches or exceeds proprietary frontier agents in 10 of 16 evaluation settings across 13 medical benchmarks spanning radiology, pathology, and clinical reasoning. Using over 25x fewer parameters than typical frontier models like Gemini-3, Meissa operates fully offline with 22x lower end-to-end latency compared to API-based deployment. Data, models, and environments are released at https://github.com/Schuture/Meissa.

Long-term memory is fundamental for personalized agents capable of accumulating knowledge, reasoning over user experiences, and adapting across time. However, existing memory benchmarks primarily target declarative memory, specifically semantic and episodic types, where all information is explicitly presented in dialogues. In contrast, real-world actions are also governed by non-declarative memory, including habitual and procedural types, and need to be inferred from diverse digital traces. To bridge this gap, we introduce Lifebench, which features densely connected, long-horizon event simulation. It pushes AI agents beyond simple recall, requiring the integration of declarative and non-declarative memory reasoning across diverse and temporally extended contexts. Building such a benchmark presents two key challenges: ensuring data quality and scalability. We maintain data quality by employing real-world priors, including anonymized social surveys, map APIs, and holiday-integrated calendars, thus enforcing fidelity, diversity and behavioral rationality within the dataset. Towards scalability, we draw inspiration from cognitive science and structure events according to their partonomic hierarchy; enabling efficient parallel generation while maintaining global coherence. Performance results show that top-tier, state-of-the-art memory systems reach just 55.2\% accuracy, highlighting the inherent difficulty of long-horizon retrieval and multi-source integration within our proposed benchmark. The dataset and data synthesis code are available at https://github.com/1754955896/LifeBench.

Large language models (LLMs) have rapidly progressed into general-purpose agents capable of solving a broad spectrum of tasks. However, current models remain inefficient at reasoning: they apply fixed inference-time compute regardless of task complexity, often overthinking simple problems while underthinking hard ones. This survey presents a comprehensive review of efficient test-time compute (TTC) strategies, which aim to improve the computational efficiency of LLM reasoning. We introduce a two-tiered taxonomy that distinguishes between L1-controllability, methods that operate under fixed compute budgets, and L2-adaptiveness, methods that dynamically scale inference based on input difficulty or model confidence. We benchmark leading proprietary LLMs across diverse datasets, highlighting critical trade-offs between reasoning performance and token usage. Compared to prior surveys on efficient reasoning, our review emphasizes the practical control, adaptability, and scalability of TTC methods. Finally, we discuss emerging trends such as hybrid thinking models and identify key challenges for future work towards making LLMs more computationally efficient, robust, and responsive to user constraints.

Recent supervised fine-tuning (SFT) approaches have significantly improved language models' performance on mathematical reasoning tasks, even when models are trained at a small scale. However, the specific capabilities enhanced through such fine-tuning remain poorly understood. In this paper, we conduct a detailed analysis of model performance on the AIME24 dataset to understand how reasoning capabilities evolve. We discover a ladder-like structure in problem difficulty, categorize questions into four tiers (Easy, Medium, Hard, and Extremely Hard (Exh)), and identify the specific requirements for advancing between tiers. We find that progression from Easy to Medium tier requires adopting an R1 reasoning style with minimal SFT (500-1K instances), while Hard-level questions suffer from frequent model's errors at each step of the reasoning chain, with accuracy plateauing at around 65% despite logarithmic scaling. Exh-level questions present a fundamentally different challenge; they require unconventional problem-solving skills that current models uniformly struggle with. Additional findings reveal that carefully curated small-scale datasets offer limited advantage-scaling dataset size proves far more effective. Our analysis provides a clearer roadmap for advancing language model capabilities in mathematical reasoning.

Equipping large language models with explicit skills has emerged as a promising paradigm for enabling autonomous agents to solve complex tasks. Agent skills can be inherently divided into general skills for broad cognitive transfer and task-specific skills for dynamic execution. However, existing skill-based reinforcement learning (RL) methods typically force a rigid choice between full externalization, which incurs prohibitive context overhead, and full internalization, which risks overfitting and knowledge conflicts. To address this dilemma, we propose Skill0.5, a novel agentic RL framework that explicitly differentiates skill treatments by combining general skill internalization with task-specific skill utilization. Driven by a dynamic, difficulty-aware router, Skill0.5 streams tasks into distinct mastery tiers to apply tailored optimization strategies: it internalizes general skills via privileged distillation to build a cognitive foundation for hard tasks, while using diagnostic probing on easy tasks to penalize shortcuts and enforce specific skill utilization. Experiments on ALFWorld and WebShop demonstrate that Skill0.5 outperforms both memory-based and skill-based RL baselines, yielding performance improvements across both in-distribution and out-of-distribution scenarios.

Gaze-referential inference--the ability to infer what others are looking at--is a critical component of a theory of mind that underpins natural human-AI interaction. In a controlled study, we evaluated this skill across 111 Vision Language Models (VLMs) using photos taken with manipulated difficulty and variability, comparing performance with that of human participants (N = 65), and analyzed behaviors using mixed-effects models. We found that 94 of the 111 VLMs failed to do better than random guessing, while humans achieved near-ceiling accuracy. VLMs even respond with each choice almost equally frequently. Are they randomly guessing? Although most VLMs struggle, when we zoom in on five of the top-tier VLMs with above-chance performance, we find that their performance declined with increasing task difficulty but varied only slightly across different prompts and scene objects. These behavioral features cannot be explained by considering them as random guessers. Instead, they likely use a combination of heuristics and guessing such that their performance is subject to the task difficulty but robust to perceptual variations. This suggests that VLMs, lacking gaze inference capability, have yet to become technologies that can naturally interact with humans, but the potential remains.

Large Language Models (LLMs) are changing the coding paradigm, known as vibe coding, yet synthesizing algorithmically sophisticated and robust code still remains a critical challenge. Incentivizing the deep reasoning capabilities of LLMs is essential to overcoming this hurdle. Reinforcement Fine-Tuning (RFT) has emerged as a promising strategy to address this need. However, most existing approaches overlook the heterogeneous difficulty and granularity inherent in test cases, leading to an imbalanced distribution of reward signals and consequently biased gradient updates during training. To address this, we propose Test-driven and cApability-adaptive cuRriculum reinfOrcement fine-Tuning (TAROT). TAROT systematically constructs, for each problem, a four-tier test suite (basic, intermediate, complex, edge), providing a controlled difficulty landscape for curriculum design and evaluation. Crucially, TAROT decouples curriculum progression from raw reward scores, enabling capability-conditioned evaluation and principled selection from a portfolio of curriculum policies rather than incidental test-case difficulty composition. This design fosters stable optimization and more efficient competency acquisition. Extensive experimental results reveal that the optimal curriculum for RFT in code generation is closely tied to a model's inherent capability, with less capable models achieving greater gains with an easy-to-hard progression, whereas more competent models excel under a hard-first curriculum. TAROT provides a reproducible method that adaptively tailors curriculum design to a model's capability, thereby consistently improving the functional correctness and robustness of the generated code. All code and data are released to foster reproducibility and advance community research at https://github.com/deep-diver/TAROT.

With the rapid improvement in the general capabilities of LLMs, LLM personalization, i.e., how to build LLM systems that can generate personalized responses or services that are tailored to distinct user personas, has become an increasingly important research and engineering problem. However, unlike many new challenging benchmarks being released for evaluating the general/reasoning capabilities, the lack of high-quality benchmarks for evaluating LLM personalization greatly hinders progress in this field. To address this, we introduce PersonaFeedback, a new benchmark that directly evaluates LLMs' ability to provide personalized responses given pre-defined user personas and queries. Unlike existing benchmarks that require models to infer implicit user personas from historical interactions, PersonaFeedback decouples persona inference from personalization, focusing on evaluating the model's ability to generate responses tailored to explicit personas. PersonaFeedback consists of 8298 human-annotated test cases, which are categorized into easy, medium, and hard tiers based on the contextual complexity of the user personas and the difficulty in distinguishing subtle differences between two personalized responses. We conduct comprehensive evaluations across a wide range of models. The empirical results reveal that even state-of-the-art LLMs that can solve complex real-world reasoning tasks could fall short on the hard tier of PersonaFeedback where even human evaluators may find the distinctions challenging. Furthermore, we conduct an in-depth analysis of failure modes across various types of systems, demonstrating that the current retrieval-augmented framework should not be seen as a de facto solution for personalization tasks. All benchmark data, annotation protocols, and the evaluation pipeline will be publicly available to facilitate future research on LLM personalization.

While generalization over tasks from easy to hard is crucial to profile language models (LLMs), the datasets with fine-grained difficulty annotations for each problem across a broad range of complexity are still blank. Aiming to address this limitation, we present Easy2Hard-Bench, a consistently formatted collection of 6 benchmark datasets spanning various domains, such as mathematics and programming problems, chess puzzles, and reasoning questions. Each problem within these datasets is annotated with numerical difficulty scores. To systematically estimate problem difficulties, we collect abundant performance data on attempts to each problem by humans in the real world or LLMs on the prominent leaderboard. Leveraging the rich performance data, we apply well-established difficulty ranking systems, such as Item Response Theory (IRT) and Glicko-2 models, to uniformly assign numerical difficulty scores to problems. Moreover, datasets in Easy2Hard-Bench distinguish themselves from previous collections by a higher proportion of challenging problems. Through extensive experiments with six state-of-the-art LLMs, we provide a comprehensive analysis of their performance and generalization capabilities across varying levels of difficulty, with the aim of inspiring future research in LLM generalization. The datasets are available at https://huggingface.co/datasets/furonghuang-lab/Easy2Hard-Bench.

Agentic coding assistants are increasingly given extra capabilities, such as browser based testing tools and design oriented system prompts, on the assumption that more capability yields better software. This study tested that assumption directly. Ninety independent agent runs built the same application, a real time retrospective board, from one detailed specification, each scored on a fixed 14 criterion functional rubric (42 point maximum) and a visual quality review. The runs spanned several model generations, two agent harnesses, two reasoning effort levels, a testing tool, and two design oriented prompts. Capability tier dominated: frontier models clustered near the ceiling while a low cost local model fell to 24 to 37 points. A criterion level analysis revealed what run totals conceal. Container deployment was the dominant defect, failing first try in 44 percent of runs, with its failure rate shifting sharply across model generations while mean totals moved less than a point. The testing tool raised cost by 42 to 68 percent without improving functional score or reliability, even on interface visible criteria. Raising reasoning effort from High to xHigh lifted first try perfect runs from 28 percent to 89 percent and cut corrective prompts about five fold, for 9 to 29 percent more cost. A design oriented prompt raised visual quality, 4.5 versus 3.0 on a 5 point scale, without lifting function, and a one paragraph paraphrase of its directive reproduced the entire lift. The practical lesson is to match the fix to the failure: most first run failures came from weak reasoning, which a stronger model or more effort prevents, not from visible flaws a checking tool would catch.

We introduce MetroLLM-Bench, a 955-case benchmark for testing language models as the policy layer of a transit kiosk. It covers six real metro systems, ranging from 37 to 414 stations, and eleven categories that include routing, fare calculation, disruptions, accessibility, and adversarial input. In each case, the model must call structured tools and submit a machine-renderable terminal state containing an outcome, a per-ticket fare quote when applicable, and a kiosk action. Fourteen deterministic scoring components form Tier 1; eight semantic-quality components form Tier 2, six of which use a language-model judge. We report Tier 1 and the combined score of both tiers. A stratified 75/25 split reserves 717 cases for training-data generation and 238 for held-out evaluation. We evaluate twenty-six models from six vendors, of which twenty-three are ranked. On the held-out partition, a 4B Qwen 3.5 student trained through parameter-efficient fine-tuning (PEFT) exceeds both GPT-5.6 tiers on Tier 1 (91.3 against 90.6 and 90.0) and matches GPT-5.4 full at maximum reasoning effort (91.4), with a 2.6 GB Q4_K_M footprint. Larger 9B and 27B students provide no further Tier 1 improvement over the 4B student at this training scale. Across the four Qwen sizes, the PEFT gain over the corresponding base model decreases from +7.03 points at 2B (three training seeds) to -0.91 at 27B; every seed shows the same direction at every size. A deterministic rule-based baseline reaches 84.6 on Tier 1, with the remaining language-model advantage concentrated in policy adaptation, compound scenarios, accessibility, and temporal reasoning. Muse Glimmer 30B leads the composite ranking, and serving configuration alone moves the Qwen 3.5-to-3.8 comparison by 2.7 Tier 1 points. The benchmark, harness, reproduction guide, and fine-tuned students are released at https://github.com/continker/metrollm-bench.

MathNet-Retrieve asks a retriever to find, for a math problem, a document stating the same problem. An LLM under one fixed prompt writes each gold document and its near-miss distractors; LLM judges filter them. We call this procedure the "recipe", training on pairs built the same way "recipe-matching", and ask how much score it buys beyond the ability the benchmark claims to test. Two models from one base, matched in rows and settings, differ only in the training file: pairs written under the benchmark's published prompt by another vendor's LLM and judge, or computer-algebra-verified pairs with no LLM anywhere. The first leads by 45 R@1 points on the easy tier. By a non-LLM paraphrase control, half to two thirds of that gap comes from the pairs being LLM-written at all: LLM rewrites under two unrelated prompts, with the verified model's negatives, recover 30 and 22 of the 45 points; back-translations with the same negatives recover almost none. The remaining 15 to 25 points appear only under the benchmark's own prompt and vanish on real duplicates no generator wrote, the same problem in two languages. The hard tier rewards the recipe's pair structure, a deep rewrite against a minimal-edit near-miss: LLM rewrites alone score zero on it, attaching negatives unlocks it, and every negative that does so costs cross-language points; the sets scoring highest on it separate near-misses no LLM wrote worse than LLM rewrites with verified negatives. MELD also moves when a model trains on pairs built its way, without losing retention; on SABER-Math the registered attack fails, and the one gain, from its LLM-written summaries, is small but holds at a matched budget. Only on MathNet-Retrieve could we pin an inversion, benchmark score up and real retention down, to one edit of a training file. We release the generator-free duplicate evaluations, the near-miss test and three trained models.

We investigate how well large language models (LLMs) generalize across different task difficulties, a key question for effective data curation and evaluation. Existing research is mixed regarding whether training on easier or harder data leads to better results, and whether those gains come on easier or harder test data. We address this question by conducting a systematic evaluation of LLMs' generalization across models, datasets, and fine-grained groups of example difficulty. We rank examples in six datasets using the outputs of thousands of different LLMs and Item Response Theory (IRT), a well-established difficulty metric in educational testing. Unlike prior work, our difficulty ratings are therefore determined solely by the abilities of many different LLMs, excluding human opinions of difficulty. With a more objective, larger-scale, and finer-grained analysis, we show that cross-difficulty generalization is often limited; training on either easy or hard data cannot achieve consistent improvements across the full range of difficulties. These results show the importance of having a range of difficulties in both training and evaluation data for LLMs, and that taking shortcuts with respect to difficulty is risky.

Tool use enables large language models to solve complex tasks through sequences of API calls, yet existing reinforcement learning approaches fail to scale to multi-step composition settings. Outcome-based rewards provide only sparse feedback, while trajectory-supervised rewards depend on annotated reference solutions, penalizing valid alternatives and limiting scalability. We propose TIER: Trajectory-Invariant Execution Rewards, a reward framework that derives supervision directly from function schemas and runtime execution, rather than from reference trajectories. The reward decomposes into format validity, schema adherence, execution success, and answer correctness, providing dense, interpretable sequence-level feedback derived from fine-grained verification of individual steps of tool use. This design allows any valid execution path to receive credit, naturally supporting multiple solution strategies and adapting to evolving tool interfaces. On DepthBench, a compositional benchmark stratified by depth (1 to 6 steps), TIER achieves >90% accuracy across steps, where trajectory-supervised rewards collapse beyond step-4. We further demonstrate consistent gains on benchmarks like BFCL v3 and NestFUL. Ablation studies confirm that all reward components are necessary, highlighting the importance of multi-level supervision for compositional reasoning.

Estimating the difficulty of a dataset typically involves comparing state-of-the-art models to humans; the bigger the performance gap, the harder the dataset is said to be. However, this comparison provides little understanding of how difficult each instance in a given distribution is, or what attributes make the dataset difficult for a given model. To address these questions, we frame dataset difficulty -- w.r.t. a model V -- as the lack of V-usable information (Xu et al., 2019), where a lower value indicates a more difficult dataset for V. We further introduce pointwise \mathcal{V-information} (PVI) for measuring the difficulty of individual instances w.r.t. a given distribution. While standard evaluation metrics typically only compare different models for the same dataset, V-usable information and PVI also permit the converse: for a given model V, we can compare different datasets, as well as different instances/slices of the same dataset. Furthermore, our framework allows for the interpretability of different input attributes via transformations of the input, which we use to discover annotation artefacts in widely-used NLP benchmarks.

As browser-use agents improve, benchmarks keep pace by collecting new tasks, websites, and applications, often making tasks longer or more novel. This makes difficulty expensive to refresh and difficult to control: when many aspects change at once, it is unclear what actually makes a task challenging. We instead construct challenging instances from tasks agents already solve, turning difficulty into a programmable property of the environment. BreakingWeb pairs every base task with an intervention condition that preserves the user instruction, latent target, and backend success criterion while changing the environment at different web stack layers. Each intervention is deterministic, detectable, and recoverable, and is annotated with the cognitive primitive it primarily loads. The benchmark contains 519 clean/intervention task pairs across seven self-hosted websites and 29 intervention families, all graded against outcomes. We evaluate six strong browser-use agents, three GUI-only agents that see only screenshots, and humans. The construction is effective: interventions cut agent pass rate by 22.9% on average and overturn nearly half of the tasks each agent solves cleanly, whereas humans lose 10.0% on a first attempt and 5.7% after one familiarisation attempt. The dominant failure is belief failure: 75% of the six agents' failures end with a declared success although the required change never happened. Our code, data and environment are publicly available at www.breakingweb.app.

Comprehensive evaluations of language models (LM) during both development and deployment phases are necessary because these models possess numerous capabilities (e.g., mathematical reasoning, legal support, or medical diagnostic) as well as safety risks (e.g., racial bias, toxicity, or misinformation). The average score across a wide range of benchmarks provides a signal that helps guide the use of these LMs in practice. Currently, holistic evaluations are costly due to the large volume of benchmark questions, making frequent evaluations impractical. A popular attempt to lower the cost is to compute the average score on a subset of the benchmark. This approach, unfortunately, often renders an unreliable measure of LM performance because the average score is often confounded with the difficulty of the questions in the benchmark subset. Item response theory (IRT) was designed to address this challenge, providing a reliable measurement by careful controlling for question difficulty. Unfortunately, question difficulty is expensive to estimate. Facing this challenge, we train a model that predicts question difficulty from its content, enabling a reliable measurement at a fraction of the cost. In addition, we leverage this difficulty predictor to further improve the evaluation efficiency through training a question generator given a difficulty level. This question generator is essential in adaptive testing, where, instead of using a random subset of the benchmark questions, informative questions are adaptively chosen based on the current estimation of LLM performance. Experiments on 22 common natural language benchmarks and 172 LMs show that this approach is more reliable and efficient compared to current common practice.

A benchmark is not merely a collection of test questions; it naturally induces an empirical measure over difficulty. In this work, we develop this measure-based view of benchmark evaluation. This perspective allows us to identify and formalize two central failure modes: measure mismatch, where the benchmark difficulty distribution differs from the intended difficulty distribution, and coverage gaps, where uncovered regions along the difficulty axis prevent benchmarks from detecting meaningful localized training progress. We propose two diagnostic metrics that capture these two failure modes and provide rigorous theoretical guarantees. Empirically, we analyze 25 popular benchmarks spanning 7 domains using carefully estimated item difficulties. Our results show that many benchmarks implicitly overweight certain difficulty regions, so their leaderboard conclusions should be interpreted with caution; others leave substantial portions of the difficulty axis uncovered, limiting their ability to monitor the model improvement process. Finally, we turn diagnosis into action by introducing BenchFill, a simple difficulty-targeted item rewriting algorithm that effectively reduces local coverage gaps.

A working citation looks like proof -- but the fact that a link resolves does not mean the cited paper supports the claim. I find that current agentic models rarely fabricate citations (over 99% resolve), yet roughly 15.9% link to the wrong paper. Existing benchmarks miss this failure mode: when a question has a fixed answer key, a model can reproduce the expected source from that key rather than independently verifying that the source supports the claim. I introduce \openbiorq{}, a retrieval-grounded agentic benchmark of 12{,}553 unsolved biomedical research questions across 12 domains that treats open questions as a faithfulness-and-abstention probe. To my knowledge, this is the first biomedical benchmark to combine an agentic setting -- where the model must issue multiple tool calls -- with unsolved questions that have no answer key. Openness is verified against real follow-up evidence rather than a model's parametric knowledge. Difficulty is empirical: I anchor it on questions that three open-weight reference models fail to answer, rather than on subjective hardness labels. On this hardest subset, held-out models from the same lineage as the difficulty anchors solve only ~17%, while three independent frontier agents (Gemini-3-Pro, Opus-4.7, GPT-5.5) span a wide 29-60% range. The benchmark is thus hard, non-saturating (the best agent still leaves ~33-40\% unsolved), and discriminating across capability tiers. Beyond difficulty, I observe agentic collapse on the hardest questions, where agents stop using their tools. For the most collapse-prone model, blocking tool access entirely barely changes its score -- so tools stop paying off exactly where they are needed most. A frozen per-question checklist raises inter-judge agreement from Spearman 0.35 to 0.82.

Clinical personality assessment screens response validity before interpreting substantive scales. LLM evaluation does not. We apply the validity scaling framework from the PAI and MMPI-3 to metacognitive probe data from 20 frontier models across 524 items. Six validity indices are operationalised: L (maintaining confidence on errors), K (betting on errors), F (withdrawing consensus-endorsed items), Fp (withdrawing correct answers), RBS (inverted monitoring), and TRIN (fixed responding). A tiered classification system identifies four models as construct-level invalid and two as elevated. Valid-profile models produce item-sensitive confidence (mean r = .18, 14 of 16 significant). Invalid-profile models do not (mean r = -.20, d = 2.17, p = .001). Chain-of-thought training produces two opposite response distortions. Two latent dimensions account for 94.6% of index variance. Companion papers extract a portable screening protocol (Cacioli, 2026e) and validate it against selective prediction (Cacioli, 2026f). All data and code: https://github.com/synthiumjp/validity-scaling-llm

In this paper, we investigate which questions are challenging for retrieval-based Question Answering (QA). We (i) propose retrieval complexity (RC), a novel metric conditioned on the completeness of retrieved documents, which measures the difficulty of answering questions, and (ii) propose an unsupervised pipeline to measure RC given an arbitrary retrieval system. Our proposed pipeline measures RC more accurately than alternative estimators, including LLMs, on six challenging QA benchmarks. Further investigation reveals that RC scores strongly correlate with both QA performance and expert judgment across five of the six studied benchmarks, indicating that RC is an effective measure of question difficulty. Subsequent categorization of high-RC questions shows that they span a broad set of question shapes, including multi-hop, compositional, and temporal QA, indicating that RC scores can categorize a new subset of complex questions. Our system can also have a major impact on retrieval-based systems by helping to identify more challenging questions on existing datasets.

The validity screen (Cacioli, 2026d, 2026e) classifies LLM confidence signals as Valid, Indeterminate, or Invalid. We test whether these classifications predict selective prediction performance. Twenty frontier LLMs from seven families were evaluated on 524 items across six cognitive tracks. Valid models show mean Type 2 AUROC = .624 (SD = .048). Invalid models show mean AUROC = .357 (SD = .231). Cohen's d = 2.81, p = .002. The tiers order monotonically: Invalid (.357) < Indeterminate (.554) < Valid (.624). Split-half cross-validation yields median d = 1.77, P(d > 0) = 1.0 across 1,000 splits. The three-tier classification accounts for 47% of the variance in AUROC. DeepSeek-R1 drops from 85.3% accuracy at full coverage to 11.3% at 10% coverage. The screen predicts the criterion. For selective prediction, the screen matters.

Existing browser agent benchmarks face a fundamental trilemma: real-website benchmarks lack reproducibility due to content drift, controlled environments sacrifice realism by omitting real-web noise, and both require costly manual curation that limits scalability. We present WebForge, the first fully automated framework that resolves this trilemma through a four-agent pipeline -- Plan, Generate, Refine, and Validate -- that produces interactive, self-contained web environments end-to-end without human annotation. A seven-dimensional difficulty control framework structures task design along navigation depth, visual complexity, reasoning difficulty, and more, enabling systematic capability profiling beyond single aggregate scores. Using WebForge, we construct WebForge-Bench, a benchmark of 934 tasks spanning 7 domains and 3 difficulty levels. Multi-model experiments show that difficulty stratification effectively differentiates model capabilities, while cross-domain analysis exposes capability biases invisible to aggregate metrics. Together, these results confirm that multi-dimensional evaluation reveals distinct capability profiles that a single aggregate score cannot capture. Code and benchmark are publicly available at https://github.com/yuandaxia2001/WebForge.

Long-horizon agents often compress interaction histories into write-time summaries. This creates a fundamental write-before-query barrier: compression decisions are made before the system knows what a future query will hinge on. As a result, summaries can cause unverifiable omissions -- decisive constraints (e.g., allergies) may be dropped, leaving the agent unable to justify an answer with traceable evidence. Retaining raw logs restores an authoritative source of truth, but grounding on raw logs by default is expensive: many queries are answerable from summaries, yet raw grounding still requires processing far longer contexts, inflating token consumption and latency. We propose TierMem, a provenance-linked framework that casts retrieval as an inference-time evidence allocation problem. TierMem uses a two-tier memory hierarchy to answer with the cheapest sufficient evidence: it queries a fast summary index by default, and a runtime sufficiency router Escalates to an immutable raw-log store only when summary evidence is insufficient. TierMem then writes back verified findings as new summary units linked to their raw sources. On LoCoMo, TierMem achieves 0.851 accuracy (vs.0.873 raw-only) while reducing input tokens by 54.1\% and latency by 60.7%.

The rapid proliferation of Large Language Models (LLMs) and diverse specialized benchmarks necessitates a shift from fragmented, task-specific metrics to a holistic, competitive ranking system that effectively aggregates performance across multiple ability dimensions. Primarily using static scoring, current evaluation methods are fundamentally limited. They struggle to determine the proper mix ratio across diverse benchmarks, and critically, they fail to capture a model's dynamic competitive fitness or its vulnerability when confronted with sequential, high-stakes tasks. To address this, we introduce the novel Competitive Swiss-System Dynamics (CSD) framework. CSD simulates a multi-round, sequential contest where models are dynamically paired across a curated sequence of benchmarks based on their accumulated win-loss record. And Monte Carlo Simulation (N=100,000 iterations) is used to approximate the statistically robust Expected Win Score (E[S_m]), which eliminates the noise of random pairing and early-round luck. Furthermore, we implement a Failure Sensitivity Analysis by parameterizing the per-round elimination quantity (T_k), which allows us to profile models based on their risk appetite--distinguishing between robust generalists and aggressive specialists. We demonstrate that CSD provides a more nuanced and context-aware ranking than traditional aggregate scoring and static pairwise models, representing a vital step towards risk-informed, next-generation LLM evaluation.

WANDR (Wide ANd Deep Research) is a benchmark of 500 realistic, challenging data-collection tasks for research agents. Each task requires a system to discover a large set of entities that satisfy specified criteria (breadth), investigate each entity through multiple coordinated web searches (depth), and return independently verifiable records with supporting sources and excerpts. Tasks are represented as qualification key hierarchies that specify the entities, relationships, evidence, and required count at each level; a hierarchy with n companies, m employees per company, and k sources per employee requires n x m x k records. This structure supports diverse workflows such as market mapping, due diligence, literature review, product comparison, and talent sourcing, with targets ranging from dozens to thousands of records. WANDR replaces static gold answer sets with task-specific judges that refetch cited pages and verify each record against its evidence, allowing evaluation of current and changing facts. Record verdicts are aggregated into soft and hard precision, recall, and F1 scores that distinguish factual quality, coverage, and hierarchical completeness. The tasks are derived from de-identified product-usage logs and produced through a semi-automated pipeline with automated checks, empirical audits, and human review where needed. We evaluate six production research systems and find that the benchmark is far from saturated: at high effort, the strongest system reaches only 0.363 soft F1 and 0.133 hard F1. Performance degrades as target volume and hierarchy depth increase, with incomplete discovery, missing enrichment, and incomplete evidence construction remaining major bottlenecks. The benchmark and evaluation harness are available at https://github.com/perplexityai/wandr.

Enhancing the reasoning capabilities of Large Language Models (LLMs) with efficiency and scalability remains a fundamental challenge in artificial intelligence research. This paper presents a rigorous experimental investigation into how difficulty-aware staged reinforcement learning (RL) strategies can substantially improve LLM reasoning performance. Through systematic analysis, we demonstrate that strategically selecting training data according to well-defined difficulty levels markedly enhances RL optimization. Moreover, we introduce a staged training methodology, progressively exposing models to increasingly challenging tasks, further amplifying reasoning capabilities. Our findings reveal significant cross-domain benefits when simultaneously training models on mathematical reasoning and code generation tasks. Notably, our proposed approach enables a 1.5B parameter model to achieve an accuracy of 42.3\% on the AIME-2024 benchmark, 89.5\% on the MATH-500 benchmark. These results underscore the efficacy of our method in advancing the reasoning proficiency of LLMs. We will open-source our datasets on GitHub and Hugging Face.

Accurate estimation of item (question or task) difficulty is critical for educational assessment but suffers from the cold start problem. While Large Language Models demonstrate superhuman problem-solving capabilities, it remains an open question whether they can perceive the cognitive struggles of human learners. In this work, we present a large-scale empirical analysis of Human-AI Difficulty Alignment for over 20 models across diverse domains such as medical knowledge and mathematical reasoning. Our findings reveal a systematic misalignment where scaling up model size is not reliably helpful; instead of aligning with humans, models converge toward a shared machine consensus. We observe that high performance often impedes accurate difficulty estimation, as models struggle to simulate the capability limitations of students even when being explicitly prompted to adopt specific proficiency levels. Furthermore, we identify a critical lack of introspection, as models fail to predict their own limitations. These results suggest that general problem-solving capability does not imply an understanding of human cognitive struggles, highlighting the challenge of using current models for automated difficulty prediction.

A prevalent assumption in LLM agent deployment holds that more structured harnesses universally improve reliability, and that higher-capability models need proportionally less structural guidance -- together implying a monotone inverse relationship between model capability tier and optimal harness complexity. We test this hypothesis through a controlled 432-run experiment crossing six models across four capability tiers with three harness conditions (light, balanced, strict) on HEAT-24, a 24-task synthetic benchmark with git-based workspace verification. Our results refute the monotone inverse relationship on two fronts. First, for the frontier chat model evaluated (Gemini 2.5 Flash), increased harness verbosity lowers VTSR by 29-38 percentage points -- a harness-complexity paradox. Second, for the frontier reasoning model evaluated (Qwen3.5-122B, extended thinking enabled), strict harness achieves the highest VTSR (91.7%) and the lowest latency, the opposite of the prediction. Within the constrained tier, a 2B model (Gemma4:e2B) matches strong-open-tier stability at 91.7% across all harnesses. Because each tier is represented by a single model in this study, these results should be interpreted as model-specific observations; harness sensitivity appears non-monotone across the models evaluated, and depends critically on model type (chat vs. reasoning). We introduce a six-label failure taxonomy showing that format_violation dominates capable-model failures while wrong_file dominates low-capability failures, and we derive practical tier-aware harness selection guidelines.

We present GPQA, a challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry. We ensure that the questions are high-quality and extremely difficult: experts who have or are pursuing PhDs in the corresponding domains reach 65% accuracy (74% when discounting clear mistakes the experts identified in retrospect), while highly skilled non-expert validators only reach 34% accuracy, despite spending on average over 30 minutes with unrestricted access to the web (i.e., the questions are "Google-proof"). The questions are also difficult for state-of-the-art AI systems, with our strongest GPT-4 based baseline achieving 39% accuracy. If we are to use future AI systems to help us answer very hard questions, for example, when developing new scientific knowledge, we need to develop scalable oversight methods that enable humans to supervise their outputs, which may be difficult even if the supervisors are themselves skilled and knowledgeable. The difficulty of GPQA both for skilled non-experts and frontier AI systems should enable realistic scalable oversight experiments, which we hope can help devise ways for human experts to reliably get truthful information from AI systems that surpass human capabilities.

Vision-Language Models (VLMs) have advanced across multimodal benchmarks but still show clear gaps in ordinal number understanding, i.e., the ability to track relative positions and generalize to large indices. We present OrdinalBench, a diagnostic benchmark that standardizes ordinal number understanding as an evaluation task for VLMs. The core task is N-th object identification, defined by a starting reference and traversal rule. Task difficulty is controlled along three axes: (i) ordinal magnitude, from small numbers to extreme cases up to 300; (ii) arrangement complexity, from single loops to maze-like paths; and (iii) object count. The benchmark provides 39,000 question-answer pairs, each annotated with a ground-truth reasoning trajectory and balanced across difficulty levels for controlled large-scale testing. Beyond answer-only evaluation, our framework requires models to generate structured stepwise traces of the counting process and provides an open evaluation toolkit that measures both final accuracy and step-level path consistency. Zero-shot evaluations of GPT-5, Gemini 2.5 Flash Lite, Qwen2.5-VL, InternVL3.5, and Molmo reveal sharp degradation under large-ordinal and complex-path conditions, highlighting weak generalization despite strong scores on standard multimodal tasks. By framing ordinal number understanding as a core target, OrdinalBench provides a reproducible benchmark and diagnostic framework for developing VLMs with stronger sequential reasoning. All data and code are available at https://ordinalbench.github.io/

Agentic large language model (LLM) systems rely on external memory for long-horizon state and concurrent multi-agent execution, but centralized indexes and heuristic partitions become bottlenecks as memory volume and parallel access grow. We present ShardMemo, a budgeted tiered memory service with Tier A per-agent working state, Tier B sharded evidence with shard-local approximate nearest neighbor (ANN) indexes, and Tier C, a versioned skill library. Tier B enforces scope-before-routing: structured eligibility constraints mask ineligible shards before routing or ANN search. We cast shard probing as masked mixture-of-experts (MoE) routing over eligible shards, probing up to B_{probe} shards via Top-B_{probe} or adaptive Top-P, and use cost-aware gating over profile/observation/session shard families; the router is trained from evidence-to-shard supervision. On LoCoMo, ShardMemo improves over the strongest baseline (GAM) by +5.11 to +6.82 F1 across question categories. Under a fixed-budget routing setting (B_{probe}=3), ShardMemo improves over cosine-to-prototype shard routing by +6.87 F1 while reducing retrieval work (VecScan 521->414, -20.5%) and p95 latency (95->76 ms). On long-context HotpotQA, ShardMemo achieves 63.41/61.88/57.95 F1 at 56K/224K/448K tokens. On ToolBench, Tier C reaches 0.97 Precision@3 and 1.94 StepRed (+10.2% and +7.2% over embedding-similarity retrieval).

How to design reinforcement learning (RL) tasks that effectively unleash the reasoning capability of large language models (LLMs) remains an open question. Existing RL tasks (e.g., math, programming, and constructing reasoning tasks) suffer from three key limitations: (1) Scalability. They rely heavily on human annotation or expensive LLM synthesis to generate sufficient training data. (2) Verifiability. LLMs' outputs are hard to verify automatically and reliably. (3) Controllable Difficulty. Most tasks lack fine-grained difficulty control, making it hard to train LLMs to develop reasoning ability from easy to hard. To address these limitations, we propose Saturn, a SAT-based RL framework that uses Boolean Satisfiability (SAT) problems to train and evaluate LLM reasoning. Saturn enables scalable task construction, rule-based verification, and precise difficulty control. Saturn designs a curriculum learning pipeline that continuously improves LLMs' reasoning capability by constructing SAT tasks of increasing difficulty and training LLMs from easy to hard. To ensure stable training, we design a principled mechanism to control difficulty transitions. We introduce Saturn-2.6k, a dataset of 2,660 SAT problems with varying difficulty. It supports the evaluation of how LLM reasoning changes with problem difficulty. We apply Saturn to DeepSeek-R1-Distill-Qwen and obtain Saturn-1.5B and Saturn-7B. We achieve several notable results: (1) On SAT problems, Saturn-1.5B and Saturn-7B achieve average pass@3 improvements of +14.0 and +28.1, respectively. (2) On math and programming tasks, Saturn-1.5B and Saturn-7B improve average scores by +4.9 and +1.8 on benchmarks (e.g., AIME, LiveCodeBench). (3) Compared to the state-of-the-art (SOTA) approach in constructing RL tasks, Saturn achieves further improvements of +8.8%. We release the source code, data, and models to support future research.

As the field progresses toward Artificial General Intelligence (AGI), there is a pressing need for more comprehensive and insightful evaluation frameworks that go beyond aggregate performance metrics. This paper introduces a unified rating system that jointly models the difficulty of individual test cases and the competency of AI models (or humans) across vision, language, and action domains. Unlike existing metrics that focus solely on models, our approach allows for fine-grained, difficulty-aware evaluations through competitive interactions between models and tasks, capturing both the long-tail distribution of real-world challenges and the competency gap between current models and full task mastery. We validate the generalizability and robustness of our system through extensive experiments on multiple established datasets and models across distinct AGI domains. The resulting rating distributions offer novel perspectives and interpretable insights into task difficulty, model progression, and the outstanding challenges that remain on the path to achieving full AGI task mastery.

Reasoning models have demonstrated impressive performance on difficult tasks that traditional language models struggle at. However, many are plagued with the problem of overthinking--generating large amounts of unnecessary tokens which don't improve accuracy on a question. We introduce approximate measures of problem-level difficulty and demonstrate that a clear relationship between problem difficulty and optimal token spend exists, and evaluate how well calibrated a variety of reasoning models are in terms of efficiently allocating the optimal token count. We find that in general, reasoning models are poorly calibrated, particularly on easy problems. To evaluate calibration on easy questions we introduce DUMB500, a dataset of extremely easy math, reasoning, code, and task problems, and jointly evaluate reasoning model on these simple examples and extremely difficult examples from existing frontier benchmarks on the same task domain. Finally, we introduce THOUGHTTERMINATOR, a training-free black box decoding technique that significantly improves reasoning model calibration.

Language-model agents are increasingly deployed under standing instructions: a system prompt, a policy file, or a skills document is placed in context, and the agent is trusted to let that document govern every action that follows. Existing benchmarks rarely test this deployment pattern directly; they measure whether an agent can complete a task, not whether a long, binding policy document constrains its behavior over an extended tool-use horizon. We present HANDBOOK_md, a benchmark of 65 agentic tasks modeled on how employees follow company handbooks. Each task places an agent in a self-contained company environment (a file workspace with mock email, chat, calendar, issue-tracking, and commerce services exposed over the Model Context Protocol) and instructs it to carry out routine professional work governed by an expert-written standard operating procedure of 20-124 pages. Tasks span five domains (finance, medical billing, insurance, logistics, and HR) and 10 fictional companies. To resist memorization, every task modifies one of 10 base handbooks, altering the specific rules and thresholds on which grading depends, so no two tasks share the same set of policies. Grading is fully deterministic: each task carries a rubric of programmatic criteria (824 in total) that check both that required actions occurred and that prohibited actions did not. Under strict grading, where a trial passes only if every criterion is satisfied, the strongest evaluated model passes 36.2% of trials, and most frontier models remain below 25%. Failures follow consistent patterns: agents let a plausible but unauthorized in-environment request override the standing policy, perform a required check and then act against its result, lose rule details over long horizons, and report compliance they did not achieve. We release the tasks, environments, and evaluation harness.

Balancing game difficulty in video games is a key task to create interesting gaming experiences for players. Mismatching the game difficulty and a player's skill or commitment results in frustration or boredom on the player's side, and hence reduces time spent playing the game. In this work, we explore balancing game difficulty using machine learning-based agents to challenge players based on their current behavior. This is achieved by a combination of two agents, in which one learns to imitate the player, while the second is trained to beat the first. In our demo, we investigate the proposed framework for personalized dynamic difficulty adjustment of AI agents in the context of the fighting game AI competition.

We present the system description for our submission towards the Key Point Analysis Shared Task at ArgMining 2021. Track 1 of the shared task requires participants to develop methods to predict the match score between each pair of arguments and keypoints, provided they belong to the same topic under the same stance. We leveraged existing state of the art pre-trained language models along with incorporating additional data and features extracted from the inputs (topics, key points, and arguments) to improve performance. We were able to achieve mAP strict and mAP relaxed score of 0.872 and 0.966 respectively in the evaluation phase, securing 5th place on the leaderboard. In the post evaluation phase, we achieved a mAP strict and mAP relaxed score of 0.921 and 0.982 respectively. All the codes to generate reproducible results on our models are available on Github.

This document presents a detailed description of the challenge on clarifying questions for dialogue systems (ClariQ). The challenge is organized as part of the Conversational AI challenge series (ConvAI3) at Search Oriented Conversational AI (SCAI) EMNLP workshop in 2020. The main aim of the conversational systems is to return an appropriate answer in response to the user requests. However, some user requests might be ambiguous. In IR settings such a situation is handled mainly thought the diversification of the search result page. It is however much more challenging in dialogue settings with limited bandwidth. Therefore, in this challenge, we provide a common evaluation framework to evaluate mixed-initiative conversations. Participants are asked to rank clarifying questions in an information-seeking conversations. The challenge is organized in two stages where in Stage 1 we evaluate the submissions in an offline setting and single-turn conversations. Top participants of Stage 1 get the chance to have their model tested by human annotators.

Knowledge benchmarks for LLMs face three issues: scaling-driven designs that do not operationalize disciplinary representativeness; flat-payment annotation that permits lazy consensus; and unaudited ranking instability under bounded test budgets. We introduce KINA, an 899-item benchmark across 261 fine-grained disciplines, with two formal results. First, we cast representativeness as a coverage-style objective over expert-elicited anchors and operationalize disciplinary representativeness through a proxy, yielding a (1-1/e) greedy approximation (Proposition 1); the guarantee applies to the proxy, not to population representativeness. Second, we prove a bonus-on-bar tournament weakly FOSD-dominates flat payment in released-review quality, with incentive-compatibility threshold B > Delta C / Delta p_min (Theorem 1). Evaluating 42 models from 13 labs, the top model, Gemini-3.1-Pro-Preview, reaches 53.17%, followed by Claude-Opus-4.6 at 49.92% and GPT-5.4 at 48.55%, leaving substantial headroom below saturation. The full leaderboard shows a tiered structure rather than a smooth total order: a small frontier tier lies above 48%, a dense strong-model tier spans roughly 38-45%, and low-performing models remain only modestly above the 10% chance baseline. Tool augmentation adds up to 5.17 points across the five tool-use evaluations, with gains varying substantially across models. We report bootstrap ranking-stability statistics to make bounded-budget variance explicit and to discourage over-interpretation of adjacent ranks.

Static coding benchmarks are fixed at release. Over time their tasks leak into training data, and stronger agents eventually saturate them. We argue that a benchmark should instead aim to be self-scaling: its tasks and oracles are created at evaluation time, so no content exists in advance to leak, and its difficulty is set by the participating agents rather than fixed at release. We introduce SWE-Duel, an arena in which two coding agents evaluate each other over mirrored Red-Blue turns. In each turn, the Red agent works on a real repository across three context-isolated sessions. It proposes and implements a new feature with feature tests, then embeds a subtle bug in that feature along with sealed bug-triggering tests. The resulting pull request is admitted only if (1) it passes the feature tests and the repository's existing tests while failing exactly the bug tests, and (2) a fresh session of the same agent can repair the bug blind, which rules out unsolvable tasks. The Blue agent then reviews the pull request without being told a bug exists. It wins only if its revision passes all three test suites, and then the two agents swap roles. We run a tournament among 11 model-harness participants on 15 repositories in five languages for 1,500 turns. The leaders win in different ways. GPT-5.5 with Codex CLI ranks first as the strongest solver, repairing 98.3% of the defects it reviews. Claude Opus 4.8 with mini-swe-agent ranks second as the strongest task author, with a 27.3% Red win rate. Finally, within one model family, tasks written by the newest version defeat a fixed-reference solver far more often than those of older versions (42.9% vs. 4.0%), giving preliminary evidence that task difficulty scales with task authoring capability.

ARC Challenge appears more difficult than ARC Easy for modern LLMs primarily due to an evaluation setup that prevents direct comparison of answer choices rather than inherent complexity. Although some researchers have quietly shifted to a more appropriate scheme over the last year, the implications of this change have yet to be widely acknowledged. We highlight this overlooked shift, show how similar evaluation practices falsely imply reasoning deficits in other benchmarks, and demonstrate that fairer methods dramatically reduce performance gaps (e.g. on SIQA) and even yield superhuman results (OpenBookQA). In doing so, we reveal how evaluation shapes perceived difficulty and offer guidelines to ensure that multiple-choice evaluations accurately reflect actual model capabilities.

Existing RAG benchmarks often overlook query difficulty, leading to inflated performance on simpler questions and unreliable evaluations. A robust benchmark dataset must satisfy three key criteria: quality, diversity, and difficulty, which capturing the complexity of reasoning based on hops and the distribution of supporting evidence. In this paper, we propose MHTS (Multi-Hop Tree Structure), a novel dataset synthesis framework that systematically controls multi-hop reasoning complexity by leveraging a multi-hop tree structure to generate logically connected, multi-chunk queries. Our fine-grained difficulty estimation formula exhibits a strong correlation with the overall performance metrics of a RAG system, validating its effectiveness in assessing both retrieval and answer generation capabilities. By ensuring high-quality, diverse, and difficulty-controlled queries, our approach enhances RAG evaluation and benchmarking capabilities.

How can we train models to perform well on hard test data when hard training data is by definition difficult to label correctly? This question has been termed the scalable oversight problem and has drawn increasing attention as language models have continually improved. In this paper, we present the surprising conclusion that current language models often generalize relatively well from easy to hard data, even performing as well as "oracle" models trained on hard data. We demonstrate this kind of easy-to-hard generalization using simple training methods like in-context learning, linear classifier heads, and QLoRA for seven different measures of datapoint hardness, including six empirically diverse human hardness measures (like grade level) and one model-based measure (loss-based). Furthermore, we show that even if one cares most about model performance on hard data, it can be better to collect and train on easy data rather than hard data, since hard data is generally noisier and costlier to collect. Our experiments use open models up to 70b in size and four publicly available question-answering datasets with questions ranging in difficulty from 3rd grade science questions to college level STEM questions and general-knowledge trivia. We conclude that easy-to-hard generalization in LMs is surprisingly strong for the tasks studied, suggesting the scalable oversight problem may be easier than previously thought. Our code is available at https://github.com/allenai/easy-to-hard-generalization

Following the recent achievement of gold-medal performance on the IMO by frontier LLMs, the community is searching for the next meaningful and challenging target for measuring LLM reasoning. Whereas olympiad-style problems measure step-by-step reasoning alone, research-level problems use such reasoning to advance the frontier of mathematical knowledge itself, emerging as a compelling alternative. Yet research-level math benchmarks remain scarce because such problems are difficult to source (e.g., Riemann Bench and FrontierMath-Tier 4 contain 25 and 50 problems, respectively). To support reliable evaluation of next-generation frontier models, we introduce Soohak, a 439-problem benchmark newly authored from scratch by 64 mathematicians. Soohak comprises two subsets. On the Challenge subset, frontier models including Gemini-3-Pro, GPT-5, and Claude-Opus-4.5 reach 30.4%, 26.4%, and 10.4% respectively, leaving substantial headroom, while leading open-weight models such as Qwen3-235B, GPT-OSS-120B, and Kimi-2.5 remain below 15%. Notably, beyond standard problem solving, Soohak introduces a refusal subset that probes a capability intrinsic to research mathematics: recognizing ill-posed problems and pausing rather than producing confident but unjustified answers. On this subset, no model exceeds 50%, identifying refusal as a new optimization target that current models do not directly address. To prevent contamination, the dataset will be publicly released in late 2026, with model evaluations available upon request in the interim.

Machine translation quality has steadily improved over the years, achieving near-perfect translations in recent benchmarks. These high-quality outputs make it difficult to distinguish between state-of-the-art models and to identify areas for future improvement. In this context, automatically identifying texts where machine translation systems struggle holds promise for developing more discriminative evaluations and guiding future research. In this work, we address this gap by formalizing the task of translation difficulty estimation, defining a text's difficulty based on the expected quality of its translations. We introduce a new metric to evaluate difficulty estimators and use it to assess both baselines and novel approaches. Finally, we demonstrate the practical utility of difficulty estimators by using them to construct more challenging benchmarks for machine translation. Our results show that dedicated models outperform both heuristic-based methods and LLM-as-a-judge approaches, with Sentinel-src achieving the best performance. Thus, we release two improved models for difficulty estimation, Sentinel-src-24 and Sentinel-src-25, which can be used to scan large collections of texts and select those most likely to challenge contemporary machine translation systems.

Recent diagnostic datasets on compositional generalization, such as SCAN (Lake and Baroni, 2018) and COGS (Kim and Linzen, 2020), expose severe problems in models trained from scratch on these datasets. However, in contrast to this poor performance, state-of-the-art models trained on larger and more general datasets show better generalization ability. In this work, to reconcile this inconsistency, we conduct an empirical analysis by training Transformer models on a variety of training sets with different data factors, including dataset scale, pattern complexity, example difficulty, etc. First, we show that increased dataset complexity can lead to better generalization behavior on multiple different generalization challenges. To further understand this improvement, we show two axes of the benefit from more complex datasets: they provide more diverse examples so compositional understanding becomes more effective, and they also prevent ungeneralizable memorization of the examples due to reduced example repetition frequency. Finally, we explore how training examples of different difficulty levels influence generalization differently. On synthetic datasets, simple examples invoke stronger compositionality than hard examples do. On larger-scale real language datasets, while hard examples become more important potentially to ensure decent data coverage, a balanced mixture of simple and hard examples manages to induce the strongest generalizability. The code and data for this work are available at https://github.com/owenzx/data4comp

Training next-generation code generation models requires high-quality datasets, yet existing datasets face difficulty imbalance, format inconsistency, and data quality problems. We address these challenges through systematic data processing and difficulty scaling. We introduce a four-stage Data Processing Framework encompassing collection, processing, filtering, and verification, incorporating Automatic Difficulty Filtering via an LLM-based predict-calibrate-select framework that leverages multi-dimensional difficulty metrics across five weighted dimensions to retain challenging problems while removing simplistic ones. The resulting MicroCoder dataset comprises tens of thousands of curated real competitive programming problems from diverse platforms, emphasizing recency and difficulty. Evaluations on strictly unseen LiveCodeBench demonstrate that MicroCoder achieves 3x larger performance gains within 300 training steps compared to widely-used baseline datasets of comparable size, with consistent advantages under both GRPO and its variant training algorithms. The MicroCoder dataset delivers obvious improvements on medium and hard problems across different model sizes, achieving up to 17.2% relative gains in overall performance where model capabilities are most stretched. These results validate that difficulty-aware data curation improves model performance on challenging tasks, providing multiple insights for dataset creation in code generation.

Collecting high-quality training examples for language model fine-tuning is expensive, with practical budgets limiting the amount of data that can be procured. We investigate whether example difficulty affects GRPO training effectiveness by comparing selection strategies (easy, medium, hard, random) across multiple models and reasoning tasks. Training on the hardest 10\% of examples (those where the base model fails most often) yields dramatic performance gains up to 47\%, while easy examples produce minimal improvements of 3-15\%. This occurs because GRPO requires outcome variance to generate learning signals; hard examples maintain mixed success/failure outcomes throughout training while easy examples quickly converge to consistent success, eliminating learning opportunities. Moreover, models trained on hard examples show superior out-of-distribution generalization, with only hard-trained models achieving meaningful gains on the AIME2025 benchmark. Our findings provide clear guidance: when budget-constrained, prioritize collecting and annotating examples where your base model struggles, as these drive nearly all learning value in GRPO fine-tuning

In precision-oriented tasks like answer ranking, it is more important to rank many relevant answers highly than to retrieve all relevant answers. It follows that a good ranking strategy would be to learn how to identify the easiest correct answers first (i.e., assign a high ranking score to answers that have characteristics that usually indicate relevance, and a low ranking score to those with characteristics that do not), before incorporating more complex logic to handle difficult cases (e.g., semantic matching or reasoning). In this work, we apply this idea to the training of neural answer rankers using curriculum learning. We propose several heuristics to estimate the difficulty of a given training sample. We show that the proposed heuristics can be used to build a training curriculum that down-weights difficult samples early in the training process. As the training process progresses, our approach gradually shifts to weighting all samples equally, regardless of difficulty. We present a comprehensive evaluation of our proposed idea on three answer ranking datasets. Results show that our approach leads to superior performance of two leading neural ranking architectures, namely BERT and ConvKNRM, using both pointwise and pairwise losses. When applied to a BERT-based ranker, our method yields up to a 4% improvement in MRR and a 9% improvement in P@1 (compared to the model trained without a curriculum). This results in models that can achieve comparable performance to more expensive state-of-the-art techniques.

Clinical protocol-execution tasks -- checking a lab value, applying a threshold, placing a correctly structured FHIR order -- are natural candidates for RL from world feedback: once clinical SMEs encode decision logic into a verifier, that verifier grades unlimited rollouts without per-episode annotation. But applying RL requires a sound feedback channel and sufficient base capability. We audit MedAgentBench v1/v2, find a 41.7\% silent-finish ceiling that makes inaction the RL dominant strategy, and construct MedAgentBench-v3 (MAB-v3) (508 tasks, 8.9\% ceiling). Training Qwen3-8B exposes two structural barriers: a capability ceiling (10/20 task types have 0\% base performance, zero gradient) and a format-knowledge barrier (3/20 types require exact clinical codes undiscoverable by exploration). Pure RL reaches 18.2\% pass@1 vs.\ 34.1\% for rule-based SFT; the 15.9~pp gap is attributable entirely to these barriers. A decision/format-knowledge/lookup taxonomy predicts RL learnability and prescribes the fix: SFT to inject codes, RL to learn conditionals.

Current AI alignment methodologies rely on human-provided demonstrations or judgments, and the learned capabilities of AI systems would be upper-bounded by human capabilities as a result. This raises a challenging research question: How can we keep improving the systems when their capabilities have surpassed the levels of humans? This paper answers this question in the context of tackling hard reasoning tasks (e.g., level 4-5 MATH problems) via learning from human annotations on easier tasks (e.g., level 1-3 MATH problems), which we term as easy-to-hard generalization. Our key insight is that an evaluator (reward model) trained on supervisions for easier tasks can be effectively used for scoring candidate solutions of harder tasks and hence facilitating easy-to-hard generalization over different levels of tasks. Based on this insight, we propose a novel approach to scalable alignment, which firstly trains the process-supervised reward models on easy problems (e.g., level 1-3), and then uses them to evaluate the performance of policy models on hard problems. We show that such easy-to-hard generalization from evaluators can enable easy-to-hard generalizations in generators either through re-ranking or reinforcement learning (RL). Notably, our process-supervised 7b RL model achieves an accuracy of 34.0\% on MATH500, despite only using human supervision on easy problems. Our approach suggests a promising path toward AI systems that advance beyond the frontier of human supervision.

Automatically estimating the complexity of texts for readers has a variety of applications, such as recommending texts with an appropriate complexity level to language learners or supporting the evaluation of text simplification approaches. In this paper, we present our submission to the Text Complexity DE Challenge 2022, a regression task where the goal is to predict the complexity of a German sentence for German learners at level B. Our approach relies on more than 220,000 pseudo-labels created from the German Wikipedia and other corpora to train Transformer-based models, and refrains from any feature engineering or any additional, labeled data. We find that the pseudo-label-based approach gives impressive results yet requires little to no adjustment to the specific task and therefore could be easily adapted to other domains and tasks.

Training large neural language models on large datasets is resource- and time-intensive. These requirements create a barrier to entry, where those with fewer resources cannot build competitive models. This paper presents various techniques for making it possible to (a) train a large language model using resources that a modest research lab might have, and (b) train it in a reasonable amount of time. We provide concrete recommendations for practitioners, which we illustrate with a case study: a T5 model for Danish, the first for this language.

Recent AI systems have achieved strong results on a wide range of benchmarks, yet these gains have not translated into economically meaningful deployment across many professional domains. We argue that this gap is largely an evaluation problem: widely used benchmarks lack sustained performance measurement on real and economically valuable workflows. This paper introduces Agents' Last Exam (ALE), a benchmark designed to evaluate AI agents on long-horizon, economically valuable, real-world tasks with verifiable outcomes. Developed in collaboration with 250+ industry experts, ALE covers non-physical industries defined with reference to O*NET / SOC 2018 (the U.S. federal occupational taxonomy). It is organized around a task taxonomy with 55 subfields grouped into 13 industry clusters covering 1K+ tasks. Current results show that the hardest tier remains far from saturated: across mainstream harness and backbone configurations, the average full pass rate is 2.6%. ALE is designed as a living benchmark: its task pool grows continuously as new workflows and industries are onboarded. More broadly, ALE is intended not merely as another leaderboard, but as an instrument for closing the gap between benchmark success and GDP-relevant impact.

While recent work has convincingly showed that sequence-to-sequence models struggle to generalize to new compositions (termed compositional generalization), little is known on what makes compositional generalization hard on a particular test instance. In this work, we investigate what are the factors that make generalization to certain test instances challenging. We first substantiate that indeed some examples are more difficult than others by showing that different models consistently fail or succeed on the same test instances. Then, we propose a criterion for the difficulty of an example: a test instance is hard if it contains a local structure that was not observed at training time. We formulate a simple decision rule based on this criterion and empirically show it predicts instance-level generalization well across 5 different semantic parsing datasets, substantially better than alternative decision rules. Last, we show local structures can be leveraged for creating difficult adversarial compositional splits and also to improve compositional generalization under limited training budgets by strategically selecting examples for the training set.

Large Reasoning Models (LRMs) benefit substantially from training on challenging competition-level questions. However, existing automated question synthesis methods lack precise difficulty control, incur high computational costs, and struggle to generate competition-level questions at scale. In this paper, we propose CoDiQ (Controllable Difficult Question Generation), a novel framework enabling fine-grained difficulty control via test-time scaling while ensuring question solvability. Specifically, first, we identify a test-time scaling tendency (extended reasoning token budget boosts difficulty but reduces solvability) and the intrinsic properties defining the upper bound of a model's ability to generate valid, high-difficulty questions. Then, we develop CoDiQ-Generator from Qwen3-8B, which improves the upper bound of difficult question generation, making it particularly well-suited for challenging question construction. Building on the CoDiQ framework, we build CoDiQ-Corpus (44K competition-grade question sequences). Human evaluations show these questions are significantly more challenging than LiveCodeBench/AIME with over 82% solvability. Training LRMs on CoDiQ-Corpus substantially improves reasoning performance, verifying that scaling controlled-difficulty training questions enhances reasoning capabilities. We open-source CoDiQ-Corpus, CoDiQ-Generator, and implementations to support related research.

We present MedConceptsQA, a dedicated open source benchmark for medical concepts question answering. The benchmark comprises of questions of various medical concepts across different vocabularies: diagnoses, procedures, and drugs. The questions are categorized into three levels of difficulty: easy, medium, and hard. We conducted evaluations of the benchmark using various Large Language Models. Our findings show that pre-trained clinical Large Language Models achieved accuracy levels close to random guessing on this benchmark, despite being pre-trained on medical data. However, GPT-4 achieves an absolute average improvement of nearly 27%-37% (27% for zero-shot learning and 37% for few-shot learning) when compared to clinical Large Language Models. Our benchmark serves as a valuable resource for evaluating the understanding and reasoning of medical concepts by Large Language Models. Our benchmark is available at https://huggingface.co/datasets/ofir408/MedConceptsQA

Recent thinking models solve complex reasoning tasks by scaling test-time compute, but this scaling must be allocated in line with task difficulty. On one hand, short reasoning (underthinking) leads to errors on harder problems that require extended reasoning steps; but, excessively long reasoning (overthinking) can be token-inefficient, generating unnecessary steps even after reaching a correct intermediate solution. We refer to this as under-adaptivity, where the model fails to modulate its response length appropriately given problems of varying difficulty. To address under-adaptivity and strike a balance between under- and overthinking, we propose TRAAC (Think Right with Adaptive, Attentive Compression), an online post-training RL method that leverages the model's self-attention over a long reasoning trajectory to identify important steps and prune redundant ones. TRAAC also estimates difficulty and incorporates it into training rewards, thereby learning to allocate reasoning budget commensurate with example difficulty. Our approach improves accuracy, reduces reasoning steps, and enables adaptive thinking compared to base models and other RL baselines. Across a variety of tasks (AIME, AMC, GPQA-D, BBEH), TRAAC (Qwen3-4B) achieves an average absolute accuracy gain of 8.4% with a relative reduction in reasoning length of 36.8% compared to the base model, and a 7.9% accuracy gain paired with a 29.4% length drop compared to the best RL baseline. TRAAC also shows strong generalization: although our models are trained on math datasets, they show accuracy and efficiency gains on out-of-distribution non-math datasets like GPQA-D, BBEH, and OptimalThinkingBench. Our analysis further verifies that TRAAC provides fine-grained adjustments to thinking budget based on difficulty and that a combination of task-difficulty calibration and attention-based compression yields gains across diverse tasks.

Recent works train agents by constructing large-scale multimodal environment pools. However, we find that simply increasing the number of multimodal environments does not always benefit. We further analyze the limitations in current multimodal environment distributions through a series of experiments. Based on these findings, we study how to build more effective training environment distributions from two dimensions: **diversity** and **difficulty structure**. For diversity, we propose **Ability-aware Environment Selection (AES)** to obtain diverse environment sets. For difficulty structure, we propose **Hierarchical Difficulty Curriculum (HDC)**, which organizes curriculum learning through two difficulty levels: harness weakening and state-scale progression. Experiments show that AES and HDC effectively improve multimodal agent training.

Predicting human item difficulty is central to educational assessment, where reliable estimates support fairness and effective test construction. Existing methods often depend on costly human calibration or item-level textual representations, providing limited evidence about the cognitive processes that make items difficult. We argue that difficulty should be viewed not only as a property of item text, but also as an observable consequence of the problem-solving burden an item induces. Large Reasoning Models (LRMs) offer scalable process evidence through reasoning traces, but such evidence must be structured to support interpretable modeling. To this end, we introduce Epi2Diff (Episode to Difficulty), a framework that maps LRM reasoning traces into cognitively grounded episode sequences. These episodes group trace segments into functional problem-solving states, enabling difficulty to be modeled through reasoning scale, effort allocation, and state transitions. Epi2Diff extracts compact episode-dynamic features and combines them with semantic item representations for human difficulty prediction. Experiments on four real-world human difficulty datasets show that Epi2Diff consistently outperforms strong baselines, including fine-tuned small language models, LLM in-context learning, and supervised LLM adaptation. On SAT-derived classification benchmarks, Epi2Diff achieves an 8.1% average relative gain over supervised LLM fine-tuning baselines. Further analyses show that harder items induce more effortful, iterative, and implementation-centered episode dynamics, rather than merely longer responses. These results demonstrate that cognitive episodes in LRM reasoning traces provide a predictive and interpretable process representation for human item difficulty, offering a new lens for educational measurement with reasoning models.

Measuring the full abilities of large language models (LLMs) requires benchmarks representing multiple tasks. We aim to create large, high-quality datasets for comparison of logical reasoning skills across several languages and of suitable difficulty for LLMs of various reasoning ability. We explore multiple ways of increasing difficulty. We generate zebra puzzles in multiple languages, themes, sizes and including 14 different clue types and 8 red herring types (uninformative clues). We find puzzle sizes 2x3 and 4x5 are sufficiently challenging for GPT-4o mini (a non-reasoning model) and o3-mini (a reasoning model), respectively. Including 5 red herrings decreases o3-mini puzzle-level accuracy on 4x5 puzzles by 15pm7 %. Scores of o3-mini on 4x5 puzzles are not significantly affected by use of English vs. Danish or the common houses theme vs. the country-specific smoerrebroed theme. We find no correlation between difficulty and the selected clue types. Datasets of 128+1024 puzzles are published as MultiZebraLogic in each of nine Germanic languages for sizes 2x3 and 4x5. We publish code for puzzle generation, designed for adaptablity into more languages and themes.

Determining when people are struggling allows for a finer-grained understanding of actions that complements conventional action classification and error detection. Struggle detection, as defined in this paper, is a distinct and important task that can be identified without explicit step or activity knowledge. We introduce the first struggle dataset with three real-world problem-solving activities that are labelled by both expert and crowd-source annotators. Video segments were scored w.r.t. their level of struggle using a forced choice 4-point scale. This dataset contains 5.1 hours of video from 73 participants. We conducted a series of experiments to identify the most suitable modelling approaches for struggle determination. Additionally, we compared various deep learning models, establishing baseline results for struggle classification, struggle regression, and struggle label distribution learning. Our results indicate that struggle detection in video can achieve up to 88.24% accuracy in binary classification, while detecting the level of struggle in a four-way classification setting performs lower, with an overall accuracy of 52.45%. Our work is motivated toward a more comprehensive understanding of action in video and potentially the improvement of assistive systems that analyse struggle and can better support users during manual activities.

We introduce Office Comprehension Bench (OCB), the first public benchmark to jointly evaluate LLM systems on Word, Excel, and PowerPoint comprehension over native file formats (.docx, .xlsx, .pptx) and their variants. OCB consists of two tracks. File Fidelity Q&A tests structural and visual perception of office artifacts - tables, charts, embedded images, formulas, and app-specific elements such as headers, speaker notes, and named ranges. Domain Q&A tests expert-level reasoning grounded in real-world industry documents across 12 professional domains, with queries requiring multi-step analysis and synthesis across documents. Each reference answer is decomposed into atomic, binary-gradable claims, and an ensemble of LLM judges scores responses against each claim independently. Even the strongest frontier system in its default reasoning mode reaches only about 59.3% on Domain Q&A; increasing thinking depth within a tier does not move performance materially, while moving to a higher product tier yields modest gains. We release the dataset, evaluation tooling, judge prompt, and a public leaderboard.

Large Reasoning Models (LRMs) have demonstrated impressive performance on complex tasks, including logical puzzle games that require deriving solutions satisfying all constraints. However, whether they can flexibly apply appropriate rules to varying conditions, particularly when faced with non-canonical game variants, remains an open question. Existing corpora focus on popular puzzles like 9x9 Sudoku, risking overfitting to canonical formats and memorization of solution patterns, which can mask deficiencies in understanding novel rules or adapting strategies to new variants. To address this, we introduce HardcoreLogic, a challenging benchmark of over 5,000 puzzles across 10 games, designed to test the robustness of LRMs on the "long-tail" of logical games. HardcoreLogic systematically transforms canonical puzzles through three dimensions: Increased Complexity (IC), Uncommon Elements (UE), and Unsolvable Puzzles (UP), reducing reliance on shortcut memorization. Evaluations on a diverse set of LRMs reveal significant performance drops, even for models achieving top scores on existing benchmarks, indicating heavy reliance on memorized stereotypes. While increased complexity is the dominant source of difficulty, models also struggle with subtle rule variations that do not necessarily increase puzzle difficulty. Our systematic error analysis on solvable and unsolvable puzzles further highlights gaps in genuine reasoning. Overall, HardcoreLogic exposes the limitations of current LRMs and establishes a benchmark for advancing high-level logical reasoning.

Math reasoning benchmarks have proliferated, yet most lack a per-item difficulty signal grounded in actual human performance. We introduce KCSAT-ML, a decade (2014-2025) of Korean College Scholastic Ability Test (KCSAT; Suneung) mathematics: 664 problems with a 339-item core set carrying official per-item error rates from nationwide cohorts of hundreds of thousands of examinees. We pair the benchmark with Difficulty-aligned Reasoning Gain (DRG): a score-orthogonal metric that asks whether a model's mistakes concentrate on the items humans found hard, or on items humans found easy. Together they expose, across a wide range of VLMs (and LLMs via OCR), three patterns: (i) low-budget accuracy collapses on the high-human-error tail at every model size; (ii) test-time scaling (TTS) raises token use roughly linearly with cohort error rate, while accuracy gains follow a non-monotonic curve; (iii) within a single family, TTS flips between anti-scaling on the hardest items and overthinking on easier ones -- two faces of the same alignment failure. On DRG, models with near-identical accuracy can sit at near-opposite values: one model gets wrong what humans also find hard, while another solves the hardest items yet fails on items humans find easy -- a contrast that aggregate accuracy hides. Our code and dataset builder will be open-sourced at https://github.com/naver-ai/KCSAT-ML.

Language models have been adopted faster than evaluation practices that reliably assess their risks, and a growing share of use is affective rather than informational. Within that traffic users disclose suicide and self-harm (SSH) crises. Existing safety benchmarks score models primarily on refusal, but refusal is not the desired behavior here; the conversation requires a substantive supportive response. We introduce DistressBench, 718 English conversations with clinician-authored user turns spanning 23 SSH subcategories, each paired with a task-specific rubric of weighted atomic criteria across seven dimensions of crisis support and a clinician reference response. A model's score is the weighted average over the criteria its reply satisfies, reading directly as the fraction of required care delivered. The best of 25 frontier models satisfies 88.4% of weighted criteria and the median model 71.4%, against 98.7% for clinician references scored on the same criteria. Pooled across all 25 models, in 35.3% of conversations where a model satisfied every criterion for recognizing the crisis, it failed to route the user toward human help. The shortfall concentrates in the dimensions requiring the model to act rather than withhold. Contrasting each dimension's median model against its best separates capability limits from choices: de-escalation reaches 74.6% in the strongest model against a median of 46.0%, and safety disclaimers 84.6% against 8.3%. Multi-turn conversations score below single-turn ones (68.5% against 74.2%). We release the corpus, rubrics, judge calibration labels, and a companion annex of 20 long-horizon red-team transcripts. These results argue for evaluating crisis safety on the care a model delivers rather than on what it declines to say, and for treating de-escalation and resource routing as deployment requirements.

Large language models exhibit a puzzling inconsistency: they solve complex problems yet frequently fail on seemingly simpler ones. We investigate whether LLMs internally encode problem difficulty in a way that aligns with human judgment, and whether this representation tracks generalization during reinforcement learning post-training. We train linear probes across layers and token positions on 60 models, evaluating on mathematical and coding subsets of Easy2HardBench. We find that human-labeled difficulty is strongly linearly decodable (AMC: rho approx 0.88) and exhibits clear model-size scaling, whereas LLM-derived difficulty is substantially weaker and scales poorly. Steering along the difficulty direction reveals that pushing models toward "easier" representations reduces hallucination and improves accuracy. During GRPO training on Qwen2.5-Math-1.5B, the human-difficulty probe strengthens and positively correlates with test accuracy across training steps, while the LLM-difficulty probe degrades and negatively correlates with performance. These results suggest that human annotations provide a stable difficulty signal that RL amplifies, while automated difficulty estimates derived from model performance become misaligned precisely as models improve. We release probe code and evaluation scripts to facilitate replication.

The measure of a machine learning algorithm is the difficulty of the tasks it can perform, and sufficiently difficult tasks are critical drivers of strong machine learning models. However, quantifying the generalization difficulty of machine learning benchmarks has remained challenging. We propose what is to our knowledge the first model-agnostic measure of the inherent generalization difficulty of tasks. Our inductive bias complexity measure quantifies the total information required to generalize well on a task minus the information provided by the data. It does so by measuring the fractional volume occupied by hypotheses that generalize on a task given that they fit the training data. It scales exponentially with the intrinsic dimensionality of the space over which the model must generalize but only polynomially in resolution per dimension, showing that tasks which require generalizing over many dimensions are drastically more difficult than tasks involving more detail in fewer dimensions. Our measure can be applied to compute and compare supervised learning, reinforcement learning and meta-learning generalization difficulties against each other. We show that applied empirically, it formally quantifies intuitively expected trends, e.g. that in terms of required inductive bias, MNIST < CIFAR10 < Imagenet and fully observable Markov decision processes (MDPs) < partially observable MDPs. Further, we show that classification of complex images < few-shot meta-learning with simple images. Our measure provides a quantitative metric to guide the construction of more complex tasks requiring greater inductive bias, and thereby encourages the development of more sophisticated architectures and learning algorithms with more powerful generalization capabilities.

This paper introduces DOoM, a new open-source benchmark designed to assess the capabilities of language models in solving mathematics and physics problems in Russian. The benchmark includes problems of varying difficulty, ranging from school-level tasks to university Olympiad and entrance exam questions. In this paper we discuss the motivation behind its creation, describe dataset's structure and evaluation methodology, and present initial results from testing various models. Analysis of the results shows a correlation between model performance and the number of tokens used, and highlights differences in performance between mathematics and physics tasks.

Large language models have demonstrated impressive performance on challenging mathematical reasoning tasks, which has triggered the discussion of whether the performance is achieved by true reasoning capability or memorization. To investigate this question, prior work has constructed mathematical benchmarks when questions undergo simple perturbations -- modifications that still preserve the underlying reasoning patterns of the solutions. However, no work has explored hard perturbations, which fundamentally change the nature of the problem so that the original solution steps do not apply. To bridge the gap, we construct MATH-P-Simple and MATH-P-Hard via simple perturbation and hard perturbation, respectively. Each consists of 279 perturbed math problems derived from level-5 (hardest) problems in the MATH dataset (Hendrycksmath et. al., 2021). We observe significant performance drops on MATH-P-Hard across various models, including o1-mini (-16.49%) and gemini-2.0-flash-thinking (-12.9%). We also raise concerns about a novel form of memorization where models blindly apply learned problem-solving skills without assessing their applicability to modified contexts. This issue is amplified when using original problems for in-context learning. We call for research efforts to address this challenge, which is critical for developing more robust and reliable reasoning models.

Medical benchmarks are indispensable for evaluating the capabilities of language models in healthcare for non-English-speaking communities,therefore help ensuring the quality of real-life applications. However, not every community has sufficient resources and standardized methods to effectively build and design such benchmark, and available non-English medical data is normally fragmented and difficult to verify. We developed an approach to tackle this problem and applied it to create the first Vietnamese medical question benchmark, featuring 14,000 multiple-choice questions across 34 medical specialties. Our benchmark was constructed using various verifiable sources, including carefully curated medical exams and clinical records, and eventually annotated by medical experts. The benchmark includes four difficulty levels, ranging from foundational biological knowledge commonly found in textbooks to typical clinical case studies that require advanced reasoning. This design enables assessment of both the breadth and depth of language models' medical understanding in the target language thanks to its extensive coverage and in-depth subject-specific expertise. We release the benchmark in three parts: a sample public set (4k questions), a full public set (10k questions), and a private set (2k questions) used for leaderboard evaluation. Each set contains all medical subfields and difficulty levels. Our approach is scalable to other languages, and we open-source our data construction pipeline to support the development of future multilingual benchmarks in the medical domain.

A central limitation of current data-free self-evolution methods for training search agents is their reliance on difficulty-based proposer rewards. These methods reward a proposer for generating questions that challenge a co-evolving solver, using solver difficulty as a proxy for question quality. Yet difficulty alone is insufficient to distinguish questions that require cross-passage evidence from those that are answerable via simpler shortcuts. In addition, measuring difficulty demands repeated solver rollouts for every candidate question, leading to substantial computational costs. In this paper, we introduce \methodname, the first self-evolving search framework that eliminates difficulty-based proposer rewards and directly optimizes for evidence necessity relative to shortcut contexts. Dr. Free samples relational chains from a knowledge graph and pairs them with aligned passages, giving question generation an explicit multi-hop structure. A generated question receives a positive information-gain reward only when the likelihood of the target answer under the complete evidence passages exceeds the maximum likelihood under all evaluated shortcut contexts. Because this signal is computed from teacher-forced likelihoods, it removes the need for pass-rate estimation and reduces proposer training time by over 7times. Experiments on seven open-domain QA benchmarks show that Dr. Free outperforms prior data-free search agents and the supervised baseline, with large improvements on multi-hop QA benchmarks.

Among the hardest tasks for humans are those found in competitive programming where problems require sophisticated algorithmic thinking, puzzle solving, and the creation of effective code. As a domain to assess language models (LMs), it has not received enough attention, though. This study presents the ICPC benchmark, which consists of 254 international collegiate programming contest (ICPC) tasks. Each problem includes official analysis, reference code, and sample, high-quality unit, and hidden tests. We are able to develop and evaluate a variety of LM inference techniques for competitive programming with these resources. With zero-shot chain-of-thought prompting, we find that o1 only achieves a 19.1\% pass@1 solve rate. With our best inference technique, which combines multi-turn self-judge with reflection and retrieval over episodic information, raises this to 42.2\%. Furthermore, we conduct a new human-in-the-loop investigation to gain a deeper understanding of the remaining difficulties. Surprisingly, we discover that o1 can solve 17 out of 18 problems that were previously unsolvable by any model or technique with just a few specific instructions. A footstep toward LMs with grounded, imaginative, and algorithmic thinking is provided by our quantitative findings and qualitative research. We open-source our code and data at https://github.com/kraritt/zolve.

The demand for Large Language Models (LLMs) capable of sophisticated mathematical reasoning is growing across industries. However, the development of performant mathematical LLMs is critically bottlenecked by the scarcity of difficult, novel training data. We introduce SAND-Math (Synthetic Augmented Novel and Difficult Mathematics problems and solutions), a pipeline that addresses this by first generating high-quality problems from scratch and then systematically elevating their complexity via a new Difficulty Hiking step. We demonstrate the effectiveness of our approach through two key findings. First, augmenting a strong baseline with SAND-Math data significantly boosts performance, outperforming the next-best synthetic dataset by uparrow 17.85 absolute points on the AIME25 benchmark. Second, in a dedicated ablation study, we show our Difficulty Hiking process is highly effective: by increasing average problem difficulty from 5.02 to 5.98, this step lifts AIME25 performance from 46.38\% to 49.23\%. The full generation pipeline, final dataset, and a fine-tuned model form a practical and scalable toolkit for building more capable and efficient mathematical reasoning LLMs. SAND-Math dataset is released here: https://huggingface.co/datasets/amd/SAND-MATH{https://huggingface.co/datasets/amd/SAND-MATH}

Estimating the difficulty of input questions as perceived by large language models (LLMs) is essential for accurate performance evaluation and adaptive inference. Existing methods typically rely on repeated response sampling, auxiliary models, or fine-tuning the target model itself, which may incur substantial computational costs or compromise generality. In this paper, we propose a novel approach for difficulty estimation that leverages only the hidden representations produced by the target LLM. We model the token-level generation process as a Markov chain and define a value function to estimate the expected output quality given any hidden state. This allows for efficient and accurate difficulty estimation based solely on the initial hidden state, without generating any output tokens. Extensive experiments across both textual and multimodal tasks demonstrate that our method consistently outperforms existing baselines in difficulty estimation. Moreover, we apply our difficulty estimates to guide adaptive reasoning strategies, including Self-Consistency, Best-of-N, and Self-Refine, achieving higher inference efficiency with fewer generated tokens.

Language Reasoning Models (LRMs) achieve strong performance by scaling test-time computation but often suffer from ``overthinking'', producing excessively long reasoning traces that increase latency and memory usage. Existing LRMs typically enforce conciseness with uniform length penalties, which over-compress crucial early deduction steps at the sequence level and indiscriminately penalize all queries at the group level. To solve these limitations, we propose \model, a dual-level framework for prefix-protected and difficulty-aware compression under hierarchical supervision. At the sequence level, prefix-protected optimization employs decaying mixed rollouts to maintain valid reasoning paths while promoting conciseness. At the group level, difficulty-aware penalty dynamically scales length constraints based on query complexity, maintaining exploration for harder questions while curbing redundancy on easier ones. Extensive experiments on DeepSeek-R1-Distill-Qwen (1.5B/7B) demonstrate that \model achieves a substantial reduction in token usage (up to 55.7\%) while simultaneously improving accuracy (up to 4.1\%) on math benchmarks, with generalization ability to code, science, and general domains.

We study the problem of controlling the difficulty level of text generated by Large Language Models (LLMs) for contexts where end-users are not fully proficient, such as language learners. Using a novel framework, we evaluate the effectiveness of several key approaches for this task, including few-shot prompting, supervised finetuning, and reinforcement learning (RL), utilising both GPT-4 and open source alternatives like LLama2-7B and Mistral-7B. Our findings reveal a large performance gap between GPT-4 and the open source models when using prompt-based strategies. However, we show how to bridge this gap with a careful combination of finetuning and RL alignment. Our best model, CALM (CEFR-Aligned Language Model), surpasses the performance of GPT-4 and other strategies, at only a fraction of the cost. We further validate the quality of our results through a small-scale human study.

Reinforcement learning (RL) post-training has enabled newer capabilities in models, such as agentic tool-use for search. However, these models struggle primarily due to limitations with sparse outcome-based rewards and a lack of training data that encapsulates questions of differing hardness, which results in models not performing deeper searches with tools to collect evidence for question-answering. To address these limitations, we introduce S^3-R1 (Synthetic data and stabilized Search R1), a framework that couples a data-centric approach with denser learning signals. We first develop a synthetic generation and curation pipeline that programmatically derives diverse, multi-hop questions from existing documents. This pipeline incorporates a retrieval-based verification step to specifically isolate questions of intermediate difficulty. We then pair this expanded training set with a reward structure that evaluates both intermediate search quality and the correctness of the final answer. This setup directly mitigates the credit assignment problems inherent to sparse rewards. Our evaluations show that S^3-R1 outperforms existing baselines by learning more effective search and synthesis strategies, yielding up to a 10% improvement in robust generalization on out-of-domain datasets.

We introduce ChemPro, a progressive benchmark with 4100 natural language question-answer pairs in Chemistry, across 4 coherent sections of difficulty designed to assess the proficiency of Large Language Models (LLMs) in a broad spectrum of general chemistry topics. We include Multiple Choice Questions and Numerical Questions spread across fine-grained information recall, long-horizon reasoning, multi-concept questions, problem-solving with nuanced articulation, and straightforward questions in a balanced ratio, effectively covering Bio-Chemistry, Inorganic-Chemistry, Organic-Chemistry and Physical-Chemistry. ChemPro is carefully designed analogous to a student's academic evaluation for basic to high-school chemistry. A gradual increase in the question difficulty rigorously tests the ability of LLMs to progress from solving basic problems to solving more sophisticated challenges. We evaluate 45+7 state-of-the-art LLMs, spanning both open-source and proprietary variants, and our analysis reveals that while LLMs perform well on basic chemistry questions, their accuracy declines with different types and levels of complexity. These findings highlight the critical limitations of LLMs in general scientific reasoning and understanding and point towards understudied dimensions of difficulty, emphasizing the need for more robust methodologies to improve LLMs.

We introduce a benchmark to evaluate the capability of AI to solve problems in theoretical physics, focusing on high-energy theory and cosmology. The first iteration of our benchmark consists of 57 problems of varying difficulty, from undergraduate to research level. These problems are novel in the sense that they do not come from public problem collections. We evaluate our data set on various open and closed language models, including o3-mini, o1, DeepSeek-R1, GPT-4o and versions of Llama and Qwen. While we find impressive progress in model performance with the most recent models, our research-level difficulty problems are mostly unsolved. We address challenges of auto-verifiability and grading, and discuss common failure modes. While currently state-of-the art models are still of limited use for researchers, our results show that AI assisted theoretical physics research may become possible in the near future. We discuss the main obstacles towards this goal and possible strategies to overcome them. The public problems and solutions, results for various models, and updates to the data set and score distribution, are available on the website of the dataset tpbench.org.

Evaluating large language models becomes increasingly challenging as their capabilities advance: benchmarks can saturate, public test sets risk contamination, and assessing harder tasks can require expensive grading or execution infrastructure. These challenges are amplified in automatic prompt optimization (APO), where evaluation is repeated throughout the search for better prompts. Studying APO therefore requires a benchmark that is cheap and deterministic to score, hard enough to leave room for improvement, and renewable as models evolve. We introduce a chess benchmark built from 1,118 Lichess puzzles to study APO for frozen LLMs: we optimize their prompts without updating their model weights. Chess combines inexpensive exact-match scoring, engine-based evaluation of alternative moves, and a renewable supply of problems with adjustable difficulty. Unlike evaluations that report only success on isolated test items, the benchmark also connects puzzle-solving gains to short game-play rollouts within the same domain. We use it to evaluate six APO algorithms on eight target models, measuring not only baseline strength but also how much each model responds to optimization and whether optimized prompts transfer across models and to game play. Chess is thus a well-suited benchmark for APO: it is (i) challenging, as even the strongest evaluated model, Gemini 3.5 Flash (used as the meta-model), solves only about 55\% of puzzles; (ii) discriminative, revealing gains, unchanged performance, and regressions across methods and models; (iii) renewable, with fresh puzzles to reduce contamination risk and adjustable difficulty to maintain headroom as models improve; and (iv) affordable, as the complete study runs for around \$800. We release the puzzles, optimization and evaluation code, and dataset-renewal scripts (https://github.com/imec-ailabs/Automatic-Prompt-Optimization-with-Chess).

In this work, we investigate how explicitly modeling problem's difficulty prior information shapes the effectiveness of reinforcement learning based fine-tuning for multimodal reasoning. Our exploration mainly comprises of following three perspective: First, through offline data curation, we analyze the U-shaped difficulty distribution of two given datasets using the base model by multi-round sampling, and then filter out prompts that are either too simple or extremely difficult to provide meaningful gradients and perform subsequent two-stage training. Second, we implement an online advantage differentiation, computing group-wise empirical accuracy as a difficulty proxy to adaptively reweight advantages estimation, providing stronger learning signals for more challenging problems. Finally, we introduce difficulty hints as explicit prompts for more complex samples in the second training stage, encouraging the model to calibrate its reasoning depth and perform reflective validation checks. Our comprehensive approach demonstrates significant performances across various multi-modal mathematical reasoning benchmarks with only 2K+0.6K two-stage training data.

Programming is a powerful and ubiquitous problem-solving tool. Developing systems that can assist programmers or even generate programs independently could make programming more productive and accessible, yet so far incorporating innovations in AI has proven challenging. Recent large-scale language models have demonstrated an impressive ability to generate code, and are now able to complete simple programming tasks. However, these models still perform poorly when evaluated on more complex, unseen problems that require problem-solving skills beyond simply translating instructions into code. For example, competitive programming problems which require an understanding of algorithms and complex natural language remain extremely challenging. To address this gap, we introduce AlphaCode, a system for code generation that can create novel solutions to these problems that require deeper reasoning. In simulated evaluations on recent programming competitions on the Codeforces platform, AlphaCode achieved on average a ranking of top 54.3% in competitions with more than 5,000 participants. We found that three key components were critical to achieve good and reliable performance: (1) an extensive and clean competitive programming dataset for training and evaluation, (2) large and efficient-to-sample transformer-based architectures, and (3) large-scale model sampling to explore the search space, followed by filtering based on program behavior to a small set of submissions.

Math reasoning is becoming an ever increasing area of focus as we scale large language models. However, even the previously-toughest evals like MATH are now close to saturated by frontier models (90.0% for o1-mini and 86.5% for Gemini 1.5 Pro). We introduce HARP, Human Annotated Reasoning Problems (for Math), consisting of 5,409 problems from the US national math competitions (A(J)HSME, AMC, AIME, USA(J)MO). Of these, 4,780 have answers that are automatically check-able (with libraries such as SymPy). These problems range six difficulty levels, with frontier models performing relatively poorly on the hardest bracket of 197 problems (average accuracy 41.1% for o1-mini, and 9.6% for Gemini 1.5 Pro). Our dataset also features multiple choices (for 4,110 problems) and an average of two human-written, ground-truth solutions per problem, offering new avenues of research that we explore briefly. We report evaluations for many frontier models and share some interesting analyses, such as demonstrating that frontier models across families intrinsically scale their inference-time compute for more difficult problems. Finally, we open source all code used for dataset construction (including scraping) and all code for evaluation (including answer checking) to enable future research at: https://github.com/aadityasingh/HARP.

Many court systems are overwhelmed all over the world, leading to huge backlogs of pending cases. Effective triage systems, like those in emergency rooms, could ensure proper prioritization of open cases, optimizing time and resource allocation in the court system. In this work, we introduce the Criticality Prediction dataset, a novel resource for evaluating case prioritization. Our dataset features a two-tier labeling system: (1) the binary LD-Label, identifying cases published as Leading Decisions (LD), and (2) the more granular Citation-Label, ranking cases by their citation frequency and recency, allowing for a more nuanced evaluation. Unlike existing approaches that rely on resource-intensive manual annotations, we algorithmically derive labels leading to a much larger dataset than otherwise possible. We evaluate several multilingual models, including both smaller fine-tuned models and large language models in a zero-shot setting. Our results show that the fine-tuned models consistently outperform their larger counterparts, thanks to our large training set. Our results highlight that for highly domain-specific tasks like ours, large training sets are still valuable.