Daily Papers

Large language models show large and seemingly arbitrary performance drops when a task is rewritten in ways that leave its requirements untouched. We recast task difficulty as a causal quantity: the treatment effect of meaning-preserving perturbations on a model. A meaning-preserving rewrite can only be observed to hurt a task the model can already solve, so we place the estimand on the principal stratum where the degradation occurs: the conditional causal difficulty (CCD), the probability that a randomly drawn twin breaks a solvable task. Because harder tasks tend to be more phrasing-sensitive, complexity enters as an effect modifier rather than a confounder, and we residualize it out to obtain complexity-residualized difficulty (CRD); zero for a task exactly as fragile as its complexity predicts, positive for anomalously fragile ones. Current benchmarks do not supply the data CCD requires, so we introduce GlassGSM and GlassMATH, augmenting GSM8K and MATH with twins built from perturbation families, raising fragility on GlassGSM by approx 12times over GSM-IC-style distractors on the same problems (3.7%to44.9%) while keeping the answer key unchanged. Our predictor Frail learns the per-perturbation break probability and reaches 0.613{pm}0.023 of the suite reliability ceiling on GlassGSM and 0.557{pm}0.023 on GlassMATH, against 0.171/0.401 for a surface-feature baseline and 0.444/0.463 for supervised baselines such as VAP3, and, at a 5% curation budget on GlassGSM, doubles random recall of fragile items, matching an oracle ranking. Additionally, cross-fitted residualization leaves CRD uncorrelated with the declared proxy in practice. Code is available at https://github.com/upwork/Conditional_Causal_Difficulty_for_LLMs

Agent evaluations are increasingly used to compare LLMs and inform deployment decisions, yet ranks can reflect not only the model but also the effects of the evaluation conditions such as the scaffolds or tasks. This makes reliability claim-dependent: an evaluation that reliably ranks deployed systems may not reliably rank underlying models. We ask which conclusions current agent evaluations reliably support and what additional evaluation would improve them. We develop a Bayesian variance-decomposition framework for sparse, imbalanced agent leaderboards and apply it to 22 benchmarks from the Holistic Agent Leaderboard and Harbor Index. The framework separates signal, performance differences relevant to the intended claim, from noise, irrelevant variation that can still change rankings. We find: (1) Reliability depends on the measurement goal. Fixed model-scaffold systems are ranked reliably (0.935-0.994), while underlying-model reliability is substantially lower (0.148-0.841). (2) Scaffold choice can change conclusions. Inter-scaffold reliability measures whether scaffolds preserve model rankings, showing that scaffold effects vary substantially across evaluations. (3) More tasks cannot resolve all uncertainty. Even infinitely many similarly constructed tasks improve model-ranking reliability of a benchmark by at most 0.097 when uncertainty is dominated by limited scaffold coverage. (4) Pooling diverse benchmarks can improve cross-task rankings at lower cost. For rankings across diverse agentic tasks, pooling benchmarks raises projected reliability from 0.44 to 0.75 at the same task budget and can reduce projected cost by up to 83\%. Evaluation design should follow the intended claim: identify what a score or ranking should mean, diagnose what limits its reliability, and spend evaluation budget on the sources of uncertainty that matter.

Red-team evaluations of AI models support some claims and not others, and the boundary between the two is calculable rather than merely a matter of judgment. We define the evidential ceiling of an evaluation as the largest factor by which one result can move belief under a fixed testing budget, derive it in closed form for the benchmark null result, and use it to locate that boundary exactly. We find that above a calculable harm rate, a benchmark of modest size certifies a category to a stated evidentiary standard, and a clean sheet is then the stronger of the two possible observations, outweighing a single reproduced failure. Below that rate, no passive benchmark of feasible size provides the specified evidence of safety under the fixed scoring rule and approximately independent trial structure. The crossing between the two regimes has a closed form. The bound is not specific to benchmarks: written in terms of a procedure's hypothesis conditioned elicitation rates, it covers adaptive and automated red teaming as well, and shows that discrimination between the hypotheses rather than attack success is what determines evidential worth. Auditing eight evaluation suites against the boundary, we find that current benchmarks are adequate for high-frequency harm categories and several orders of magnitude short for rare, catastrophic ones. Safety benchmarks are not uninformative. They are informative about a specific and computable set of propositions, and the discipline they need is to state which.

LLM agents increasingly rely on community-contributed skills that expand an agent's operational capability set. We study a core safety problem in agentic AI systems: whether individually safe skills can compose into unsafe installed skill sets. We present SkillReact, a compositional security measurement framework with three components: a deterministic static-composition benchmark, a two-rater LLM-assisted human-adjudication pipeline, and an action-based exploitability harness. On 1,520 ClawHub skills, 651 pass individual inspection and form 211,575 pairs; the benchmark flags 22.25% of these as structural candidates. We treat this raw rate as a recall-oriented scanner ceiling and calibrate it against human judgment: in a pattern-stratified audit, roughly one in five flagged pair-pattern hits survives as a real compositional risk (population-weighted validity 18.2%, our headline result), implying about 14K genuine risk memberships in a single registry that per-skill scanning misses by construction, since every pair is individually safe. An action-based harness then probes when these candidates become model-issued tool calls, and finds realization gated by host-model disposition: on an anchor-conditioned dropper subset, Haiku-4-5 issues the dropper-stage tool call on all 39 direct-prompt trials (36 of them the full download-then-execute chain, 3 download-only), Opus-4-7 stops at the download, and Sonnet-4-6 refuses outright. A control that holds the request fixed and varies only the installed skills finds compliance highest with no skills installed: a composition fixes which capabilities are reachable, while the host model decides whether to use them. Together these motivate install-time compositional checks and capability isolation as complements to per-skill scanning.

Many deployments must compare candidate language models for safety before a labeled benchmark exists for the relevant language, sector, or regulatory regime. We formalize this setting as benchmarkless comparative safety scoring and specify the contract under which a scenario-based audit can be interpreted as deployment evidence. Scores are valid only under a fixed scenario pack, rubric, auditor, judge, sampling configuration, and rerun budget. Because no labels are available, we replace ground-truth agreement with an instrumental-validity chain: responsiveness to a controlled safe-versus-abliterated contrast, dominance of target-driven variance over auditor and judge artifacts, and stability across reruns. We instantiate the chain in SimpleAudit, a local-first scoring instrument, and validate it on a Norwegian safety pack. Safe and abliterated targets separate with AUROC values between 0.89 and 1.00, target identity is the dominant variance component (η^2 approx 0.52), and severity profiles stabilize by ten reruns. Applying the same chain to Petri shows that it admits both tools. The substantial differences arise upstream of the chain, in claim-contract enforcement and deployment fit. A Norwegian public-sector procurement case comparing Borealis and Gemma 3 demonstrates the resulting evidence in practice: the safer model depends on scenario category and risk measure. Consequently, scores, matched deltas, critical rates, uncertainty, and the auditor and judge used must be reported together rather than collapsed into a single ranking.

Existing benchmarks measure capability -- whether a model succeeds on a single attempt -- but production deployments require reliability -- consistent success across repeated attempts on tasks of varying duration. We show these properties diverge systematically as task duration grows, and that pass@1 on short tasks is structurally blind to this divergence. We introduce a reliability science framework for long-horizon LLM agents with four metrics: Reliability Decay Curve (RDC), Variance Amplification Factor (VAF), Graceful Degradation Score (GDS), and Meltdown Onset Point (MOP). We evaluate 10 models across 23,392 episodes on a 396-task benchmark spanning four duration buckets and three domains. Key findings: (1) reliability decay is domain-stratified -- SE GDS drops from 0.90 to 0.44 while document processing is nearly flat (0.74 to 0.71); (2) VAF bifurcates by capability tier -- high VAF is a capability signature, not an instability signal; (3) capability and reliability rankings diverge substantially, with multi-rank inversions at long horizons; (4) frontier models have the highest meltdown rates (up to 19%) because they attempt ambitious multi-step strategies that sometimes spiral; and (5) memory scaffolds universally hurt long-horizon performance across all 10 models. These results motivate reliability as a first-class evaluation dimension alongside capability.

Given a black-box AI system and a task, at what confidence level can a practitioner trust the system's output? We answer with a reliability level -- a single number per system-task pair, derived from self-consistency sampling and conformal calibration, that serves as a black-box deployment gate with exact, finite-sample, distribution-free guarantees. Self-consistency sampling reduces uncertainty exponentially; conformal calibration guarantees correctness within 1/(n+1) of the target level, regardless of the system's errors -- made transparently visible through larger answer sets for harder questions. Weaker models earn lower reliability levels (not accuracy -- see Definition 2.4): GPT-4.1 earns 94.6% on GSM8K and 96.8% on TruthfulQA, while GPT-4.1-nano earns 89.8% on GSM8K and 66.5% on MMLU. We validate across five benchmarks, five models from three families, and both synthetic and real data. Conditional coverage on solvable items exceeds 0.93 across all configurations; sequential stopping reduces API costs by around 50%.

Benchmark gains are often treated as evidence of greater LLM capability. Yet the same gain can reflect different changes in model behavior. A model may reach new answers, or produce answers that were already within reach. Aggregate scores do not distinguish these changes question by question. We establish a question-level audit under fixed budgets, temperatures, and answer formats. A question is realized when the default deployment procedure produces the correct answer. A question is reachable when a specified probe finds that answer within a fixed budget. We first test whether inference-time layer routing can expand reachability. Under a matched budget, random routes match or exceed structured search in all 43 model and task settings. Answer-blind procedures retain almost none of this gain, which instead requires access to the correct answer. We then ask why reachable answers sometimes fail to appear. Across six cases spanning 0.5B to 31B, silencing one identified MLP block repairs 68 to 92 percent of a predefined failure set. We next test whether training closes the gap by expanding reachability. In five of six matched evaluations, deployed performance rises while the reachable ceiling remains flat or falls. For DAPO, the deployed score rises by 14.7 points while the reachable ceiling falls by 13.3 points. Across the settings we audit, realization and reachability therefore do not always change together. Claims of capability expansion should report both realized performance and reachability under matched evaluation conditions. Code is available at https://github.com/LiZaiyuan0619/reachability-not-realization

Agent skill libraries are becoming routable software assets: a retrieved skill can contribute instructions, scripts, resource bindings, and execution assumptions to an agent. This makes retrieval failures more specific than broad irrelevance. A system can find the right capability family yet expose the wrong same-capability representative. We study this failure as same-capability risk-exposure retrieval. Each benchmark unit pairs a helpful skill with a query-specific risky sibling that shares the capability family but differs on an execution-controlling contract, such as the required resource, precondition, procedure, or artifact. We introduce SameCapRisk-Bench, an auditable benchmark with 1,190 skill-risk units and 1,686 evaluation query cases: 694 marked-sibling units under public library pressure and 496 hard role-flip units where the same two skills swap helpful/risky roles across paired queries. The release records admission evidence, cue/leakage checks, source hashes, family relations, and fixed candidate pools. The benchmark reports helpful ranking together with harmful sibling rate (HSR@K), the top-K exposure of the marked risky sibling. On this benchmark, public SkillRouter, SkillRet, and R3-Skill retrieve helpful skills at high Recall@3 (0.848--0.888) but also expose marked risky siblings frequently (HSR@3 0.346--0.372). A fully public score-and-cluster pipeline lowers HSR@3 to 0.128--0.182, with Recall@3 of 0.713--0.776. Under a benchmark-trained reference scorer, public text-cluster and controlled resolvers reach HSR@3 0.012 and 0.007; the latter attains Recall@3 0.833. Skill retrieval should therefore report both capability matching and same-family risk exposure, with HSR serving as a targeted exposure certificate for fixed skill libraries.

Practitioners make three prompt-design decisions with almost no controlled evidence behind them: how to format instructions and context (markdown, plain text, prose, or tabular), how many simultaneous instructions a system prompt can carry before compliance degrades, and how much context a model can hold before recall and honesty degrade. We report two controlled experiments crossing all three factors on one held, contamination-free synthetic corpus (the "Book of Veyra," 8,780 uniquely-named entities, deterministically regenerable from a fixed seed), evaluated across five models. Experiment 1 (960 calls/model) measures instruction-following decay as rule count N grows from 10 to 160, crossed with four formats and system-prompt vs. user-turn placement. Perfect-response rate collapses to zero by N=80 for every model, format, and placement. Placement produces effects at least as large as format at N=160 in most models, but the direction is model-specific. No model shows a reliable markdown advantage; one 35B model favors plain text instead. Experiment 2 (5,520 calls/model) measures recall accuracy, false-premise sycophancy, and absent-fact fabrication across a 2k-to-512k-token context ladder in the same four formats. Recall stays near ceiling through 64-128k tokens, then degrades sharply and format-dependently: one model's accuracy spread reaches 48 points at 128k tokens. Fabrication never occurs (0/5,760 probes), and sycophancy stays negligible (<=8.3%). What rises sharply near each model's context ceiling is outright refusal to answer (0% to 79-90%), distinct from sycophancy or fabrication. Neither pre-registered format ordering holds, and token overhead (+22% to +37% over plain text) further changes which format is preferable where accuracy spread is genuine. We release the full harness, corpus generator, and raw results (VeyraBench): https://github.com/iNetanel/veyrabench

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.

Agent harness evolution enables recursive self-improvement (RSI) without updating model weights. We examine whether higher benchmark scores yield reliable, efficient, and transferable improvement. We evaluate four methods across harnesses, general and coding tasks. Repeated evaluations, controlled agent expansion, component interventions, and admission replays reveal recurring patterns. All sixteen evolved agents gain previously uncovered tasks while losing previously covered ones. In one general-task case, reliability rises by 5.2 points despite losing ten previously covered tasks. Most reliability gains also increase deployment cost. Specific changes usually outperform generic reminders, but their benefits can change when combined. The highest-scoring agent does not yield the largest mean improvement in any of our four pools, and stricter admission can discard useful alternatives. Transfer selectively preserves gains: coverage usually falls on the general-task target across both harness types and paradigms, but rises on the coding target with the outer-layer-only harness. These findings motivate preservation-aware evaluation, tested decision rules, and explicit applicability conditions. As a constructive follow-up, EvoAudit turns these diagnostics into revisions of select, analyze, mutate, and gate, achieving the best performance.

We introduce Tencent WorkBuddy Bench, a multi-domain evaluation suite for coding agents; this report documents its construction methodology, scoring protocol, and a cross-model leaderboard. At its core is a unified evaluation framework for constructing and running distribution-informed coding-agent tasks across four work domains - Code, Web, Office, and Security. Rather than adapting public issue text, every task is reverse-engineered from a real commit, pull request, or business scenario and rewritten as a short, colloquial, role-played request, so that a task's prompt is not recoverable by web-searching the underlying issue, pull request, or commit thread. Because the dataset is released openly - task directories, environment images, evaluation harness, tests, and reference solutions - contamination resistance rests on this construction together with dataset versioning rather than on secrecy. The four subsets - repository-level engineering, front-end development, office and business workflows, and red-/blue-team security - probe complementary facets of real work, each with its own verification style. All are packaged in a uniform task-directory format and run, under a uniform and reproducible protocol, on two agent harnesses (CodeBuddy Code and Claude Code); the full open release makes the benchmark reproducible end to end and directly auditable, since any third party can re-run each task and inspect its content. Because each subset uses a different scoring instrument, scores are not comparable across subsets and the suite reports no suite-wide average. We report a cross-model leaderboard across several model families.

Reinforcement learning is a natural way to post-train LLM agents for long-horizon interactive tasks judged only by end-of-task verification, yet a shared belief holds that outcome-only RL soon hits a ceiling on small open models. Recent work therefore compensates around the training with denser rewards, SFT priors, skill libraries, curated memory, or multi-agent orchestration. We argue the ceiling is an artifact of two failures of common practice. Signal starvation: group-relative RL with sparse outcome-only rewards yields a gradient only when a task's rollout group mixes successes and failures, so under-scaled exploration silences exactly the hardest, most instructive tasks. Policy drift: squeezing many updates out of a small task pool degrades the policy itself, as an unanchored objective lets the sampling distribution collapse exactly when saturation has already made informative groups rare. We present CANOPY (Coverage-ANchored On-PolicY RL), a minimalist protocol attacking both directly: scale same-task exploration until the natural signal reappears, keep every update on-policy, KL-anchored, and confined to the agent's own action tokens, then cash in an enlarged interaction budget at test time. On AppWorld, a long-horizon interactive coding benchmark, a Qwen3-14B policy trained with CANOPY through environment interaction alone--without task-specific supervision, auxiliary credit signals, or elaborate agent scaffolding--topped the public leaderboard (Feb. 2026; Test-Normal TGC 86.9, Test-Challenge 67.6), and the same design principles lift Qwen3.5-9B on SWE-bench Verified by 16.6 points. Agentic RL alone thus internalizes long-horizon capability directly into a small open model; we plan to release the complete training stack at https://github.com/AlibabaResearch/SignalCoverageRL.

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.

Advanced LLMs have achieved near-ceiling instruction-following accuracy on benchmarks such as IFEval. However, these impressive scores do not necessarily translate to reliable services in real-world use, where users often vary their phrasing, contextual framing, and task formulations. In this paper, we study nuance-oriented reliability: whether models exhibit consistent competence across cousin prompts that convey analogous user intents but with subtle nuances. To quantify this, we introduce a new metric, reliable@k, and develop an automated pipeline that generates high-quality cousin prompts via data augmentation. Building upon this, we construct IFEval++ for systematic evaluation. Across 20 proprietary and 26 open-source LLMs, we find that current models exhibit substantial insufficiency in nuance-oriented reliability -- their performance can drop by up to 61.8% with nuanced prompt modifications. What's more, we characterize it and explore three potential improvement recipes. Our findings highlight nuance-oriented reliability as a crucial yet underexplored next step toward more dependable and trustworthy LLM behavior. Our code and benchmark are accessible: https://github.com/jianshuod/IFEval-pp.

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.

Generative CAD models are expected to remain behaviorally correct after parameter edits, so increasing the number of edit checks is often treated as a direct route to more reliable evaluation. Under a fixed budget, however, auditing each program more thoroughly reduces the number of tasks and independent generations that can be evaluated, which can ultimately make model-level estimates less accurate. We study this phenomenon and the conditions under which it arises. We decompose behavioral evaluation into three evidence levels: task templates, stochastic generations, and within-program edits. We define an average failure risk that is invariant to audit depth, and combine three-level variance with measured execution costs to analyze the tradeoff between deeper edit auditing and broader independent coverage. Experiments across two CAD environments and five generation systems show that the value of deeper auditing depends on where evaluation uncertainty originates. When template heterogeneity or generation stochasticity dominates, additional edit checks can increase total estimation error; when within-program state variation is large and generation is expensive, deeper auditing is more valuable. Variance and cost estimates from calibration predict the direction of this change and provide a diagnostic basis for allocating evidence on held-out tasks. These results show that the thoroughness of program inspection can diverge from the reliability of model evaluation, and they help determine whether the next unit of budget should be spent on a new task, a new generation, or additional edit checks.

Autonomous language-model agents increasingly rely on installable skills and tools to complete user tasks. Static skill auditing can expose capability surface before deployment, but it cannot determine whether a particular invocation is unsafe under the current user request and runtime context. We therefore study skill invocation auditing as a continuous-risk estimation problem: given a user request, candidate skill, and runtime context, predict a score that supports ranking and triage before a hard intervention is applied. We introduce STARS, which combines a static capability prior, a request-conditioned invocation risk model, and a calibrated risk-fusion policy. To evaluate this setting, we construct SIA-Bench, a benchmark of 3,000 invocation records with group-safe splits, lineage metadata, runtime context, canonical action labels, and derived continuous-risk targets. On a held-out split of indirect prompt injection attacks, calibrated fusion reaches 0.439 high-risk AUPRC, improving over 0.405 for the contextual scorer and 0.380 for the strongest static baseline, while the contextual scorer remains better calibrated with 0.289 expected calibration error. On the locked in-distribution test split, gains are smaller and static priors remain useful. The resulting claim is therefore narrower: request-conditioned auditing is most valuable as an invocation-time risk-scoring and triage layer rather than as a replacement for static screening. Code is available at https://github.com/123zgj123/STARS.

Inference-time multi-agent LLM scaling lacks a shared unit: counting nominal agents conflates cost with independent evidence. We derive a two-parameter scaling law R(N) = N_eff/N = 1/(1+c(N-1)N^{-β}) where the regime exponent β classifies any configuration into one of three asymptotic regimes -- hard-ceiling at 1/c (β= 0), sublinear at N^β/c (0 < β< 1), or linear (βge 1), and a mean-field theorem predicts that peer count k and rounds τ during agent debate enter the dynamics only through their product kτ. The law applies at two levels: answer diversity and correctness redundancy. Across 44 (model times task times condition) cells spanning peer debate, self-correction, random-noise placebo, self-consistency, three open-weight families (Qwen, Llama, Ministral) at scales from 7B to 32B with a frontier API check (Gemini), thinking models, heterogeneous teams, and sparse communication, the functional form fits every condition at R^2 > 0.99; only (c, β) shifts. On free-form math, dense peer influence collapses the answer-level regime from sublinear into hard-ceiling; correctness-level fits remain hard-ceiling throughout. Three findings have practical implications. (i)~Thirty dense debating agents produce no more answer diversity than one on MMLU-Hard. (ii)~A noise placebo tracks self-correction on free-form math and at 4times scale, so within homogeneous teams the gain commonly attributed to ``debate'' comes from re-evaluation, not peer content. (iii)~A single N le 5 pilot predicts the N=30 structural ceiling, and within the configurations tested only architectural diversity (heterogeneous teams) lowers c and escapes the hard-ceiling regime, communication-mode interventions do not.

Autonomous agents increasingly perform bounded software tasks under an orchestrator that retries, resumes, and budgets them. The machinery such orchestrators reach for is the service mesh's: retry, timeout, and error-rate circuit breaking. We report a failure study of a production agentic software-delivery platform over 147 numbered incidents spanning 81 runs, each with a measured cost and, in most cases, a mutation proof reproducing the failure. All three assumptions those primitives rest on are violated in practice, and we quantify the consequences: a loop of fifty-four consecutive successful tool calls no error-rate breaker could see; a progress signal constant by construction, guaranteeing a false trip on the third repair round and driving one run from six of six components to three; twenty-one events accumulated across six invocations of one delegation, making a correct, idempotent component unwinnable; a misrouted failure that woke five components for a two-component fault, leaving three bystanders regressing working code; and twelve incidents in which the enforcement layer blocked correct work, the most expensive costing 107 agent turns and zero accepted writes. We find one cross-cutting cause and its dual. Identity adequacy: in five separate subsystems an identity that failed to discriminate produced a confident wrong answer, and two of them derived the corrective rule independently. Evidence adequacy: a reliability decision may be taken only on evidence capable of moving, attributable to what it measures, and deterministic under identical conditions. From the findings we derive seven reliability primitives whose enforcement unit is the delegation rather than the message, and specify the controlled evaluation the study motivates but does not constitute.

When can an agent failure be caught? An audit is usually limited by the record rather than by the method. CatchBench therefore puts one auditor's question to three information states: the declared configuration before a run (PRE), a growing prefix of its trace (LIVE), and the finished trace (POST). Prior benchmarks fix one of these states or vary the telemetry; to our knowledge none scores all three under one task-method interface. Each state admits different questions, so seven task contracts carry their own labels and metrics rather than one leaderboard. Four are evidential; three are Gold-derived mechanism diagnostics. The release scores 72 entrants, from rule scanners and structural models to eleven LLM judges across nine model families (GPT, Claude, Gemini, Gemma, Llama, Qwen, DeepSeek, Mistral, Nova), over 1187 declared configurations and 1162 recorded runs. Most of the arena does not order: 56 of 138 registered contrasts separate, and the rest are published unresolved rather than ranked. The two sharpest results cut against our own data. One rule ignores every name and permission; it flags each capability declared after the first. On one of six configuration sources it reaches a perfect F1, so a score there measures how the corpus was built rather than how well a method reasons. Our admissibility bar then rejected one injected substrate and withheld evidential status from the other. A benchmark number is therefore not interpretable until the process behind its labels is published and tested for the shortcut it may leave. We report both, and regenerate every ordering from released predictions with no model call.

We study reliability in autonomous language-model agents that translate user mandates into validated tool actions under real capital. The setting is DX Terminal Pro, a 21-day deployment in which 3,505 user-funded agents traded real ETH in a bounded onchain market. Users configured vaults through structured controls and natural-language strategies, but only agents could choose normal buy/sell trades. The system produced 7.5M agent invocations, roughly 300K onchain actions, about $20M in volume, more than 5,000 ETH deployed, roughly 70B inference tokens, and 99.9% settlement success for policy-valid submitted transactions. Long-running agents accumulated thousands of sequential decisions, including 6,000+ prompt-state-action cycles for continuously active agents, yielding a large-scale trace from user mandate to rendered prompt, reasoning, validation, portfolio state, and settlement. Reliability did not come from the base model alone; it emerged from the operating layer around the model: prompt compilation, typed controls, policy validation, execution guards, memory design, and trace-level observability. Pre-launch testing exposed failures that text-only benchmarks rarely measure, including fabricated trading rules, fee paralysis, numeric anchoring, cadence trading, and misread tokenomics. Targeted harness changes reduced fabricated sell rules from 57% to 3%, reduced fee-led observations from 32.5% to below 10%, and increased capital deployment from 42.9% to 78.0% in an affected test population. We show that capital-managing agents should be evaluated across the full path from user mandate to prompt, validated action, and settlement.

Recursive self-improvement (RSI) asks whether an AI system can improve the process that produces AI systems, so that the next system inherits the improvement. That process is the training algorithm: a better objective or update rule improves the compute-capability exchange rate for every subsequent run, including the one that produces the next agent. Whether RSI is feasible therefore turns on whether an agent can design training algorithms. No benchmark isolates that ability: existing suites are won by collecting data or by tuning hyperparameters, and none tells a change to how a run is executed apart from a change to how the model learns. We present AI4AI-Bench, 10 frozen research repositories spanning 10 training algorithm families. In each task, an agent has 4 hours on one B300 to rewrite the training algorithm; its code is then rerun from scratch for up to 12 hours and scored by a fixed evaluator hidden from the agent, against the repository's original algorithm under the same procedure. Because the 10 metrics are incommensurable, every task is mapped onto one scale on which 0 is an uninformative model, 0.1 is the algorithm the repository ships, and 1.0 is the task optimum. Across 29 configurations of 6 systems on all 10 tasks the mean score is 0.166, and the best system reaches 0.250: even the strongest closes under a fifth of the distance between the algorithm that was already there and the optimum. The submissions show where that distance went: most never change how the model learns at all, and the minority that do average 0.226 against 0.126 for the rest. More reasoning effort mostly buys the willingness to go there, taking that minority from 8% of submissions to 64% and the mean score from 0.094 to 0.196. We release the task suite, the evaluators and every scored submission, so that the measurement can be repeated as these systems change.

The BITEM team entered both subtasks of the NTCIR-19 R2C2 task with a single agentic pipeline, in which a model searches, reads and records evidence over a movie corpus while an orchestrator holds the record and rules on what may be submitted. A claim is admitted only once an entailment cascade has checked it against the passage it cites, and an answer is released only once enough checked evidence stands behind it. Each question is run three or four times, every pass retrieving from a corpus stripped of what the earlier passes have already seen. The confidence filed with each answer is computed by the orchestrator from what the run leaves behind and is never asked of the model, which is offered no way to rate itself. The two retrieval runs placed 4th and 5th of 22, pooling the passes was worth 0.0709 nDCG@20, and the gain was largest on the multi-hop and post-processing-heavy questions, where the organisers rank the pooled run top of the field. Sixteen of the 25 answer runs were built on passages these two runs supplied, 12 of them filed by other teams. HMR rewards a system whose confidence is high where it answers right and low where it answers wrong. The pipeline reached an accuracy of 0.9219, 6th of 25, while the confidence filed with those answers gave an HMR of 0.4915, 13th. A few rules crafted over those same recorded signals, with no further model call and no further retrieval, raise that to an accuracy of 0.9375, 5th, and an HMR of 0.6985, 9th. Ranking on HMR alone can reward a system for answering wrongly with low confidence, so we propose accHMR, the accuracy multiplied by HMR, which reports the reward in proportion to the accuracy, and on which the revised rules would have scored 0.6549, 5th. For future work, fitting a model on the numbers the pipeline already produces, rather than writing such rules by hand, would be a real step forward.

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.

We study KV cache eviction under a shared globally capped decode-time harness. Seven policies (LRU, H2O, SnapKV, StreamingLLM, Ada-KV, QUEST, Random) share a prompt-boundary vulnerability: without structural protection, they collapse to near-zero quality on six pure-transformer models (F1leq0.064). Reserving 10\% of cache at each boundary recovers 69--90\% of the C{=}2{,}048 reference-ceiling quality on seven LongBench models at C{=}256 (13\% retention); a ten-model panel spans 68--98\%. An attention-mass pilot (Qwen2.5-3B, N{=}30) suggests why: the position-0 sink holds {sim}75% of prefix mass, while other boundary tokens sit near {sim}0.41{times} uniform expectation, so attention scorers retain the sink but still drop structurally critical tokens. With protection, simplified score-isolation variants are TOST-equivalent to LRU at K{=}32 (Δ{=}0.02); at K{=}8, attention policies pairwise converge yet beat LRU by 0.011--0.021 F1 across C{=}256 and C{=}512. Faithful Ada-KV/QUEST add {sim}0.03--0.04 F1 on Mistral-7B and Phi-3.5 beyond simplified variants. A NIAH-32K regime-transfer pilot on Qwen3-4B (decode vs.\ prefill, C{in}{512,2048}) shows near-identical protection lifts (ratio 0.99--1.00). At 64K, protection helps but recovery is modest; faithful per-head scoring matches full-cache ceiling on Gemma-3-4B at 6.3\% retention only when the model already supports strong 64K retrieval without eviction. Overall: protection dominates; scoring differences are secondary once boundaries are guarded; per-head allocation gives a further modest gain.

Multi-step enterprise agent tasks fail in a characteristic way: single-pass inference has no checkpoint between deciding an answer and committing to it. We study one production system (Leni) whose architecture installs such checkpoints: verification loops (execute, observe, compare, correct) staffed by lightweight task-specialized post-trained models. We evaluate the unmodified production configuration on three public benchmarks stressing distinct failure modes: SpreadsheetBench Verified (silent computation error), BullshitBench v2 (premise confabulation), and the GAIA validation split (cascade error over long tool chains). The full system improves over its frontier base model by +11.0 percentage points on SpreadsheetBench (91.25% vs 80.25%, n=400, p<0.001), +7 to +10 percentage points on BullshitBench (98% vs 91%, n=100), and roughly +15 points on GAIA validation (75.2% pass@1, n=165; 83.0% best-of-k). Our central contribution is a decomposition of that uplift: most of it comes from scaffolding, routing, and specialist models rather than from the verification step itself, whose isolated contribution is small (+1.5 points) but concentrated at the top of the score distribution, where it converts otherwise-failing tasks. We instrument the loop end-to-end, yielding an empirical verifier confusion matrix (catch rate about 0.20, fix rate 0.75, no false-alarm regressions) that grounds a compounding-reliability model. Specialist-swap ablations suggest that the loop's value depends on who observes it: replacing the small trained verifier with the generating frontier model eliminates most rescues. A valid-premise control shows zero over-rejections in 100 expert-level questions.

The standard paradigm for solving coding tasks via large language models (LLMs) is to generate-then-rank programs, where the latter step uses a verifier in the ranking process. The growing consensus is that a comprehensive verifier (e.g., a full test suite) should be prioritized over an outcome reward model (ORM) whenever possible, with little consideration given to the trade-offs involved. We aim to challenge this assumption by systematically exploring the tradeoff between speed and accuracy. We find that ORMs play a crucial role in scaling verification through trading accuracy for speed, even when a comprehensive verifier is available. Their value becomes especially apparent when used in a generate-prune-then-rank approach, where a faster but less accurate verifier removes incorrect solutions prior to ranking -- leading to a system that is 11.65x faster while only being 8.33% less accurate than the full test suite. We analyze the generate-prune-then-rank approach and show that it works by filtering out incorrect but highly ranked solutions. These findings enable the design of scalable and accurate program ranking systems.

Large language models for code generation increasingly rely on synthetic data, where both problem solutions and verification tests are generated by models. While this enables scalable data creation, it introduces a previously unexplored bottleneck: the verification ceiling, in which the quality and diversity of training data are fundamentally constrained by the capabilities of synthetic verifiers. In this work, we systematically study how verification design and strategies influence model performance. We investigate (i) what we verify by analyzing the impact of test complexity and quantity: richer test suites improve code generation capabilities (on average +3 pass@1), while quantity alone yields diminishing returns, (ii) how we verify by exploring relaxed pass thresholds: rigid 100% pass criteria can be overly restrictive. By allowing for relaxed thresholds or incorporating LLM-based soft verification, we can recover valuable training data, leading to a 2-4 point improvement in pass@1 performance. However, this benefit is contingent upon the strength and diversity of the test cases used, and (iii) why verification remains necessary through controlled comparisons of formally correct versus incorrect solutions and human evaluation: retaining diverse correct solutions per problem yields consistent generalization gains. Our results show that Verification as currently practiced is too rigid, filtering out valuable diversity. But it cannot be discarded, only recalibrated. By combining calibrated verification with diverse, challenging problem-solution pairs, we outline a path to break the verification ceiling and unlock stronger code generation models.

Large language model (LLM) agents now execute long, tool-using tasks where final outcome checks can arrive too late for intervention. Online warning requires lightweight prefix monitors over heterogeneous traces, but hand-authored event schemas are brittle and deployment-time LLM judging is costly. We introduce PrefixGuard, a trace-to-monitor framework with an offline StepView induction step followed by supervised monitor training. StepView induces deterministic typed-step adapters from raw trace samples, and the monitor learns an event abstraction and prefix-risk scorer from terminal outcomes. Across WebArena, τ^2-Bench, SkillsBench, and TerminalBench, the strongest PrefixGuard monitors reach 0.900/0.710/0.533/0.557 AUPRC. Using the strongest backend within each representation, they improve over raw-text controls by an average of +0.137 AUPRC. LLM judges remain substantially weaker under the same prefix-warning protocol. We also derive an observability ceiling on score-based area under the precision-recall curve (AUPRC) that separates monitor error from failures lacking evidence in the observed prefix. For finite-state audit, post-hoc deterministic finite automaton (DFA) extraction remains compact on WebArena and τ^2-Bench (29 and 20 states) but expands to 151 and 187 states on SkillsBench and TerminalBench. Finally, first-alert diagnostics show that strong ranking does not imply deployment utility: WebArena ranks well yet fails to support low-false-alarm alerts, whereas τ^2-Bench and TerminalBench retain more actionable early alerts. Together, these results position PrefixGuard as a practical monitor-synthesis recipe with explicit diagnostics for when prefix warnings translate into actionable interventions.

Background: Agent skills are increasingly deployed as modular, reusable capability units in AI agent systems. Medical research agent skills require safeguards beyond general-purpose evaluation, including scientific integrity, methodological validity, reproducibility, and boundary safety. This study developed and preliminarily evaluated a domain-specific audit framework for medical research agent skills, with a focus on reliability against expert review. Methods: We developed MedSkillAudit (skill-auditor@1.0), a layered framework assessing skill release readiness before deployment. We evaluated 75 skills across five medical research categories (15 per category). Two experts independently assigned a quality score (0-100), an ordinal release disposition (Production Ready / Limited Release / Beta Only / Reject), and a high-risk failure flag. System-expert agreement was quantified using ICC(2,1) and linearly weighted Cohen's kappa, benchmarked against the human inter-rater baseline. Results: The mean consensus quality score was 72.4 (SD = 13.0); 57.3% of skills fell below the Limited Release threshold. MedSkillAudit achieved ICC(2,1) = 0.449 (95% CI: 0.250-0.610), exceeding the human inter-rater ICC of 0.300. System-consensus score divergence (SD = 9.5) was smaller than inter-expert divergence (SD = 12.4), with no directional bias (Wilcoxon p = 0.613). Protocol Design showed the strongest category-level agreement (ICC = 0.551); Academic Writing showed a negative ICC (-0.567), reflecting a structural rubric-expert mismatch. Conclusions: Domain-specific pre-deployment audit may provide a practical foundation for governing medical research agent skills, complementing general-purpose quality checks with structured audit workflows tailored to scientific use cases.

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.

Managing modern computing infrastructure has become a steadily harder problem due to the ever-increasing complexity. Recent advances in AI agents create a timely opportunity to automate infrastructure management tasks, but it remains unclear how well such agents can handle real-world infrastructure complexity. We present InfraBench, a benchmark suite for evaluating AI agents on realistic infrastructure tasks across the full system stack and full operational lifecycle with fine-grained risk assessment. Experiments with 15 agent-model configurations show that even the strongest agent cannot secure a full score across all tasks. Mean effective scores range from roughly 40% to 88% (with per-configuration standard errors of 6-12 points), repeating every task three times reveals that top configurations still pass only a fraction of their attempts, and per-check scoring exposes a general failure pattern: agents may routinely satisfy short-term objectives while leaving non-durable changes, broken distributed invariants, unsafe side effects, and uncleaned state behind. INFRABENCH, including its live leaderboard, tasks, and evaluation harness, is publicly available at infraben.ch.

LLM agents are judged by what they deliver, and in professional work what gets delivered is a project, not a task. A project pays off only if the whole package holds, and one weak part can limit the value of the rest. Yet existing benchmarks mostly test agents on one task at a time, leaving a gap between the capability we measure and the outcome that makes agents useful. To address this gap, we introduce DeliverBench, which tests whether agents can carry a project from brief to final delivery. Its 11 projects, spanning financial due diligence, legal contract drafting, and drug safety reporting, are reconstructed with 13 domain experts from real materials and chain 150 subtasks through 2-15 rounds of new materials and client feedback. Because no reference outputs are injected, agents inherit their own errors from one task to the next. Expert rubrics then score every checkpoint as well as the final package. Across seven model-harness systems, even the best system scores only 65.0 out of 100 on final delivery and passes 50.0% of subtasks at a 90-point threshold. The failures are often failures of carrying work forward: in one audit project, all seven systems addressed the review findings, but only one preserved traceability. Together, these results suggest that task-level capability does not simply compose into delivery, which also depends on keeping the work coherent as it evolves.

Measuring complexity from a model's own code is biased by failure: a hard prompt can produce a short, broken program that looks simple. We instead score each prompt's structure on six dimensions from the prompt alone, never from generated code or test results. We draw 5,000 Python prompts across six bands of a preliminary single-rater score, and four out-of-panel LLM judges rescore every prompt (composite ICC(2,1) = 0.872). Twenty-one models answer every prompt (105,000 generations), and an independent LLM auditor, checked against known-answer cases, reviews each unit-test verdict. Averaged over models, pass rate has a nonmonotone breakpoint at composite 14.0: 83.4% at or below, 90.1% above. That jump is not a law of difficulty: with task-type fixed effects the breakpoint stays at 14.0 but the adjusted jump shrinks from 6.8 to 0.4 points, and neither of the benchmark's two prompt sources reproduces it on its own. Nineteen models pass more often above their own breakpoint and two less often; one of those declines comes from API calls that returned nothing. In post hoc follow-up, Kimi K2.5 shows the clearest decline, at 14.25, under a prompt-source control and within the later source alone. Among failed generations with measurable output complexity, 25.0% pair a prompt composite above 8 with output complexity of at most 10, twice the share among passing ones. The pooled breakpoint moves to 11.25 on raw test outcomes. Human agreement with the index is moderate, paraphrase and cross-language rescoring preserve its ordering, and overidentification tests reject the six dimensions as instruments, so we make no causal claim. The contribution is a prompt-side way to locate where each model's pass rate changes and to test how much that depends on the prompt mix and on how success is scored.

Large language models are increasingly deployed as autonomous agents executing multi-step workflows in real-world software environments. However, existing agent benchmarks suffer from three critical limitations: (1) trajectory-opaque grading that checks only final outputs, (2) underspecified safety and robustness evaluation, and (3) narrow modality coverage and interaction paradigms. We introduce Claw-Eval, an end-to-end evaluation suite addressing all three gaps. It comprises 300 human-verified tasks spanning 9 categories across three groups (general service orchestration, multimodal perception and generation, and multi-turn professional dialogue). Every agent action is recorded through three independent evidence channels (execution traces, audit logs, and environment snapshots), enabling trajectory-aware grading over 2,159 fine-grained rubric items. The scoring protocol evaluates Completion, Safety, and Robustness, reporting Average Score, Pass@k, and Pass^k across three trials to distinguish genuine capability from lucky outcomes. Experiments on 14 frontier models reveal that: (1) trajectory-opaque evaluation is systematically unreliable, missing 44% of safety violations and 13% of robustness failures that our hybrid pipeline catches; (2) controlled error injection primarily degrades consistency rather than peak capability, with Pass^3 dropping up to 24% while Pass@3 remains stable; (3) multimodal performance varies sharply, with most models performing poorer on video than on document or image, and no single model dominating across all modalities. Beyond benchmarking, Claw-Eval highlights actionable directions for agent development, shedding light on what it takes to build agents that are not only capable but reliably deployable.

The inherent rigidity of fixed-size benchmarks makes them an inefficient tool for model evaluation. Diverse evaluation objectives, including model ranking, model selection and testing throughout development, demand varying levels of statistical power. The mismatch between fixed sample sizes and these diverse needs results in either excessive computational cost or compromised reliability - a critical concern for model evaluation. To overcome these limitations, we call for adoption of sequential testing in our field. We provide an adaptive evaluation framework, that provides a principled way to navigate the trade-off between efficiency and reliability in model evaluation. Our framework combines the established statistical paradigm of sequential testing with stopping criteria tailored to common evaluation needs such as diminishing returns detection, and minimum detectable effect size. We demonstrate its ability to adaptively manage the efficiency-reliability trade-off on the Open VLM Leaderboard, including, for example, a 80% reduction in computational cost compared to fixed-size evaluation (with a 2.5-point CI width allowance) while maintaining statistical significance.

AI evaluation suites often report many scores without checking whether those scores carry independent information. We introduce Effective Dimensionality (ED), the participation ratio of a centered benchmark-score spectrum, as a fast, population-conditional upper-bound diagnostic of measurement breadth. Applied at per-instance granularity to 22 benchmarks across 8 domains and more than 8,400 model evaluations, ED reveals substantial redundancy: the six-score Open LLM Leaderboard behaves like roughly two effective measurement axes (ED = 1.7), BBH and MMLU-Pro are near-interchangeable (rho = 0.96, stable across seven subpopulations), and measurement breadth varies more than 20x across current benchmarks. We show that relative ED rankings are stable under matched-dimension controls and that ED can flag redundant suite components, monitor performance-conditional compression, and guide benchmark maintenance. Because binary spectra overestimate absolute latent dimensionality, we interpret ED as a screening statistic rather than a literal factor count and complement it with null, reliability, and saturation analyses. We provide a 22-benchmark reference atlas and a four-step diagnostic workflow that benchmark maintainers can run with a score matrix and a few lines of code.

Clinical decision-support systems are expert systems whose recommendations clinicians act on directly, yet they are usually cleared on one aggregate accuracy number from a held-out test set. That number says nothing about input reliability under encoding shifts, subgroup gaps, threshold sensitivity, or operational feasibility. We present RISED, a pre-deployment evaluation framework operationalising five dimensions (Reliability, Inclusivity, Sensitivity, Equity, Deployability) through BCa bootstrap 95% confidence intervals, literature-grounded thresholds, and Holm-Bonferroni-corrected PASS / FAIL / INCONCLUSIVE verdicts; Equity is a proxy-dependence diagnostic rather than a gating test. Applied to seven cohorts spanning 35 years (n from 303 to 99,492), RISED surfaces failures invisible to AUROC: on Diabetes 130, Reliability passes by three orders of magnitude (PSS = 0.0004) while Inclusivity (AUC parity gap = 0.262) and Sensitivity (max threshold-flip rate 49.1%) fail decisively; both NHIS cohorts reproduce this. NHANES 2021-2023, with a complete feature profile, achieves INCONCLUSIVE verdicts; BRFSS 2024 produces the suite's most severe Sensitivity failure (max threshold-flip rate 64.2%) after instrument rotation removed hypertension and cholesterol. The pattern recurs on credit- and income-prediction cohorts, confirming domain-agnosticity; a multi-model check shows the failures are data-driven, not model-specific. RISED ships as an open-source Python package complementing TRIPOD+AI, FUTURE-AI, and Fairlearn with the structured numerical evidence those standards require but do not prescribe.

Sequential agent search accumulates changes from its current champion but discards alternative branches; independent proposals preserve breadth but restart from the root. Loreley instead retains complete repository states in a Quality-Diversity (QD) archive and samples them as parents or supplies them as context for later edits. Candidates are Git commits produced in isolated worktrees and judged by a project-supplied evaluator. We compare configured Loreley QD, sequential champion editing, and independent root proposals in a matched Zstandard experiment: seven paired blocks and 48 physical candidate jobs per policy and block (1,008 total), with root-only initialization and each policy's native concurrency. Validation selected a winner at each budget checkpoint; an agent-hidden holdout measured the fixed candidate. At 48 jobs, QD was 0.135% below Sequential Champion (95% BCa interval for the paired effect: -0.556% to +0.161%) and 0.320% above Independent Root (-0.082% to +0.686%). Neither contrast established a QD advantage; Sequential had the highest observed 48-job mean and median. Archive retention and later sampling did occur. Four of seven final QD winners had a non-incumbent state in their primary-parent ancestry under a retrospective one-incumbent rule applied only to the observed QD stream. Including inspiration edges raised the count to six, without showing that supplied context caused an edit. Three earlier capability campaigns produced generation-4, multi-file improvements in two Python libraries and a separate Zstandard revision. Loreley engaged the intended stepping-stone mechanism, but the controlled experiment did not show an endpoint benefit at 48 jobs.

Tool-using LLM agents can violate the very policies they are deployed to enforce while appearing to complete the task successfully. In policy-permissive environments, a tool may execute any well-formed call even when the corresponding state transition is forbidden by domain policy. The result is a silent wrong state (a booking cancelled, a passenger count changed, a claim acted on without verification) that neither the tool nor the agent's self-report exposes. We study this failure mode in the τ^2-bench airline domain. On a budget agent, 78% of observed failures are silent wrong-state failures with no tool error, and the aggregate failure rate is reproducible across disjoint seeds, not sampling noise. We then evaluate a lightweight intervention: deterministic, read-only pre-execution gates that inspect the proposed call and current state before allowing a write. A four-gate suite raises full-benchmark success from 29.6% to 42.0% on gpt-4o-mini (+12.4pp; paired task-level bootstrap P=0.0012), and the lift reproduces on a disjoint 15-seed set (+12.3pp; P=0.0008). The effect is concentrated where the gates fire: on the 26/50 firing tasks, success rises by +19.2pp, while movement on the 24 non-firing tasks does not exclude zero. Two negative controls (a self-enforcing retail domain and BFCL) bound the mechanism: gates help when tools are policy-permissive and add little where tools already self-enforce. As suggestive evidence, not a central claim, the same failure mode persists at the frontier: gpt-5.2 at default reasoning still attempts policy-violating writes, and the same suite improves success from 61.2% to 71.6% (+10.4pp; P=0.020; n=5, no replication). The contribution is a bounded evaluation and reliability result: deterministic gates do not guarantee task success, but they can deterministically prevent a known class of silent policy-violating writes at the action boundary.

Enterprise agent programs are moving from prototypes into production, where reusable skills, tools, and workflow packages must be reviewed with evidence rather than prose. Current gates often scan these artifacts for structure, style, and security, but they do not answer the deployment question: does the capability package help a live agent complete enterprise tasks under the same model, sandbox, and grading policy? We present ACES (Agentic Continuous Evaluation of Skills), a repository-native framework for evaluating skills and product capability packages as executable agent artifacts. ACES runs paired live trials with and without a target skill, normalizes trajectories into the Agent Trajectory Interchange Format (ATIF), grades six default runtime metrics, and reports Skill Lift: the target skill's added value for a fixed task, harness, workspace, and scorer. The same protocol supports product-owned task suites that compare baseline, skill, bundle, team-skill, and plugin targets. On 145 real skills from internal enterprise repositories and public catalogs, scan-only gates surface useful authoring issues but measure complementary facets (structural versus LLM-judge Spearman ρ= 0.14). Across 947 scored paired cases from 58 of 64 production skills and four primary harnesses, mean composite Skill Lift is 0.2134 (95\% paired-case CI [0.1967, 0.2301]); mean outcome-only lift, the average of accuracy and goal accuracy, is 0.1799. Composite lift is positive in 72.8\% of paired cases. The largest process-metric gains appear in skill execution, behavior check, and skill efficiency---signals about discovery, routing, workflow following, and tool use that document scans cannot observe. An open-source implementation of the methodology is available in NVIDIA SkillEvaluator.

As long-horizon coding agents produce more code than any developer can review, oversight collapses onto a single surface: the automated test suite. Reward hacking naturally arises in this setup, as the agent optimizes for passing tests while deviating from the users true goal. We study this reward hacking phenomenon by decompose software engineering tasks into three parts: (i) a natural language description of the specification (ii) visible validation tests that exercise specified features in isolation, and (iii) held-out tests that compose those same features to simulate real-world usage. Based on the specification and the visible validation test suites, a genuine agent would be able to generate a solution that can also pass all of the held-out tests. Therefore we use the gap in pass rates on these two suites to quantify reward hacking. Based on this methodology, we introduce SpecBench, a benchmark comprising 30 systems-level programming tasks ranging from short horizon tasks like building a JSON parser to ultra long horizon tasks like building an entire OS kernel from scratch. Large-scale experiments reveal a consistent pattern: while every frontier agent saturates the visible suite, reward hacking persists, with smaller models exhibiting larger gaps on holdout suites. The gap also scales sharply with task length: it grows by 28 percentage points for every tenfold increase in code size. Failures range from subtle feature isolation to deliberate exploits, including a 2,900-line hash-table "compiler" that memorizes test inputs. SpecBench offers a principled testbed for measuring whether coding agents build genuine working systems or merely game the test suites developers hand them.

For deploying foundation models, practitioners increasingly need prescriptive scaling laws: given a pre training compute budget, what downstream accuracy is attainable with contemporary post training practice, and how stable is that mapping as the field evolves? Using large scale observational evaluations with 5k observational and 2k newly sampled data on model performance, we estimate capability boundaries, high conditional quantiles of benchmark scores as a function of log pre training FLOPs, via smoothed quantile regression with a monotone, saturating sigmoid parameterization. We validate the temporal reliability by fitting on earlier model generations and evaluating on later releases. Across various tasks, the estimated boundaries are mostly stable, with the exception of math reasoning that exhibits a consistently advancing boundary over time. We then extend our approach to analyze task dependent saturation and to probe contamination related shifts on math reasoning tasks. Finally, we introduce an efficient algorithm that recovers near full data frontiers using roughly 20% of evaluation budget. Together, our work releases the Proteus 2k, the latest model performance evaluation dataset, and introduces a practical methodology for translating compute budgets into reliable performance expectations and for monitoring when capability boundaries shift across time.

Rufus-Air is an open and reproducible post-training recipe on GLM-4.5-Air-Base (106B-A12B), organized as a serial pipeline of eight stages: SFT, Reasoning RL, Coding RL, Instruction-Following RL, General Agent, Coding Agent, Search Agent, and RLHF. We document the data, reward design, infrastructure, stage order, and stagewise results needed to reproduce the recipe. Stages progress from basic to advanced capabilities and from hard, verifiable rewards to softer judge-based signals. Training builds on open-source components and public data, much of it used as released, without new human annotation or an in-house distillation teacher. Our main findings are that (i) diverse, high-quality SFT establishes a strong capability floor; (ii) difficulty filtering keeps RL prompts within a productive learning range; (iii) reward reliability provides a practical principle for ordering stages; and (iv) infrastructure and engineering choices are part of the recipe, not just an implementation detail. Rufus-Air improves over the official GLM-4.5-Air post-trained release and is competitive with similarly sized open models.

The integration of AI-powered coding agents into Continuous Integration/Continuous Delivery (CI/CD) pipelines has fundamentally altered how software verification is conducted. While these agents successfully automate the test generation, current evaluation benchmarks (e.g., SWE-bench) largely focus on pass-rates rather than the intrinsic quality of the generated tests. This raises the possibility of "stealth technical debt", in which test suites pass execution but do not offer comprehensive coverage or semantic value. We address this methodological gap through a large-scale, empirical comparison of 204,673 test artifacts which comprises of 24,941 human-authored files and 179,732 agent-generated files; sourced from the AIDev dataset. Using the Abstract Syntax Tree (AST) parsing with Python's naive ast module, we implemented a "white-box" static analysis framework to evaluate three quality dimensions: Assertion Strength (RQ1), Edge-Case Coverage (RQ2), and Flakiness Potential (RQ3). Our results present a nuanced inversion of traditional assumptions. AI agents performed better than humans in Edge-Case Coverage, with almost twice the variety of boundary checks (Variety Score: 0.62 vs 0.32) and a higher frequency of null-safety testing (13.40% vs. 8.3%), even though human developers had a slight advantage in Assertion Strength (88.1% strong assertions vs. 85.37% for agents). But this thoroughness comes at a price: due mostly to their reliance on file I/O and non-deterministic logic, agent-generated tests exhibited a higher risk of flakiness (Candidate Rate: 0.41 vs. 0.30). These findings suggest that while AI agents excel at rigorous boundary testing, they lack the "environmental awareness" needed to write stable, hermetic tests.

Per-field accept/review with selective risk at most alpha -- accept a field only if the error rate among accepted fields is controlled -- is the trust contract document-extraction systems need, and the natural procedure silently violates it on real documents. On 13,859 genuine claude-sonnet-5 fields from 800 CORD receipts (49.0% correct) we diagnose three failure modes: document clustering (design effect 1.84-2.45), score-refit leakage (coverage 0.416 at risk 0.127, violating alpha=0.10 in 95% of splits), and a tie-mass pathology (a degenerate score collapses the threshold grid, 0.030 to 0.001). We organize the fixes as a validity ladder, guarantee form stated per tier. A fit/val split protocol restores expected-selective-risk control for a learned fusion: coverage 0.318 at risk 0.096 at nominal alpha=0.10, no tolerance band (production variant 0.326) -- an on-average point whose realized risk exceeds alpha in 47.5% of resplits, not a certificate. Mondrian Learn-then-Test with exact binomial tails yields per-group PAC certificates: field-iid 0.171 at risk 0.068, cluster-corrected 0.140, doc-iid 0.060 -- the only tier matching documents, honestly near-vacuous today. Support-bin, the pre-specified provenance taxonomy, wins every rigor tier on the sonnet CORD capture (p<1e-4, Bonferroni-corrected) -- a win that does not replicate on the same documents under haiku or qwen -- while on higher-accuracy corpora pooled thresholds win: conditioning helps exactly where pooled cannot certify, subsumed by a learned score elsewhere. A frozen-configuration confirmation on selection-untouched claude-haiku-4-5 held at both risk levels, and a blind three-annotator human-gold audit verifies the practical tier's accepted-set risk at 1.3% against its 10% budget (Fleiss' kappa=0.83; labels err one-sidedly pessimistic). Released Apache-2.0 with seed-pinned, regression-gated procedures.

Benchmarks shape scientific conclusions about model capabilities and steer model development. This creates a feedback loop: stronger benchmarks drive better models, and better models demand more discriminative benchmarks. Ensuring benchmark reliability is therefore essential for trustworthy evaluation and meaningful progress. In this work, we study benchmark reliability from a distributional perspective and introduce benchmark harmony, which measures how uniformly a model's performance is distributed across the subdomains of a benchmark. We posit that high harmony is a desirable benchmark property, indicating that the aggregate metric reflects uniform competence across subdomains. Across 19 multiple-choice benchmarks and five model families, we map each benchmark onto a mean-variance plane of harmony computed across models, where high mean and low variance signal more reliable evaluation. Our analysis shows that less harmonious benchmarks can give misleading results, since overall accuracy may be disproportionately influenced by specific subdomains. For instance, ARC-Easy is overwhelmed by questions on Biological Concepts, overshadowing other critical subdomains such as Geography, Physics, Chemistry, and Environmental Science. By recommending that harmony should be reported alongside accuracy, we reframe evaluation from simple performance averages to a more robust, distributionally reliable measurement of performance.

Agent benchmarks are increasingly used to compare large language models (LLMs) and guide deployment decisions, yet benchmark scores are meaningful only if they measure model capability rather than properties of the evaluation pipeline. We identify a double measurement confound: execution-critical decisions are performed by a fixed scaffold instead of the model, while the scorer evaluates outputs using criteria that may not reflect task correctness. We unify these issues within a measurement-theoretic framework that characterizes when benchmark scores can be interpreted as evidence of model capability, and instantiate it with an audit-and-repair protocol that (i) transfers execution-critical decisions from the scaffold to the model, (ii) replaces shape-based evaluation with seeded ground-truth scoring, and (iii) reports reliability beyond the mean through worst-case and tail-risk metrics. Experiments on ComtradeBench show that the joint intervention transforms a nearly flat leaderboard into a reliability spectrum that distinguishes both average performance and robustness across seeds. Applying the audit to existing benchmarks further shows that scorer validity is benchmark-specific, whereas scaffold ownership is an uncontrolled axis wherever we probed it. Our results suggest that benchmark scores should be interpreted together with their scaffolding level, scoring criterion, and reliability profile, providing a practical framework for more valid evaluation of LLM agents.

LLM-judged benchmarks are increasingly used to evaluate complex model behaviors, yet their design introduces failure modes absent in conventional ground-truth based benchmarks. We argue that without tight objectives and verifiable constructions, benchmark rankings can produce high-confidence rankings that are in fact largely noise. We introduce two mechanisms to diagnose these issues. Schematic adherence quantifies how much of a judge's overall verdict is explained by the explicit evaluation schema, revealing unexplained variance when judges deviate from their own rubric. Psychometric validity aggregates internal consistency and discriminant validity signals to quantify irreducible uncertainty in any benchmarking run. Applying these tools to Arena-Hard Auto, we find severe schema incoherence and factor collapse across popular judges: for example, unexplained variance exceeding 90 percent for DeepSeek-R1-32B and factor correlations above 0.93 for most criteria. We also show that the ELO-style aggregation used by Arena-Hard Auto collapses and masks genuine ranking uncertainty. Our results highlight design failures that undermine validity and offer actionable principles for building better-scoped, reliability-aware LLM-judged benchmarks. We release our code at https://anonymous.4open.science/r/judgment-to-noise-947D/README.md

Autonomous research agents produce competitive solutions and professional-looking manuscripts, yet their outputs contain verifiability failures undetectable by surface-level evaluation: fabricated citations, unreproducible scores, and method descriptions that diverge from the implementation. We address this through three contributions. First, Chain-of-Evidence (CoE), a verifiability framework requiring every claim to be traceable to its evidence source. Second, ScientistOne, an end-to-end autonomous research system that maintains evidence chains by construction throughout literature review, solution discovery, and paper writing. Third, CoE Audit, a post-hoc audit whose four integrity checks -- score verification, specification violation, reference verification, and method-code alignment -- apply uniformly to all systems. Across 75 papers spanning five systems and five frontier research tasks, every baseline exhibits at least one systematic failure mode: hallucinated reference rates reach 21%, score verification passes in as few as 42% of papers, and method-code alignment ranges from 20% to 80%. ScientistOne achieves zero hallucinated references (0/337), perfect score verification (12/12), and the highest method-code alignment (14/15), while matching or exceeding human expert performance on all five tasks. ScientistOne further generalizes to six additional tasks spanning medical imaging, fine-grained recognition, 3D perception, and language modeling, achieving state-of-the-art on Parameter Golf and gold medals on MLE-Bench tasks where baselines fail entirely.

Large language models (LLMs) are entering legal workflows, yet we lack a jurisdiction-specific framework to assess their baseline competence therein. We use India's public legal examinations as a transparent proxy. Our multi-year benchmark assembles objective screens from top national and state exams and evaluates open and frontier LLMs under real-world exam conditions. To probe beyond multiple-choice questions, we also include a lawyer-graded, paired-blinded study of long-form answers from the Supreme Court's Advocate-on-Record exam. This is, to our knowledge, the first exam-grounded, India-specific yardstick for LLM court-readiness released with datasets and protocols. Our work shows that while frontier systems consistently clear historical cutoffs and often match or exceed recent top-scorer bands on objective exams, none surpasses the human topper on long-form reasoning. Grader notes converge on three reliability failure modes: procedural or format compliance, authority or citation discipline, and forum-appropriate voice and structure. These findings delineate where LLMs can assist (checks, cross-statute consistency, statute and precedent lookups) and where human leadership remains essential: forum-specific drafting and filing, procedural and relief strategy, reconciling authorities and exceptions, and ethical, accountable judgment.

Agent benchmarks typically report only final outcomes: pass or fail. This threatens evaluation credibility in three ways. First, scores may be inflated or deflated by shortcuts and benchmark artifacts, misrepresenting capability. Second, benchmark performance may fail to predict real-world utility due to scaffold limitations and recurring failure modes. Finally, capability scores may conceal dangerous or catastrophic actions taken by the agent. We argue that log analysis -- the systematic tracking and analysis of the inputs, execution, and outputs of an AI agent -- is necessary to overcome these validity threats and promote credible agent evaluation. In this paper, we (1) present a taxonomy of threats to credible evaluation documented through log analysis, and (2) develop a set of guiding principles for log analysis. We illustrate these principles on tau-Bench Airline, revealing that pass^5 performance was under-elicited by nearly 50% and surfacing deployment failure modes invisible to outcome metrics. We conclude with pragmatic recommendations to increase uptake of log analysis, directed at diverse stakeholders including benchmark creators, model developers, independent evaluators, and deployers.

Coding-agent benchmarks are built for broad and comprehensive comparisons, not frequent development decisions. Individual runs vary, full suites are expensive, and the benchmark harness may differ from the harness used in practice. In a resampling analysis of DeepSWE's published trials, only 19.5% of tasks (22 of 113) had a fifth-percentile Pearson correlation of at least 0.50 with full-benchmark performance. The paper presents DeltaSelect, an open-source method that identifies tasks whose one-run results consistently track full-benchmark performance using Pearson correlation, maps fractional verifier results to a common score using linear regression, and selects a fixed task set within a dollar budget. DeltaSelect is intended for repeated baseline-versus-candidate comparisons during development, not model rankings. In a gpt-5.6-luna low-reasoning case study, DeltaSelect was used to revise custom skills and instructions. Across 13 evaluations, the recorded cost was USD 27.86 at rates published August 16, 2026. The adopted version cost 58.1% less than the initial version (USD 1.75 versus USD 4.18; p=0.008), while the calibrated score was higher (42.36% versus 36.46%; published-analog variance p=0.326).

Computer-use agents pay full frontier inference to re-derive routines their user has already performed, because an agent's memory today records what the user said, not what the user did. We compile passively captured screen activity into agent memory with a deterministic, zero-model pipeline: it segments a local capture stream into typed activity frames, bounded episodes carrying application, site, timing, input volume, and evidence pointers back to the raw rows, with no model in the loop, so the output is byte-identical, cacheable, and mechanically auditable. On one professional's single-user corpus of 128,756 frames over 51 active days, the compiler reduces a day of raw capture to a prompt-ready context block 86x smaller in 68 ms, and an agent reading that block answers questions about the day at 98.4% accuracy (Wilson 95% CI 91.7-99.7%) against an independent oracle, versus 66-80% for an LLM summary of the same capture, a mid-tier model reading the block matching a frontier one. The same compiler doubles as a demand-side cost instrument. Read off passive, pre-delegation human activity rather than agent rollouts, it supplies two parameters that agent-cost models assume but, to our knowledge, have not measured: the Routine Overhead Ratio R and the routine recurrence h. We report first values of R, a modeled upper bound, at 60-343x, and a delegable recurrence of 9.0% in-sample and 7.7% out-of-sample, for a realistic all-fleet token ceiling near 8%; a compiled routine replays deterministically with the model out of the loop, demonstrated live at zero model tokens on a guard-matched hit. Schema, compiler, and evaluation harness are open.

Outcome RL raises pass@1 but not pass@k at large k (Yue et al., 2025). But every pass@1 is scored under a token budget. Every recipe also trains under one. On physics and math problems at fixed token budgets, we compare the pass@1 of 20 released RL models (1.5B to 32B) with that of the thinking models they started from. We ask whether each pass@1 gain is changed behaviour or new knowledge. We divide each change in pass@1 into three parts. Finishing measures answering before the budget runs out. Reliability measures being right once finished. Reach measures problems RL solves that its base answered wrongly in all eight tries, minus problems the base solves that RL answered wrongly in all eight. We find four things. (1) Where the base often hits the token budget before answering, most of the change, gain or loss, is finishing (82% at 4k and 8k). Given more tokens, the base closes most of the finishing gap itself. (2) Where the base has enough tokens, the gain is reliability: the RL model is graded right more often, even on problems both finish. (3) By our definition, RL rarely solves what its base cannot. Reach adds at most 2.1 percentage points on any of our four physics and math sets. We checked 144 physics problems that only the RL model solved. Of these, 3.5% hold up when resampled or read. In a random sample, nine of ten new RL steps a human confirmed also appear in the base's samples. (4) In a small-scale RL run on an 8B model, the gain on new problems is finishing, with reliability flat. Supplying each problem's physical setup was not shown to improve on plain RL.

The economic value of inference depends on how capacity and task information are distributed across stages of AI production. We study these organizational margins using controlled workflow experiments on externally verified software-engineering tasks. In two matched resource panels, direct execution records the same success rate of 59.6 percent at logical-token ceilings of 12,000 and 24,000, while success under information-constrained planning rises from 36.2 to 51.2 percent. The planning disadvantage narrows by 15.0 percentage points (95 percent task-cluster bootstrap interval: 4.2 to 25.8). A strict read-only planning campaign varies whether the planner sees the task issue. At 12,000 tokens, issue access raises success by about 16 percentage points over issue-hidden planning. Compared with direct execution, task-informed planning is about 10 points lower at 12,000 tokens; at 24,000 tokens, it shows a 29.6-point advantage. In the resource panels, direct execution uses substantially less than either ceiling, while the planning workflow's binding rate falls from 46.2 to 0.8 percent and downstream execution accounts for 89.9 percent of the increase in total use. Scale determines the capacity available to a system; workflow and information structure shape the productive value

Tool-using AI agents are increasingly deployed across enterprise software systems, yet widely used benchmarks primarily evaluate nominal task completion, conflating baseline planning competence with operational fault recovery. We introduce UndoBench, a benchmark spanning 36 base workflows and 36 fault scenarios across 8 enterprise domains, decoupling task competence from recovery capability via counterfactual paired trials under identical seeds alongside wire-level effect-history and environment-state oracles. On 12 held-out TEST workflows across two open-weight models, two frameworks, and three recovery paradigms (5,760 executions / 2,880 paired trials) in the frozen lost-acknowledgment study, nominal competence reached 83.54% while conditional recovery success rate (CRSR) fell to 46.72%, with naive retry producing duplicate external effects in 53.33% of trials. Extensions to commercial API models reproduced this competence-recovery separation. Evaluations across complementary execution boundaries show that recovery is phase-dependent: before mutation, methods perform similarly without duplicate effects among capable trials; during partial mutation, naive retry, per-call idempotency, and zero-privilege journaling collapse on the evaluated composite workflows; after commit but before acknowledgment, verification and server-side idempotency substantially improve safety. These findings demonstrate that evaluating nominal completion alone masks critical, phase-dependent recovery vulnerabilities in autonomous agents.

We develop and counterfactually evaluate a resolution-aware risk-design framework (PIRAP) for perpetual futures whose underlying tracks a single binary prediction-market probability through resolution. The framework specifies six components: an index estimator combining mid-price, depth-weighted mid, and time-decayed VWAP; jump-aware tiered margin sized against bounded-event terminal-collapse magnitude; leverage compression schedule contracting toward resolution; resolution-aware funding rule with boundary-aware correction; a multi-stage halt protocol; and an eligibility framework. Two formal non-portability propositions establish that standard basis-only funding paired with continuous-vol static margin fails on bounded-event underlyings. Empirical evaluation uses Polymarket's PMXT v2 archive for 2026-04-21 to 2026-04-27 (13,298-market analysis sample passing adequacy gates from 61,087 ingested; 13,115 resolved within the empirical window for E3). E1 evaluates two pre-registered stylized facts; E2 conducts counterfactual replay across three engine configurations; E3 isolates the resolution-zone protocol's contribution. Results are mixed. Five pre-registered floors: stylized-fact floors (boundary depth asymmetry, terminal-jump magnitude) PASS; welfare-side directional floors (final-hour liquidation -6%, drawdown -5.1% pooled, median PnL +14%) two FAIL one PASS; E3 mechanic floors (final-hour liquidation -80% by halt construction PASS; bad-debt frequency +2.4% FAIL). Three of five materiality floors fail: the framework as specified does not validate deployment, but the empirical record establishes a halt-versus-margin scope distinction (halt addresses execution-channel risk; terminal-jump bad-debt remains margin-side) and documents a pre-emption trade-off constraining the dynamic-margin component. The paper concludes with structural recommendations and explicit non-deployable status.

LLM agents are rapidly becoming production software, deployed to handle customer service, adjudicate disputes, and operate internal systems. Notably, the work of building them is increasingly handed to coding agents, yet existing benchmarks say little about whether an AI system can deliver one under the conditions of a real client engagement. We introduce τ^τ-bench (pronounced hyper-tau-bench), a benchmark that makes agent construction the task. A developer agent is given the records a business actually keeps, a client who holds requirements, a production API that operations must run through, a codebase to inherit, and limits on serving cost and models: the same starting point a real engagement provides. From these it must deliver a complete customer-service agent, scored by deploying that agent against held-out simulated users. Across 53 tasks spanning four domains, the strongest configuration, Claude Opus 5 under Claude Code, passes just 23.9% of evaluation simulations. Meanwhile, an expert-authored reference ceiling scores 82.2%. The failures mirror ones human agent developers see: models issue shallow queries in place of deep comprehension of the records, communicate almost nothing to the client, and experiment too little with agent architecture and serving spend, shipping the first design that runs. We aim for τ^τ-bench to turn the work of cooperative agent building into a measurable target for coding agents.

AI tools are accelerating scientific publication while the systems that review it struggle to keep up, and independent verification of published research has become both harder and more important. As manual replication is slow and expensive, a growing line of work uses coding agents to automate parts of the process. Existing efforts are largely packaged as benchmarks with companion agents that only run inside the benchmark's own pipeline, and no general-purpose replication tool exists. We present VERITAS, a domain-agnostic replication framework built around CLI coding agents. Given a paper, a code repository, or both, VERITAS extracts the paper's claims, runs the methodology while resolving issues as they arise, and judges each claim against the evidence from experiment runs. The pipeline returns an importance-weighted Replication Score, a severity-rated log of every fix applied, and the patched codebase. We evaluate VERITAS on CORE-Bench and ReplicationBench, 65 papers spanning computer science, social science, medicine, and astrophysics. Against two strong Claude Code baselines on the same model and host environment, VERITAS achieves state-of-the-art performance and leads on every metric on both benchmarks.

Leaderboards rank long-horizon agents by their final outputs. Yet a higher score alone does not establish whether two systems are comparable or which stage accounts for the difference. Unequal evidence, inputs, or budgets can affect scores, and statistical corrections do not remove this mismatch. We introduce compositional controllability to address these questions. A comparison window covers one stage, several stages, or the whole agent. Our central result bounds the gap between observed and controlled score differences using only nuisance outside the window. This yields an admissibility test applied before scores are inspected. Inadmissible comparisons are refused. For admissible pairs, an ordering is certified only when the score gap exceeds the combined sampling and nuisance radii; otherwise, it remains undecided. These decisions give each system a rank interval. We introduce BioLitBench, a benchmark of 2,042 biomedical articles represented as structured claim graphs. Among seven published pipelines, a conventional statistical analysis declares a winner in 14 of 21 pairwise comparisons. Yet the top-ranked system alone received the target review's bibliography. To isolate pipeline performance, our test requires matched inputs and a fixed backbone model. It refuses 11 of the 21 comparisons, including every comparison involving the top-ranked system. Seven of the 14 conventional conclusions fall within these refused pairs. The same comparison windows support stage-level training. We train SCRIBE on Qwen3.8-27B using rewards measured at each stage's exit. Under matched evidence, SCRIBE achieves a certified rank interval of [1,2], with certified advantages over all evaluated published pipelines and the evaluated Claude and OpenAI agents. Under same pool, SCRIBE matches the strongest published retriever and is certified above three published pipelines.

Language model agents are post-trained and deployed through a harness, the prompts, tools, and control logic that present a task to the model and execute its outputs. Harnesses are often revised and judged by how an already trained model performs under them, which assumes that what the model learned under one harness carries over to the next. We ask what a policy learns from its harness and what survives when the harness changes. Under KL-regularized control, we show that the log odds of an optimal policy recover how strongly its harness favors each action over the others, so a trained policy carries the preferences of its harness. Deployed return then equals the ceiling of the harness, the best return it offers, minus the mismatch of the policy, the return lost to its own suboptimal decisions. Training changes only the mismatch, while changing the harness moves both terms. Comparing harnesses with a fixed model therefore mixes what each harness offers with how well that model uses it. In MiniWoB++, post-training makes a model's preferences specific to its training harness. Reversing the order of the observation fields leaves every optimal decision unchanged yet lowers a trained model's success from 85.0% to 19.2%. Conversely, every new harness that a fixed-model comparison rejects but that raises the ceiling outperforms the original agent after further training under it, and ceiling and mismatch measured beforehand predict these gains. For coding-agent harnesses on Terminal-Bench Lite, ceiling minus mismatch estimated with a stronger model as a stand-in for the optimum reproduces the accuracy ranking, while either term alone does not. Code is available at https://github.com/2020-qqtcg/HRP.

Existing LLM safety evaluations rely on binary attack-success rates and domain-agnostic taxonomies, leaving regulated Banking, Financial Services, and Insurance (BFSI) deployments exposed to failures elicited through legally or professionally plausible framing. We introduce RAHS (Risk-Adjusted Harm Score), a risk-sensitive metric jointly capturing disclosure severity, disclaimer mitigation, and inter-judge agreement, and FinRedTeamBench, a 989-prompt benchmark spanning seven BFSI risk areas and 34 sub-categories mapped to regulatory frameworks. Evaluation uses an ensemble of three heterogeneous LLM judges, validated against human experts, and an adaptive multi-turn red-teaming pipeline. On nine open-weight models, RAHS preserves separation under near-ceiling ASR, ranking is stable under hyperparameter sweeps, and multi-turn pressure drives not only more jailbreaks but more operationally severe disclosures, exposing failure modes that single-turn, domain-agnostic evaluations cannot reveal.

Long-horizon terminal agents may solve a task in one run yet fail across repeated runs. Stochastic errors can propagate through actions and context, creating a gap between capability and reliable execution. We present RiseHarness, a framework for iteratively improving agent harnesses with model weights fixed. It focuses on flipping tasks with both successful and failed runs under the same configuration, using successful runs as references for failure analysis. RiseHarness combines agent trajectories with internal harness traces from a runtime monitor to examine how model actions interact with command execution, feedback, context management, and completion decisions. A Harness Handbook links this evidence to harness behaviors and source code, helping identify recurring problems in mechanisms shared across tasks. RiseHarness then searches other harness frameworks for relevant implementations and adapts them as design references to the current model and execution protocol. Candidate changes undergo behavioral checks and repeated task evaluation; versions with measured gains seed the next iteration. This loop integrates suitable mechanisms from different frameworks into a single harness and uses new execution evidence to guide further changes. We evaluate task success and reliability across models, ablate harness execution evidence and external research, and test transfer to unseen benchmarks with the same model. RiseHarness treats a specific model as the unit of harness adaptation, studying how changes to shared interaction mechanisms can make existing capabilities more reliable across tasks.

In passage reranking, response ranking and multi-document question answering, LLMs can score several candidate documents or responses together in one prompt, each still receiving its own score. Such scorers are selected on ranking quality, but their scores determine a decision: what a score threshold retains, a reader answers, or which chosen/rejected pair enters preference training. Because the candidates share that prompt, reordering them changes their scores. The same query over the same candidates should still yield the same decision. However, equal ranking quality does not imply equal decisions: on passage reranking, five trained scorers within 0.010 nDCG@10 retain sets that overlap by only 0.66-0.84 when reordered. No prompt-time change we test resolves that dependence: the only one that improves ranking quality does not measurably improve decision stability. We introduce order-consistency SFT (OC-SFT), which attenuates it in the weights by penalizing disagreement between a candidate's scores across orderings. It holds ranking quality and leads every decision-stability measure among trained scorers on all three tasks. It is also more stable on 12 base models than order-averaged distillation, which trains on labels averaged across permutations. One OC-SFT permutation retains sets that overlap more than ten averaged off-the-shelf permutations. A comparison of such scorers should therefore report what a threshold retains and a reader answers, not ranking quality alone. Code is available at https://github.com/thomsonreuters/presentation-dependence.

Repository-level performance-optimization benchmarks such as GSO, SWE-Perf and SWE-fficiency evaluate coding agents by applying patches to real repositories and comparing runtime against unoptimized baselines and official reference patches. Their leaderboard scores are increasingly used as evidence of coding-agent progress, but those scores can conflate runtime instability, benchmark-specific scoring rules, and how many tasks are already solved by at least one public submission. We audit these issues across the three benchmarks. First, we replay the official reference patches for 740 code optimization tasks across four common types of Google Cloud machines. Most benchmark tasks can be replayed, but their reference patches satisfy the original benchmark validity rules in every cross-machine replay for only 39/102 GSO tasks, 11/140 SWE-Perf tasks, and 411/498 SWE-fficiency tasks; SWE-Perf is especially fragile because many reference patches produce close-to-zero runtime changes. Second, we show that public submission rankings depend strongly on the benchmark scoring rule. Among eight public submissions shared by GSO and SWE-fficiency, the official rankings disagree on 9 of 28 pairwise submission comparisons, and SWE-fficiency's leaderboard scoring rule assigns the worst ten tasks overly high score weights of 58.5%-82.8%. Third, looking across 10 public submissions for each task, we find that at least one submission matches or beats the reference patch on 85.3% (384/450) of replay-valid GSO and SWE-fficiency tasks, and beats the unoptimized base code on 99.8% (449/450). Our study complements leaderboard scores by identifying tasks with more reliable performance signals, quantifying per-task score contributions, and exposing the remaining performance gaps that are hidden by aggregate rankings.

Risk-budgeted throughput forecasting converts a point forecast of LEO satellite capacity into a conservative bound whose overestimation rate is held at a user-specified budget, so that admission control can act on it safely. We reproduce the state of the art, BG-CFQS, on the three Starlink traces its authors used and on an independent fourth dataset, and report four findings. First, its risk guarantee is reproducible but conditional on the calibration split: under a random split we recover every published quantity, including risk-pass on 3 of 3 datasets, but under a contiguous temporal split, which is what a deployed terminal faces, it holds on 1 of 3. Second, independent of any split, the method cannot serve a risk budget below 0.15, the low end of its candidate quantile set. On all four datasets the achieved rate is identical at budgets of 0.05, 0.10 and 0.15, overshooting by 1.4x to 3.7x at the tightest. Third, static conformal calibration cannot improve a ranking, only shift an operating point; we confirm this exactly on latency spike detection, where online calibration improves AUPRC by up to 51%. Fourth, group-conditional online recalibration splits into two mechanisms. The online part replicates: it tracks the budget within 1% from 0.05 to 0.35, where static calibration is 16 to 21% off. The group-conditional part does not: on three of four datasets, conditioning on any available covariate, including measured serving-satellite geometry, is worse than not conditioning, and our attempt to predict which datasets benefit failed. What is new here is empirical, not algorithmic: we claim the findings, not the methods.

A paraphrase-quality audit of MathCheck (ICLR 2025) detected 4 semantically incorrect paraphrases in 129 groups (3.1%); removing them drops GPT-4o from rank 2 to rank 4 and elevates Claude Haiku and DeepSeek V3 above it; these ranking changes are invisible to any single-model evaluation. Cross-model unanimity found these errors automatically (>= 3/4 models for MathCheck; >= 6/9 for our primary evaluation) for under $10; in our own dataset the same protocol found that 47% of auto-generated connective-variation paraphrases were semantically incorrect. That flaw compounds a deeper measurement gap: Claude Haiku 4.5 achieves 86% accuracy yet SCR=50%, meaning half its theorems are answered differently under semantically equivalent restatements, while aggregate accuracy across 9 models spans only 86-96% yet Semantic Consistency Rates (SCR) span 50-82% -- a 32-point gap invisible to standard benchmarks. Formally, for any target ranking over 9 frontier models there exists a weighting over paraphrase families that realizes it (No-Free-Benchmark corollary), because no model Pareto-dominates all families -- so benchmark designers who select families are implicitly choosing which model wins. FormInv supplies the audit protocol (replicated on external benchmarks at 100% recall), SCR and per-theorem Cochran's Q as primary invariance measures evaluated on 9 models across 366-811 items (on Lean4-verified theorems), and FormInvSelector for regime-aware model selection.

Agent benchmarks often evaluate only final answers even when agents run on stateful runtimes. We argue this under-specifies what is being evaluated: the proper unit is a declared model-plus-runtime configuration whose failures can occur in evidence acquisition, runtime routing, safety boundaries, or repeated execution. We present ClawProBench, a trace-aware benchmark for runtime-native agent evaluation instantiated on OpenClaw, a live agent runtime with workspace tools and native surfaces for browsing, memory, messaging, scheduling, skills, and subagents. ClawProBench defines two tracks: a 102-scenario full profile with live workspace and native-runtime routing tasks, and a frozen 68-scenario holdout with closed-world JSON output contracts for robust ranking. Trials are scored from execution traces via a safety-gated formula combining correctness, process quality, and efficiency, preserving failure evidence for audit. Our anonymous artifact includes benchmark definitions, scoring code, manifests and sanitized traces. We evaluate 68 configurations on the full profile and 37 on holdout. The top safety-gated average trace score is 0.7671. Native-runtime tasks underperform workspace-live tasks (0.5238 vs. 0.6415). On holdout, pass@k-any outperforms strict three-trial pass (0.6638 vs. 0.2890), while full-profile and holdout rankings show weak alignment (Spearman 0.1300). Rankings based purely on correctness differ substantially from process-aware, safety-gated and strict-pass views. Final-answer leaderboards may hide native-surface weaknesses, one-off successes and trace-local agent failure modes.

This paper provides an experimentally verified formal law for calculating the uplift that diversity of thought provides in Large Language Model (LLM) ensembles. From first principles, we derive an exact decomposition of LLM ensemble lift into rescue and damage masses, which yields a compact heuristic for calculating uplift. From this we extract the metrics which predict ensemble performance: an accuracy adjusted correctness correlation, ϕ_{adj}, together with the accuracy gap and collective accuracy of the pair. We test the law on 767,520 inferences from ten open-weight models over two graduate-level science benchmarks, together with a novel agentic cybersecurity benchmark in which each model conducts digital-forensics investigations by multi-turn tool use in a network-isolated sandbox (23,520 graded trials including abstentions); all votes are released openly. Calibrated once on SuperGPQA at a 40:60 vote split, the heuristic predicts lift on the calibration set with Spearman's ρ=0.84 and, with its coefficients frozen, transfers to two datasets never used in calibration (ρ=0.51 on GPQA Diamond and 0.84 on the forensic tasks), whilst the measured swap mass tracks realised lift with R^2ge 0.96 throughout. Raw ϕ has almost no predictive power (R^2le 0.09 throughout); the accuracy-adjusted ϕ_{adj} is markedly superior (R^2=0.67 on SuperGPQA), and the heuristic combining these metrics is the most stable pre pooling predictor across the three datasets.

Frontier model progress is often measured by academic benchmarks, which offer a limited view of performance in real-world professional contexts. Existing evaluations often fail to assess open-ended, economically consequential tasks in high-stakes domains like Legal and Finance, where practical returns are paramount. To address this, we introduce Professional Reasoning Bench (PRBench), a realistic, open-ended, and difficult benchmark of real-world problems in Finance and Law. We open-source its 1,100 expert-authored tasks and 19,356 expert-curated criteria, making it, to our knowledge, the largest public, rubric-based benchmark for both legal and finance domains. We recruit 182 qualified professionals, holding JDs, CFAs, or 6+ years of experience, who contributed tasks inspired by their actual workflows. This process yields significant diversity, with tasks spanning 114 countries and 47 US jurisdictions. Our expert-curated rubrics are validated through a rigorous quality pipeline, including independent expert validation. Subsequent evaluation of 20 leading models reveals substantial room for improvement, with top scores of only 0.39 (Finance) and 0.37 (Legal) on our Hard subsets. We further catalog associated economic impacts of the prompts and analyze performance using human-annotated rubric categories. Our analysis shows that models with similar overall scores can diverge significantly on specific capabilities. Common failure modes include inaccurate judgments, a lack of process transparency and incomplete reasoning, highlighting critical gaps in their reliability for professional adoption.

Leaderboards rank frontier models on independent axes but do not reveal whether capabilities reinforce or trade off across releases -- and at the frontier, this interaction is the more informative signal. We decompose paired SWE-bench and GPQA Diamond scores into a population coupling trend and per-release residual (h-field) that diagnoses capability emphasis and identifies which measurement or stress test is most informative next. Across 34 models from 10 labs (2024--2026), capabilities cooperate (r = +0.72, p < 10^{-6}), but cooperation varies by lab and over time: DeepSeek reversed from reasoning-rich to coding-first (h: +11.2 to -4.7, 15.9-pp swing); Google maintains consistent reasoning emphasis; Anthropic oscillates between coding excursions and recovery. Cooperation is not static -- it cascades. Six open-weight architectures confirm a second capability transition at 30--72B, and SWE-bench is now saturating while HLE and instruction-following retain discriminatory spread -- signaling the next axis rotation. We provide a three-level playbook (locate, diagnose, rotate), a per-lab measurement-priority table, and seven falsifiable predictions with timestamped criteria for the next 12 months of frontier releases. Per-lab coupling slopes vary 5times (Google 1.15 vs. DeepSeek 0.23), quantifying how efficiently each recipe converts coding gains into reasoning. Five April 2026 releases confirm the diagnostic out of sample (r rises from +0.72 to +0.75). An interactive dashboard provides phase classification with actionable recommendations, h-field diagnostics, per-lab coupling trajectories, ODE-based scaling predictions, benchmark rotation guidance, self-steering demo, and live tracking of all seven predictions: https://zehenlabs.com/cape/.

We study whether demonstration-curation metrics that detect defective training episodes also improve the downstream behavior-cloning policy that trains on the curated data. On a contact-rich LIBERO pick-and-place benchmark with a controlled structural defect (early gripper release during the carry phase), we find that the two quantities are sharply decoupled. The metric with the highest defect-detection AUROC (0.804) produces the worst curated policy (13.3% task success), while a metric with a substantially lower AUROC (0.638) produces a policy that nearly matches the oracle trained on ground-truth clean data (90.0% vs. 93.3%). We further show that five of the seven metrics we evaluate exploit episode length as a trivial proxy for the defect label, a confound that inflates reported AUROCs to near-perfect values and disappears once episode length is controlled. Across all conditions, the contaminated baseline succeeds on only 3.3% of rollouts, and the two best curation methods close this to within 3 percentage points of the 93.3% oracle ceiling. Our results argue that curation methods should be evaluated by the policy they produce, not the defects they flag, and that any curation benchmark must control for episode length before reporting detection accuracy. We release the testbed, all metric implementations, and the evaluation pipeline.

Corruption studies, the primary tool for evaluating chain-of-thought (CoT) faithfulness, identify which chain positions are "computationally important" by measuring accuracy when steps are replaced with errors. We identify a systematic confound: for chains with explicit terminal answer statements, the dominant format in standard benchmarks, corruption studies detect where the answer text appears, not where computation occurs. A within-dataset format ablation provides the key evidence: on standard GSM8K chains ending with "the answer is X," removing only the answer statement, preserving all reasoning, collapses suffix sensitivity ~19x at 3B (N=300, p=0.022). Conflicting-answer experiments quantify the causal mechanism: at 7B, CC accuracy drops to near-zero (<=0.02) across five architecture families; the followed-wrong rate spans 0.63-1.00 at 3B-7B and attenuates at larger scales (0.300 at Phi-4-14B, ~0.01 at 32B). A within-stable 7B replication (9.3x attenuation, N=76, p=7.8e-3; Qwen3-8B N=299, p=0.004) provides converging evidence, and the pattern replicates on MATH (DeepSeek-R1-7B: 10.9x suffix-survival recovery). On chains without answer suffixes the same protocol identifies the prefix as load-bearing (Delta=-0.77, p<10^-12). Generation-time probes confirm a dissociation: the answer is not early-determined during generation (early commitment <5%), yet at consumption time model outputs systematically follow the explicit answer text. The format-determination effect persists through 14B (8.5x ratio, p=0.001) and converges toward zero at 32B. We propose a three-prerequisite protocol (question-only control, format characterization, all-position sweep) as a minimum standard for corruption-based faithfulness studies.

Tool-augmented LLM agents call APIs whose intermediate outputs, such as presigned URLs, session tokens, and OAuth state parameters, are observation contracts: artifacts whose later use is constrained by the external system that produced them. We show that observation contract compliance (preserving the temporal validity and byte-level integrity) is an emergent, regression-prone capability: it is neither guaranteed by general tool-use ability nor consistently improved by larger or newer models. To measure this, we introduce ContractBench, a benchmark of 33 dual-axis tasks that probe two orthogonal failure modes no existing benchmark evaluates: validity failures (using an artifact after expiry) and integrity failures (corrupting an artifact's bytes through the observation-to-action pipeline). Our evaluation is deterministic and programmatic, with a virtual clock controlling time and SHA-256 hashes verifying byte integrity. We assign each outcome a failure label drawn from real-world API specifications. We evaluate 38 models and report four findings: (i) no evaluated model clears 80%, with Claude-Opus-4.6 leading at 77.8%, revealing that current frontier models still fail to comply with observation contracts; (ii) a sharp within-family capability cliff in Qwen 3.5 between 4B (0%) and 9B (56.6%), smoothing to 70.7% at 397B-A17B: what emerges across the cliff is mid-trajectory restraint, not tool-call competence; (iii) non-monotonic scaling across the GPT-5 family: agentic post-training can erode compliance through sycophancy-driven regression; (iv) our failure taxonomy works as an actionable in-context reward signal, yielding +7.1 pp on 42 paired GPT-5.1 failures.

Objective. Clinical AI documentation systems require evaluation methodologies that are clinically valid, economically viable, and sensitive to iterative changes. Methods requiring expert review per scoring instance are too slow and expensive for safe, iterative deployment. We present a case-specific, clinician-authored rubric methodology for clinical AI evaluation and examine whether LLM-generated rubrics can approximate clinician agreement. Materials and Methods. Twenty clinicians authored 1,646 rubrics for 823 clinical cases (736 real-world, 87 synthetic) across primary care, psychiatry, oncology, and behavioral health. Each rubric was validated by confirming that an LLM-based scoring agent consistently scored clinician-preferred outputs higher than rejected ones. Seven versions of an EHR-embedded AI agent for clinicians were evaluated across all cases. Results. Clinician-authored rubrics discriminated effectively between high- and low-quality outputs (median score gap: 82.9%) with high scoring stability (median range: 0.00%). Median scores improved from 84% to 95%. In later experiments, clinician-LLM ranking agreement (tau: 0.42-0.46) matched or exceeded clinician-clinician agreement (tau: 0.38-0.43), attributable to both ceiling compression and LLM rubric improvement. Discussion. This convergence supports incorporating LLM rubrics alongside clinician-authored ones. At roughly 1,000 times lower cost, LLM rubrics enable substantially greater evaluation coverage, while continued clinical authorship grounds evaluation in expert judgment. Ceiling compression poses a methodological challenge for future inter-rater agreement studies. Conclusion. Case-specific rubrics offer a path for clinical AI evaluation that preserves expert judgment while enabling automation at three orders lower cost. Clinician-authored rubrics establish the baseline against which LLM rubrics are validated.

We introduce DoGBENCH (Documentation Generation Benchmark), to our knowledge, the first benchmark for generating and maintaining real user-facing software documentation. It asks whether an agent can produce documentation that experienced technical writers would accept in review. The benchmark contains 292 items from open source projects, including Helm, PostHog, and Mautic. Each item gives the agent a pre-change repository and a trigger, such as a code pull request or a reported documentation gap. The agent must first decide whether the documentation needs an update. For items that need one, the agent must produce an acceptable patch in one attempt. For items that do not need updates, the agent must abstain. Task-specific rubrics, validated with project maintainers, score each patch on accuracy, completeness, reader guidance, placement, and repository conventions. The composite score combines patch quality with correct abstention, and a score of 100 means an agent meets every requirement for the task. Scores should not be interpreted as a percentage of an expert's capability. We evaluated seven agents. The highest-scoring agent reached 47.3 out of 100 on the 117-item held-out split. In a separate audit of 1,267 patches, the most common failure modes were task-completion gaps (45.5%), technical inaccuracies (36.6%), and incomplete conceptual or reference coverage (32.5%). Analysis of the corresponding trajectories identified three key patterns associated with these failures: (1) describing interfaces without examining how readers use them (36.0%), (2) missing decisive evidence and filling the gaps with plausible assumptions (33.1%), and (3) stopping after finding the first plausible documentation surface and leaving other affected pages stale (30.1%).

Benchmarks measure whether a model is correct. They do not measure whether a model is reliable. This distinction is largely academic for single-shot inference, but becomes critical for agentic AI systems, where a single rephrased prompt can trigger cascading failures in multi-step execution. Yet this form of instability is not captured by existing evaluations. Hallucinations live in variance: they arise when semantically equivalent prompts activate inconsistent internal pathways, producing divergent outputs. Consistent but incorrect outputs reflect bias or missing knowledge; confident guessing reflects calibration failure. Neither constitutes hallucination under this definition. When error is variance-dominated, reducing redundant pathways improves reliability without adding knowledge. We formalize this through Semantic Stability (SS), measured via Paraphrase Consistency (PC@k): generate k paraphrases, greedy decode each, compute mode agreement. SS is a diagnostic for variance-driven unreliability, not a method for improving correctness. We show that a dense Qwen3-0.6B agrees with itself only 23.8% of the time; at 32% sparsity, agreement jumps to 55.9%. A phase diagram reveals the sweet spot where variance reduction outpaces bias accumulation, and regimes where stability collapses onto wrong answers.

Competition retrospectives are useful when they explain what a leaderboard measured, how hidden evaluation changed conclusions, and which design patterns were rewarded. We revisit the CODS 2025 challenge, a privacy-aware Codabench competition on industrial multi-agent orchestration built on . We combine final rank sheets, a 300-submission server log, 149-team registrations, best-submission exports, the organizer winners report, the companion system paper, and verified planning-track source trees. Five results stand out. First, the public planning leaderboard saturates at 72.73\%, and richer prompts do not improve that peak. Second, hidden evaluation changes the story: public and private scores correlate moderately in planning (r{=}0.69) but negatively in execution (r{=}{-}0.13), with several 45.45\% public execution systems reaching 63.64\% on the hidden set. Third, the term is numerically almost inert in the official composite -- combined on a 0--1 scale with 0--100 percentage scores, it contributes at most 0.05 points per track, and rescaling would swap the top two teams. Fourth, the competition is operationally account-based but substantively team-based: 149 registered teams reduce to 24 with non-zero public scores and 11 fully ranked, while 52.3\% of deduplicated registrations list multiple usernames. Fifth, successful execution methods mostly improve guardrails -- response selection, contamination cleanup, fallback, and context control -- rather than novel agent architectures. These findings identify which behaviors the evaluation rewarded, and motivate scale-aware composites, skill-level diagnostics, and versioned artifact release.

A coding agent combines a model with a harness, which decides what the model sees, which tools it can use, and how the work continues. We ask whether changing the harness changes the result when the model and task stay fixed. We compare two configurations of the same harness on three coding benchmarks. The control supplies the full conversation in time order, while the treatment keeps the same record but mechanically shortens older tool results as the context fills and responds to repeated or stalled work. Under tight context, the treatment raises mean per-task fail-to-pass fraction (F2PF) in all three pressure comparisons and increases complete solutions on SWE-bench Verified and SWE-bench Pro. The tight-window Verified comparison uses 169 tasks, a 20,480-token window, and a fixed 480-second attempt endpoint; on this cohort, treatment raises mean per-task F2PF from 28 percent to 49 percent and complete solutions from 43 to 72. Without model-specific retuning, the same frozen treatment also raises both endpoints on the same cohort for three additional models with different designs. In the wide-window Qwen3.6 comparisons, observed arm outcomes are close on Verified and Pro, while FeatureBench retains a higher mean per-task F2PF under treatment. On the wide-window Verified cohort, treatment also serves fewer prompt tokens per turn. Because changing the harness changed what unchanged model weights could accomplish, coding-agent evaluations should treat the model and harness together as the tested solver.

False retry permission (FRP) occurs when an agent retries a timed-out tool call without support from the visible contract. RetryBench separates this error from realized recovery. Successful protected recovery does not imply reliable rejection of unsupported retries. In a 720-response key control, valid keys usually enabled safe progress, but offered changed keys elicited unsupported retries in Sol (1/80), Terra (55/80), and Luna (60/80). A later 240-response control removed the candidate and required models to supply a key. Luna still made 8/80 unsupported dispatches; Sol and Terra made none in 80 responses each. Accepting a displayed bad key therefore cannot explain all observed errors, although changed output requirements and conditions preclude a pure display-effect comparison. Removing the weak-contract label increased retries in Sol and Terra's coded-action conditions in a six-endpoint study. A separate 1,536-response canonical-action block also found a positive Terra contrast (115/128 versus 124/128 retries), concentrated in two task/order conditions; the other five contrasts remained unresolved after multiplicity correction. An explicit-contract cue control elicited deferral in all 762 returned responses, leaving recovery unresolved. These results support reporting permission and recovery together across declared interface variants. The evidence covers four purposive tasks, fixed GPT endpoints, and one recovery decision; it identifies neither internal mechanisms nor episode-level safety.

Coding agents often finish a session by returning one squashed patch that mixes feature implementation, bug fixes, refactorings, tests, and configuration edits. While the final code may be correct, collapsing unrelated edits into one patch removes the history structure needed for review, selective revert, and later maintenance. We study retrospective commit-history reconstruction: given a completed squashed change, an agent groups its hunks into commits and materializes a replayable commit sequence. We formalize the task as hunk-to-commit partitioning with a replay requirement and build AtomicCommitBench, containing 800 real consecutive-commit episodes from 10 Python projects. Because multiple decompositions may be reasonable, we evaluate outputs using complementary metrics: PPAR for replay validity, ARI for reference-based grouping quality, and TCR for failure containment on scoreable modified-test episodes. Natural retrospective reconstruction proves substantially harder than replay checking or synthetic tangling. Although nearly all non-random methods achieve replay validity (PPAR >= 0.988), grouping quality ranges from 0.03 to 0.46 ARI. Matched synthetic composites are much easier than real same-author squashed diffs (+0.333 ARI). In our evaluation, the GPT-5.4 setup driven by Codex CLI (0.46 ARI) and the GLM-5 setup driven by Claude Code (0.43 ARI) outperform MiniMax (0.31) and Kimi (0.29). Qualitative analysis identifies same-file lumping and support-hunk drift as recurring failure modes. Dependency-Aware Commit Evidence (DACE) improves the lower-scoring setups by 0.05 to 0.08 ARI, indicating that dependency cues and hunk-role information help agents avoid locality-driven grouping errors. AtomicCommitBench enables evaluation of the commit histories produced by coding agents alongside the final code.

We propose claim-level falsification as a principle for test-time scaling and instantiate it through Claim-Level Reliability Assessment (CLR), a training-free framework that reallocates test-time compute from additional solution sampling to targeted verification. Since whole-trace evaluation often obscures decisive errors due to signal dilution from routine tokens, CLR condenses each reasoning trace into a compact set of decision-critical claims, thereby isolating its logical anchors. Furthermore, recognizing the inherent difficulty of generating entirely correct solutions under fixed model capabilities, CLR shifts the focus to semantic falsification. This approach exploits a fundamental asymmetry between solution construction and claim refutation. Constructing a valid solution requires a flawless reasoning path, whereas refuting an incorrect claim requires identifying only a single decisive flaw. This targeted search for negative evidence systematically compresses the survival space of high-confidence incorrect traces, effectively suppressing erroneous consensus via nonlinear reliability scoring. Across four LLMs and four reasoning benchmarks under matched budgets, CLR generally improves upon pass@1 and self-consistency. On GPT-OSS-20B/CMIMC25, for instance, CLR exceeds pass@1 by 27.15 percentage-points and raises self-consistency accuracy from 77.50\% to 82.19\% with 37.0\% fewer tokens.

Robots that store past experiences must select which one to reuse in a new scene. Most systems select by visual similarity, and most evaluations report only the success of the selected experience. That number does not show whether the selection was good: a rule can score well by repeatedly using one broadly transferable experience, or poorly because its preferred experience is weak. Since robots increasingly adapt by reuse rather than retraining, a score that describes the library rather than the rule misleads what the field builds next. We contribute an audit methodology: execute every stored experience in every query scene, over two manipulation tasks, three reuse mechanisms, and libraries of K=3, 10, and 50. Because every alternative's outcome is known, a score can be traced to per-scene selection or to library quality. The audited rules select by nearest-neighbor distance in five visual embeddings, from raw pixels to CLIP. (1) One fixed experience, chosen with hindsight, captures 30-58% of the gap between random selection and an oracle; per-scene selection competes for the remaining 0.07-0.15 in success rate. (2) At Kge10, visual rules concentrate on one experience 1.5-3 times more than the oracle does, and their scores then follow that experience's quality. (3) Wherever a rule differs significantly from a shuffle that keeps its selection rates but pairs them with scenes at random, the rule is worse, for every learned image policy. (4) Visual distance predicts well whether a given pair will succeed (AUROC up to 0.96), yet ranks the candidates within one scene no better than chance for four of five embeddings at K=50 (AUROC 0.45-0.52). Exhaustive execution is usually infeasible, so the audit reduces to two cheap reports any study can give: the distribution of selected experiences, and the success of the best single experience in hindsight.

Agent Skills package procedural instructions and checks for use by general-purpose agents, but loading a skill does not guarantee that every requirement remains active throughout a long tool-using trajectory. We study this problem in a production-derived, white-box code-audit workflow. Holding the task and 24 artifact checks fixed, we vary the surrounding context and classify where failures first become visible: lost requirements, editing drift, failed checking, or non-agent evaluator/runtime failures. Codex with gpt-5.4-mini passes 8/10 runs in a 10,991-character clean context but only 3/10 in both a 299,140-character relevant context and an equal-length irrelevant context. This 50-percentage-point difference is large but remains trend-level under two-sided Fisher tests (p = 0.0698). Requirement coverage nevertheless stays above 92% in both long conditions, showing that a few omissions can invalidate an otherwise complete artifact. A second task passes all clean and long runs, so the evidence does not support a universal context-length threshold. A detailed external checklist passes 10/10 runs, compared with 5/10 for a generic self-check (p = 0.0325). Coding-agent scaffolds may help by selecting a smaller working set, but they do not eliminate failures. We do not introduce context rot or a new general monitoring method; we provide a bounded failure classification and empirical case study for white-box code auditing.

Self-consistency has emerged as a popular technique for improving large language model accuracy on reasoning tasks. The approach is straightforward: generate multiple reasoning paths and select the most common answer through majority voting. While this reliably boosts accuracy, it remains unclear whether these gains reflect genuine improvements in reasoning quality. We investigate a fundamental question that has not been studied before: does inference scaling improve reasoning faithfulness? We conduct a comprehensive empirical study across four frontier models (GPT-5.2, Claude Opus 4.5, Gemini-3-flash-preview, and DeepSeek-v3.2) on 100 GSM8K mathematical reasoning problems. Our analysis employs bootstrap confidence intervals, McNemar's tests for paired comparisons, and Cohen's d effect sizes to quantify the effects rigorously. The results reveal striking differences across models that challenge common assumptions about self-consistency. GPT-5.2 shows the expected pattern: accuracy improves from 78% to 90% at N=5, with faithfulness remaining relatively stable (0.540 to 0.510). Claude Opus 4.5 tells a completely different story. Its accuracy actually drops from 78% to 74.3% while faithfulness jumps dramatically from 0.270 to 0.891 at N=5. DeepSeek-v3.2, already at 98% accuracy, shows ceiling effects with modest faithfulness gains (0.440 to 0.541). Gemini-3-flash improves from 81% to 86% accuracy with a slight faithfulness decrease (0.260 to 0.212). Problem difficulty analysis reveals that GPT-5.2 solves 82% of hard problems while breaking only 13% of easy ones. Claude, in contrast, breaks 23% of easy problems, explaining its accuracy decrease. These findings matter for practitioners: self-consistency is not universally beneficial, and teams should test their specific models before deployment. We release our code and provide practical recommendations for navigating these tradeoffs.

Existing benchmarks for tool-using LLM agents primarily report single-run success rates and miss reliability properties required in production. We introduce ReliabilityBench, a benchmark for evaluating agent reliability across three dimensions: (i) consistency under repeated execution using pass^k, (ii) robustness to semantically equivalent task perturbations at intensity ε, and (iii) fault tolerance under controlled tool/API failures at intensity λ. ReliabilityBench contributes a unified reliability surface R(k,ε,λ), action metamorphic relations that define correctness via end-state equivalence rather than text similarity, and a chaos-engineering-style fault injection framework (timeouts, rate limits, partial responses, schema drift). We evaluate two models (Gemini 2.0 Flash, GPT-4o) and two agent architectures (ReAct, Reflexion) across four domains (scheduling, travel, customer support, e-commerce) over 1,280 episodes. Perturbations alone reduce success from 96.9% at ε=0 to 88.1% at ε=0.2. Rate limiting is the most damaging fault in ablations. ReAct is more robust than Reflexion under combined stress, and Gemini 2.0 Flash achieves comparable reliability to GPT-4o at much lower cost. ReliabilityBench provides a systematic framework for assessing production readiness of LLM agents.

OpenClaw-style agent stacks turn language into privileged execution: LLM intents flow through tool interception, policy gates, and a local executor. In parallel, skill marketplaces such as skills.sh make capability acquisition as easy as installing skills and CLIs, creating a growing capability supply chain. Together, these trends shift the dominant safety failure mode from "wrong answers" to execution-induced loss, where untrusted prompts, compromised skills, or narrative manipulation can trigger real trades and irreversible side effects. We propose Survivability-Aware Execution (SAE), an execution-layer survivability standard for OpenClaw-style systems and skill-enabled agents. SAE sits as middleware between a strategy engine (LLM or non-LLM) and the exchange executor. It defines an explicit execution contract (ExecutionRequest, ExecutionContext, ExecutionDecision) and enforces non-bypassable last-mile invariants: projection-based exposure budgets, cooldown and order-rate limits, slippage bounds, staged execution, and tool/venue allowlists. To make delegated execution testable under supply-chain risk, we operationalize the Delegation Gap (DG) via a logged Intended Policy Spec that enables deterministic out-of-scope labeling and reproducible DG metrics. On an offline replay using official Binance USD-M BTCUSDT/ETHUSDT perpetual data (15m; 2025-09-01--2025-12-01, incl. funding), SAE improves survivability: MDD drops from 0.4643 to 0.0319 (Full; 93.1%), |CVaR_0.99| shrinks from 4.025e-3 to ~1.02e-4 (~97.5%), and DG loss proxy falls from 0.647 to 0.019 (~97.0%). AttackSuccess decreases from 1.00 to 0.728 with zero FalseBlock in this run. Block bootstrap, paired Wilcoxon, and two-proportion tests confirm the shifts. SAE reframes agentic trading safety for the OpenClaw+skills era: treat upstream intent and skills as untrusted, and enforce survivability where actions become side effects.

Handwritten STEM exams capture open-ended reasoning and diagrams, but manual grading is slow and difficult to scale. We present an end-to-end workflow for grading scanned handwritten engineering quizzes with multimodal large language models (LLMs) that preserves the standard exam process (A4 paper, unconstrained student handwriting). The lecturer provides only a handwritten reference solution (100%) and a short set of grading rules; the reference is converted into a text-only summary that conditions grading without exposing the reference scan. Reliability is achieved through a multi-stage design with a format/presence check to prevent grading blank answers, an ensemble of independent graders, supervisor aggregation, and rigid templates with deterministic validation to produce auditable, machine-parseable reports. We evaluate the frozen pipeline in a clean-room protocol on a held-out real course quiz in Slovenian, including hand-drawn circuit schematics. With state-of-the-art backends (GPT-5.2 and Gemini-3 Pro), the full pipeline achieves approx8-point mean absolute difference to lecturer grades with low bias and an estimated manual-review trigger rate of approx17% at D_{max}=40. Ablations show that trivial prompting and removing the reference solution substantially degrade accuracy and introduce systematic over-grading, confirming that structured prompting and reference grounding are essential.

GPU training jobs fail often, roughly two in five on large production clusters, yet the operator typically learns of a failure only by reconnecting hours later. Experiment trackers require editing the training script and maintaining a cloud connection; the scheduler's mail hook delivers a single status line with no cause and no logs. GPUAlert is a command-line wrapper that monitors any training command at the process boundary, and with no change to that command, emails a structured notification on completion carrying a classified failure cause, durable logs, and output artifacts. The tool is organized around three reliability primitives: a pre-launch log guarantee that establishes the durable destination before the child process can crash, notifier isolation that makes the wrapper's exit code a pure function of the child's status regardless of whether the email succeeds, and a non-silent artifact budget that bounds attachment size without ever dropping output silently. We release a labelled corpus of 474 GPU training logs across 15 failure classes and a reproducible evaluation harness. On the twelve hardware-reproduced classes, the ordered-rule classifier reaches 0.997 macro-F1, against 0.830 for unordered keyword matching and 0.133 for exit-code inspection. Wrapper overhead is a constant approximately 3ms per job; the pre-launch guarantee preserves a log where a shell redirect yields nothing; and across all 15 failure modes the wrapper returns the child's exit code unchanged even when the SMTP relay is unreachable.

Direct-decision models turn text into low-latency structured labels and scores, making them attractive for classification and automatic evaluation. Yet reliability requires more than accuracy: a model must also use the ordinal decision scale supplied by the user faithfully. We analyze JEV~1.13 and three open KEV models. Our investigation begins with ANLI, where JEV assigns 38.8\% of all predictions and 51.3\% of errors to Neutral despite 74.95\% accuracy, nearly balanced gold labels, and balanced candidate positions. Across 36 ordinal datasets, final decisions use only 67--76\% of the effective gold support, versus 87--102\% on four nominal tasks. Randomizing candidate order weakens but does not remove this compression. Holding items and source scores fixed while balancing gold support and positions, we refine scales from K=2 to 14; utilization falls for every model and reaches 26--75\% at K=14, although candidate probabilities remain broad for most models. Targeted BA-LoRA post-training raises gold-relative utilization from roughly 47\% to 86\% on eight supervised scales at both KEV sizes, showing that the compression is learned and modifiable rather than an immutable architectural limit. We call this ordinal scale-utilization bias: decision-stage candidate-space compression distinct from accuracy, gold imbalance, fixed position, and candidate count alone. The code and data are available at https://github.com/Glax147/jev_ordinal_scale_bia

AI systems have become increasingly capable of dangerous behaviours in many domains. This raises the question: Do models sometimes choose to violate human instructions in order to perform behaviour that is more useful for certain goals? We introduce a benchmark for measuring model propensity for instrumental convergence (IC) behaviour in terminal-based agents. This is behaviour such as self-preservation that has been hypothesised to play a key role in risks from highly capable AI agents. Our benchmark is realistic and low-stakes which serves to reduce evaluation-awareness and roleplay confounds. The suite contains seven operational tasks, each with an official workflow and a policy-violating shortcut. An eight-variant shared framework varies monitoring, instruction clarity, stakes, permission, instrumental usefulness and blocked honest paths to support inferences regarding the factors driving IC behaviour. We evaluated ten models using deterministic environment-state scorers over 1,680 samples, with trace review employed for audit and adjudication purposes. The final IC rate is 86 out of 1,680 samples (5.1%). IC behaviour is concentrated rather than uniform: two Gemini models account for 66.3% of IC cases and three tasks account for 84.9%. Conditions in which IC behaviour is indispensable for task success result in the greatest increase in the adjusted IC rate (+15.7 percentage points), whereas emphasising that task success is critical or certain framing choices do not produce comparable effects. Our findings indicate that realistic, low-nudge environments elicit IC behaviour rarely but systematically in most tested models. We conclude that it is feasible to robustly measure tendencies for dangerous behaviour in current frontier AI agents.

Reliable confidence estimation is increasingly central to the trustworthy deployment of language models: a calibrated estimate of the probability that an output is correct decides what to ship, what to escalate, and what to retry. Existing confidence estimators, however, share one design premise: they only read the current inference process, either by introspecting on it, scoring its token probabilities, or resampling it. We argue that the current inference is not a sufficient basis for confidence. We propose XConf (eXperiential Confidence): estimating confidence together with the model's accumulated experience. The experience is stored as a record of the model's own graded past episodes, each holding the task, the model's reflection, its stated confidence, the outcome, and a lesson written once the grade arrived. Given a new task, XConf's Recall stage retrieves past episodes on similar tasks met with a similar stated confidence, and reads off their historical success rate; its Reflect stage shows the model this record, has it name its recurring failure mode, and restate a confidence now informed by its own track records. Our estimator is format-general, requiring no logit access or weight updates, and costs only one answer generation. Across nine benchmarks spanning reasoning, coding, multimodal QA, and interactive agents, and four models from three families, XConf beats or matches ten-sample self-consistency in discrimination (AUROC) on 23 of 24 comparisons, with much lower calibration error (ECE), at a tenth of the generation cost. Used for selective prediction, abstaining on the 10% least-confident episodes raises the delivered success rate by up to 8.7 points on agent tasks. We therefore see experiential confidence estimation as a new paradigm for future general-purpose confidence estimation.

Clinical research over electronic health records depends on derived variables whose definitions sit outside the schema, in item codes, time windows, missingness rules, and output-grain choices that change cohorts and downstream conclusions. We introduce M4Bench, 28 clinical derivation tasks across 15 task families on MIMIC-IV and eICU, to test whether AI coding agents can reproduce such derivations and how clinician-reviewed agent skills change their results. Its controls decompose the value of skills into specification completeness, delivery format, and schema familiarity. In a Docker-isolated Codex campaign (GPT 5.4 Mini and GPT 5.5), skills raise reward by 0.136 over the no-skill baseline, with 26 of 28 tasks improving. Specification completeness accounts for most of this gain: on an eight-task control subset, an operational spec alone raises reward by 0.162, and skills add only 0.024 on top, indistinguishable from zero. On a ten-task subset, decoy skills score like no skill and raw reference SQL nearly matches targeted skills, arguing against generic context or the skill format as the explanation. Schema obfuscation and restructuring preserve 86% and 70% of the skill gain (GPT 5.4 Mini, 26 MIMIC-IV tasks), a descriptive robustness check rather than a bound on memorization. A descriptive Claude Code replication on all 28 tasks (Opus 4.7 and Sonnet 4.6, 560 runs) is directionally consistent (+0.120, 21 of 28 tasks improving). These gains are partial agreement with the reference, not completed tasks: only 25 of 280 with-skill and 13 of 280 no-skill Codex runs pass every strict check, and skills barely change which rows agents return (key F1 +0.003). Skills are released as a versioned, clinician-reviewable markdown delivery format, with a family-level reference-validity audit. M4Bench releases tasks, skills, evaluation code, and sanitized per-run evidence as an open-source benchmark.

Patient world models are increasingly proposed as simulators of clinical trials, and recent surveys rank their capabilities on a ladder from temporal prediction to counterfactual rollouts. Position on that ladder is currently assigned by reading papers rather than by measurement. We turn the counterfactual rung into a measured quantity. From 30,119 completed phase 2/3 trials with results posted to ClinicalTrials.gov we construct 9,996 audited within-trial contrasts from 7,207 trials, and we score a simulator on how much of its accuracy depends on knowing which intervention was tested. The instrument is an intervention-blind reference: the same model, trained on the same trials, with every intervention-identifying field withheld. Interventional skill is the error the full model removes relative to that reference. On a temporal split (878 test contrasts from 655 trials) a registry-trained gradient-boosting simulator removes a fifth of the error of a trivial baseline, but its interventional skill is -0.12 (95% CI -0.21 to -0.05): withholding the intervention makes it better, and no alternative featurisation, stratum, temporal cut or same-molecule retrieval yields an interval above zero. A frontier language model reading only the registered protocol has interventional skill 0.20 [0.08, 0.31]; smaller and open-weight models sit at the limit of detection. Perturbation tests show the large model uses the intervention as such: swapping the arms reverses the sign of its prediction in 86% of contrasts. The blind reference is itself not drug-blind (a language model recovers the withheld drug class 45% of the time from the blind prompt), so interventional skill is a conservative lower bound. A reliability layer shows that ranking by predicted error lowers it but that calibrated error control does not transfer under temporal shift. Benchmark, code, prompts and per-call model outputs are released.

Large language model (LLM) agents increasingly rely on skills, structured documents that specify when to act, which procedure to follow, and which tools are allowed. Existing evaluations mostly judge the quality of a skill or its contribution to task success, leaving unexamined whether an agent can recognize a relevant skill and apply it on its own. We introduce Skill-Use, a benchmark that evaluates skill use under progressive disclosure, where an agent sees only a skill's name and short description and must retrieve the full procedure before following it. Skill-Use separates three facets of skill use. Trigger measures whether the agent invokes the relevant skill, Compliance measures how faithfully it follows the prescribed procedure, and Boundary measures whether it avoids forbidden operations. A Skill-Use (SU) score combines the three and credits execution only after the skill is triggered. Skill-Use pairs 79 real skills with 177 executable tasks across nine domains, each grounded in real files, run in an isolated Docker sandbox, and scored by a trajectory-based rubric. Evaluating eight LLMs under two agent harnesses, we find that reliable skill use remains out of reach, as the strongest configuration reaches an SU of only 0.613. Triggering and procedural compliance fail as independent bottlenecks, and both scores and model rankings shift with the harness, so skill use behaves as a capability conditioned on the harness rather than a fixed property of the model.

LLM agents act in external environments where each action changes the state that later decisions condition on, and where a single wrong step can waste interaction budget or trigger irreversible side effects long before the final failure is observed. Reliable deployment therefore requires step-level confidence estimation: a calibrated probability that each proposed action is productive, available before the action is executed. Existing LLM confidence estimators are designed to score a response from the given prompt, but agent confidence also depends on execution consequences: whether similar actions in similar situations actually advanced the task after the environment responded. We introduce the \method (\methodshort), a self-evolving critic framework in which an LLM critic accumulates evidence from its own past judgments and their observed consequences. After each trajectory, a hindsight LLM that sees the full execution feedback votes on whether each step was productive. The resulting pseudo-labels populate a memory bank from which related productive and unproductive experiences are retrieved into the critic's prompt whenever a similar step recurs. \methodshort requires no training and uses no ground truth step labels. Across three agent benchmarks and three critic backbones, \methodshort attains the best calibration (ECE and Brier) and ranking (AUC) in every dataset--critic combination, reducing ECE by up to 54% relative to the strongest training-free baseline.

A probability forecast or probabilistic classifier is reliable or calibrated if the predicted probabilities are matched by ex post observed frequencies, as examined visually in reliability diagrams. The classical binning and counting approach to plotting reliability diagrams has been hampered by a lack of stability under unavoidable, ad hoc implementation decisions. Here we introduce the CORP approach, which generates provably statistically Consistent, Optimally binned, and Reproducible reliability diagrams in an automated way. CORP is based on non-parametric isotonic regression and implemented via the Pool-adjacent-violators (PAV) algorithm - essentially, the CORP reliability diagram shows the graph of the PAV- (re)calibrated forecast probabilities. The CORP approach allows for uncertainty quantification via either resampling techniques or asymptotic theory, furnishes a new numerical measure of miscalibration, and provides a CORP based Brier score decomposition that generalizes to any proper scoring rule. We anticipate that judicious uses of the PAV algorithm yield improved tools for diagnostics and inference for a very wide range of statistical and machine learning methods.

People overthink; language models over-sample, and the extra effort can talk both into a worse answer. Reasoning systems answer a hard question by sampling it many times (test-time scaling), and the more they draw, the more often a correct answer turns up somewhere, so coverage, the fraction of problems with at least one correct try, climbs and appears to be progress. But a deployed system must return one answer, and choosing it, not knowing which try is right, is selection; selection is capped, and past a point extra samples only make the model surer of a confident mistake, even as every draw adds cost. The gap between climbing coverage and stalled selection, the identifiability gap, is the answer a model can produce but not pick. So the real question is not whether to sample but how far, and the answer is: not far. For picking an answer, the vote has already settled within a few dozen draws, the modal ceiling; for scoring a benchmark, sooner still, the correlation ceiling. Beyond that, extra draws cost compute and add nothing, and can even make the answer worse. This paper turns the cutoff into a single number, the effective number of samples, that any sampling run already reveals. The bottleneck is recognizing a right answer, not generating one.

Evaluating language models fairly is becoming harder as static benchmarks available on the internet risk contamination by training data. This makes it unclear whether models are truly reasoning or just recalling answers. In this paper, we introduce BeyondBench, an evaluation framework that avoids this problem by using algorithmic problem generation. Unlike traditional benchmarks that risk contamination from internet-scale training data, BeyondBench creates mathematically grounded problems on the fly, ensuring each test remains fresh and uncontaminated. Our framework covers 44 algorithmic tasks with a total of 117 variations, grouped into three difficulty levels: the Easy Suite (29 tasks) for basic arithmetic and statistics, the Medium Suite (5 tasks, 49 variations) for sequence patterns and reasoning, and the Hard Suite (10 tasks, 68 variations) tackling NP-complete and constraint satisfaction problems. Each task generates problems from a combinatorial space larger than 10^15 unique instances, with solutions verified deterministically by mathematical proofs. We evaluated 101 language models, including 85 open-source and 16 closed-source models, spanning sizes from 0.5B to 141B parameters and multiple quantization schemes. Our results show consistent reasoning deficiencies across model families, with performance degrading sharply as problem complexity increases from polynomial to exponential. In our Hard Suite evaluations, models such as Gemini-2.5-pro, Llama-3.3-70B, and Qwen2.5-72B achieved average accuracies of 56.38%, 26.91%, and 33.60%, respectively. Moreover, we observe that performance drops drastically without tool usage, with GPT-5, GPT-5-mini, and GPT-5-nano showing a decline of 16.81%, 28.05%, and 47.59% accuracy on the hard suite. Our leaderboard is publicly available at https://ctrl-gaurav.github.io/BeyondBench/

Banking assistants must use account-specific information to answer requests and, in many cases, take actions through tools. Evaluating only the final response misses important errors. An assistant may ask for information it already has, rely on stale context, select the wrong account, or write an invalid value after stating the correct one. We introduce IndicBankBench, a 799-case benchmark for Indian retail banking spanning five operational domains, a capability/refusal domain, and twenty primary axes. Cases are evaluated at four stages: safety, action and tool use, response adequacy, and advisory quality. Tool use and most safety checks are deterministic. A narrow resolver handles only ambiguous confirmation-before-write cases, while a separate LLM judge evaluates semantic response adequacy. We run every case three times and report strict pass^3, which requires success on all trials. Across the eleven evaluated models, strict reliability ranges from 43.7% to 58.2%, whereas at-least-once success ranges from 60% to 74%. This gap shows that at-least-once success can overstate dependable banking behavior. The case-level diagnostics also distinguish systems that ask unnecessary questions from those that act but fail to reconcile customer context or fully resolve the request. We release the cases, mock environment, and evaluation harness.

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.

Safety benchmark scores provide incomplete evidence of deployment readiness: aligned language models often adhere to rigid rules even when a situational update flips which action is safe. We term this failure brittle safety. To diagnose it, we introduce context-flip evaluation, testing 12 models across a safety benchmark (PacifAIst) and two commonsense controls using paired variants where the nominally safe action produces harm. Three findings emerge. First, brittle safety is safety-specific: all 12 models exhibit a safety-commonsense gap (mean +17.4 pp). Baseline accuracy fails to predict brittleness: among models above 90% baseline accuracy, brittleness rates range from 13.7% to 90.0%. Second, failures stem from policy override rather than miscomprehension: despite acknowledging the context change in every case, models persist via three distinct mechanisms that vary by update type and model family. Third, on a hand-audited probe of catastrophic consequence-flip scenarios, standard action-level guardrails catch none, while a state-aware validator catches all without false alarms on correct interventions. This indicates that action-level content moderation is systematically blind to consequence-flips, motivating state-aware architectural alternatives. We release our protocol, perturbed benchmarks, and deployment probe.

Reusable Web-agent verification requires a rule to follow changing requirements and support subsequent execution. We introduce AgentProofBench, which connects paired tests of a frozen acceptance rule to execution from identical saved openings. The resource spans 97 tasks, 79 families, and four Web services. Experiments with DeepSeek Flash/Pro establish two linked measurements. Across 48 families, complete-panel and decision-replay tests retain 82/96 Flash and 79/96 Pro source programs; paired reuse-rebuild BA intervals are [-5.47, 1.57] and [-10.16, 0.00] percentage points. Across 22 execution families, deployment-state panel correctness separates program assignments whose two saved openings are both delivered correctly: 32/33 assignments with error-free panels, versus 1/11 with panel errors. All reference-defined program-gate regressions and withheld deliveries occur in the latter group. Neutral continuation attains the highest correct delivery under the tested generic feedback, providing a concrete baseline for measuring the value added by a gate. Released instructions, states, programs, reference rules, and trajectories connect each test failure to its execution consequences. AgentProofBench makes reusable verification assessable at both the requirement and policy levels.

LLM agents fail mid-episode -- they loop, cascade tool errors, drift off goal, fabricate results, or silently absorb corrupted content -- and the standard remedy, judging every step with a second LLM, costs more than the agent itself. We ask how much detection is achievable from observable step telemetry alone, using monitors costing microseconds per step and trained only on healthy runs. On 2,823 committed agent episodes across three frameworks, three local models (qwen2.5 7b/3b, llama3.1 8b) and a commercial API (gemini-2.5-flash), a one-class echo-state-network ensemble with CUSUM alarms detects 0.71 of failures at a 5% false-alarm budget (AUROC 0.872). Its advantage over a memoryless baseline is a monotone function of post-onset horizon (+0.09 at <=3 steps, +0.40 at >=9), predicting its own failure region out of sample on AFTraj-2K. Ranking transfers with no retraining to two corpora from other groups (AFTraj-2K 0.745, ATBench 0.779). Monitors carry two burdens: a per-deployment healthy null (they do not transfer -- AUROC 0.527 cold against 0.885 recalibrated) and a residual false-alarm rate. We add a layer carrying neither: deterministic verification, which recomputes a run's stated total from the tool results it actually received and confirms every required call was made. Head-to-head it catches 60% of failures (96% with the coverage check) at 0 of 63 false positives against the monitor's 54% at 17%, transfers unchanged to llama3.1:8b (110 of 110 at 0 of 10), and trips on 0 of 1825 healthy episodes. Detection is then closed into repair: each flagged run is rolled back and re-run live, recovering 45% of failures against a 16% resampling control (p=0.0005) and lifting task success from 52% to 73% for about one extra model call per run. The system runs at ~200 microseconds per step, three orders of magnitude below a judge call. Code, traces and results are released.

Agent Skills package reusable natural language procedures with executable resources, enabling software agents to acquire task specific capabilities without model adaptation. Automatically generating such Skills can improve task performance, yet evaluating a candidate solely from its artifact or final task outcome leaves unresolved which actions the equipped agent will perform and which side effects those actions will produce. We present TRUSS, an evidence guided framework for generating functionally effective and safety reliable Agent Skills. TRUSS first inspects functional claims against source and domain evidence while evaluating the complete artifact under nine predefined safety properties. Candidates admitted by this static gate are loaded by a shadow agent inside a Controllable Execution Environment, where brokered tools expose requested actions to policy enforcement and record their results as provenance preserving execution traces. Functional failures and property violations are linked back to the responsible Skill content and used to guide iterative refinement. We evaluate TRUSS on 168 SkillInject artifacts, 155 SkillSafetyBench cases, and all 187 tasks in SkillGenBench. TRUSS achieves 100.00\% precision and recall in vulnerability detection. Repair reduces attack success from 38.71\% to 19.35\% with GPT 5.5 and from 46.45\% to 29.68\% with GPT 5.4, with zero attack regression. For Skill generation, TRUSS raises task effectiveness from 17.11\% without Skills to 52.94\%, while increasing the benchmark Security rate from 50.80\% to 100.00\%. These results show that execution evidence can expose behavioral failures missed by artifact inspection and can guide Skill generation toward jointly verified functional and safety outcomes.

Language-model agents are increasingly used as persistent coworkers that assist users across multiple working days. During such workflows, the surrounding environment may change independently of the agent: new emails arrive, calendar entries shift, knowledge-base records are updated, and evidence appears across images, scanned PDFs, audio, video, and spreadsheets. Existing benchmarks do not adequately evaluate this setting because they typically run within a single static episode and remain largely text-centric. We introduce , a benchmark for coworker agents built around multi-turn multi-day tasks, a stateful sandboxed service environment whose state evolves between turns, and rule-based verification. The current release contains 100 tasks across 13 professional scenarios, executed against five stateful sandboxed services (filesystem, email, calendar, knowledge base, spreadsheet) and scored by 1537 deterministic Python checkers over post-execution service state; no LLM-as-judge is invoked during scoring. We benchmark seven frontier agent systems. The strongest model reaches 75.8 weighted score, but the best strict Task Success is only 20.0\%, indicating that partial progress is common while complete end-to-end workflow completion remains rare. Turn-level analysis shows that performance drops after the first exogenous environment update, highlighting adaptation to changing state as a key open challenge. We release the benchmark, evaluation harness, and construction pipeline to support reproducible coworker-agent evaluation.

Self-evolving agents violate the assumption behind most learning-theoretic guarantees: the data, evaluator, components, and hypothesis space are produced by the policy being updated. We present SEA, an architecture that confines self-modification to a small steering adapter and a versioned harness around a frozen base model and admits each modification only through an anytime-valid gate that emits an auditable certificate against a fixed error budget. Five loop controllers compose published guarantees; because such gates can only select among behaviors the frozen base already produces, five verifier-in-the-loop mechanisms -- best-of-N, micro-step search, self-authored reproduction oracles, search-layer control, and self-repair -- supply the dense, grader-free signal the gates require, computed from the issue text alone. On a 52-instance SWE-bench Verified subset across four base models, base capability is the dominant, confound-free effect, and on two strong base models a deliberate no-op-composite control isolates the suite's contribution at +4 and +5 (Glm~5.2 24to28; Gpt 29to34, the 65% best), with event logs confirming that its mechanisms fire and prevent regressions. Results are single-run on expensive evaluations; confirming run-to-run variance and adapting the per-task algorithm mix are future work.

Pass@k is widely used to report the reasoning performance of LLMs, but it often produces unstable and potentially misleading rankings, especially when the number of trials (samples) is limited and computational resources are constrained. We present a principled Bayesian evaluation framework that replaces Pass@k and average accuracy over N trials (avg@N) with posterior estimates of a model's underlying success probability and credible intervals, yielding stable rankings and a transparent decision rule for differences. Evaluation outcomes are modeled as categorical (not just 0/1) with a Dirichlet prior, giving closed-form expressions for the posterior mean and uncertainty of any weighted rubric and enabling the use of prior evidence when appropriate. Theoretically, under a uniform prior, the Bayesian posterior mean is order-equivalent to average accuracy (Pass@1), explaining its empirical robustness while adding principled uncertainty. Empirically, in simulations with known ground-truth success rates and on AIME'24/'25, HMMT'25, and BrUMO'25, the posterior-based procedure achieves faster convergence and greater rank stability than Pass@k and recent variants, enabling reliable comparisons at far smaller sample counts. The framework clarifies when observed gaps are statistically meaningful (non-overlapping credible intervals) versus noise, and it naturally extends to graded, rubric-based evaluations. Together, these results recommend replacing Pass@k for LLM evaluation and ranking with a posterior-based, compute-efficient protocol that unifies binary and non-binary evaluation while making uncertainty explicit. Source code is available at https://github.com/mohsenhariri/scorio

Executable Agent Skills combine natural-language instructions and scripts into reusable packages for LLM agents, and revising them requires fixing errors without breaking correct behavior. Existing benchmarks do not systematically distinguish documentation repair, script repair, and preservation when evaluating skill self-evolution. We introduce SkillScriptBench, a 350-task benchmark designed to evaluate these capabilities separately. From a survey of over 35,000 GitHub-hosted Skill roots, we select 100 packages and construct 150 repair tasks. Each task pairs a package containing injected script faults with a maintenance request and executable checks of the required behavior. A complementary controlled track contains 200 tasks from 50 packages, each evaluated under the same maintenance request in four states: clean, documentation faults, script faults, and faults in both. Across four LLMs, methods that edit both documentation and scripts can repair script faults but do not consistently outperform Markdown-only revision on documentation repair or preservation. We therefore introduce AST-Guided Skill Revision, which uses abstract syntax trees and calling relationships to link maintenance requirements to relevant code locations. It restricts script edits to these locations and updates the documentation to match the revised scripts. Averaged across models, this revision stage yields absolute gains in repair success of 21.9% for Raw Package and 27.7% for CoEvoSkills on faulty packages. Absolute gains in the proportion of tasks solved in all three runs reach 20.8% and 31.5%, respectively, indicating more consistent repair success across repeated runs.

Recursive self-improvement requires models not only to acquire useful training experience, but also to decide when their current experience has ceased to be useful. Existing methods often focus on what experience to generate next, while paying less attention to when fresh experience should be introduced. Many also construct training targets through self-consistency or majority voting, making supervision depend on the model's own uncertain predictions. We address these two questions separately: training dynamics determine when to acquire new experience, while the answers to new tasks remain independently checkable. A-RSI monitors reward variation and accuracy in training logs to refresh a task pool when its learning signal declines, without consulting validation or test scores. A static library of model-authored programs produces multimodal questions and derives their reference answers; we specify the task-generation contract, automatically verify chosen properties of the reward and refresh rules, and separately audit the generated answers. The training loop uses no external training examples or human-written answer labels. On the primary eleven-benchmark suite, A-RSI raises the performance average from 57.77 to 61.05 with 800 fresh tasks.

LLM-as-judge ("grader") components are now standard in evaluation harnesses, including safety evaluations where a pass/fail verdict may gate downstream deployment decisions. A widespread assumption is that setting the grader's sampling temperature to 0 makes grading deterministic. We test this assumption against a real safety-evaluation codebase (Japan AISI's open-source aisev) and show it fails on two levels. First, the harness invokes its grader without setting temperature or seed; the underlying provider silently applies its default of 1.0, so items near the decision boundary flip pass/fail across identical runs (per-item disagreement up to ~50% over 20 runs). Second, pinning temperature=0 reduces but does not eliminate flips: across 690 API calls spanning two providers, three model tiers, and five sampling configurations, 1-2 of 7 borderline items remain non-reproducible even under forced greedy decoding (top_k=1). Claude Opus 4.7/4.8 has since deprecated temperature entirely, rendering the primary mitigation inapplicable to newer model generations. These findings expose a structural gap: evaluation harnesses that report single-run verdicts without variance or grader-disagreement metrics can present noise as a safety property. We release a reproduction harness (690 calls, 7 conditions) and recommend that harnesses treat grader disagreement as a first-class health metric alongside the scores themselves.

Human supervisors rely on coding agents' reports to assess progress, verify repairs, and decide when to intervene. But a supervisor cannot investigate a failure they were never told occurred. Failed attempts, workarounds, and unresolved risks may disappear from routine reports, leaving supervisors unaware of issues that warrant scrutiny. We conduct CTXGAP, a measurement study quantifying the gap in task-state recoverability between agents' accumulated user-facing reports and their full recorded execution contexts. Across 37 coding tasks and 4,799 valid probes, agents answer 93.6% of task-state questions on demand; full-context readers recover 92.7%, while report-only readers recover 52.0%, a 40.7-point gap. At the final checkpoints of 30 verified fixes, the gap remains 34.8 points. Human annotation identifies states mentioned without enough supporting detail as the largest category of unrecovered information, and model-assessed supervision-relevant states retain substantial gaps. Full terminal scrollback improves recovery to 73.6% but still trails full context. Under matched report lengths, context-informed rewrites improve recovery by about eight points, without a detectable targeted-over-natural advantage. Positive gaps persist across harnesses and independent DeepSeek execution and measurement. CTXGAP establishes report-based task-state recoverability as a complementary evaluation dimension.

We introduce the first version of KWBench (Knowledge Work Bench), a benchmark for unprompted problem recognition in large language models: can an LLM identify a professional scenario before attempting to solve it. Existing frontier benchmarks have saturated, and most knowledge-work evaluations to date reduce to extraction or task completion against a specification. KWBench targets the step before that: recognizing the governing structure of the situation from raw inputs alone. The benchmark contains 223 tasks sourced from practitioners across acquisitions, contract negotiations, clinical pharmacy, organizational politics, fraud analysis, and incentive design. Each task encodes a formal game-theoretic pattern (principal-agent conflict, signaling, mechanism design failure, strategic omission, coalitional dynamics, strategic interdependence) and carries structured ground truth recording the expert reading of the situation and the anticipated failure modes. Models receive raw data and a task prompt with no indication of problem type. Scoring is a three-tier rubric gated by a mandatory conjunctive check. Mandatory criteria encode the predicted wrong paths. We evaluate 16 models. The best model passes on 27.9% of tasks. The top two models agree on only 31.7% of their passes. Among the top 8, 44 tasks are solved by exactly one model; routing across the top 8 covers 50.7% of the benchmark, nearly double the best single model. Conditional on passing, quality scores converge (approx 83% across models); unconditional scores do not. Same models articulate the relevant game-theoretic concept correctly when asked, then fail to apply it unprompted. We release KWBench to shift how frontier models are evaluated on knowledge work, scoring them on whether they recognize the right problem from the situation alone, not only on how well they execute once the problem has been framed for them.

Fine-tuning large language models (LLMs) on reasoning benchmarks via reinforcement learning requires a specific reward function, often binary, for each benchmark. This comes with two potential limitations: the need to design the reward, and the potentially sparse nature of binary rewards. Here, we systematically investigate rewards derived from the probability or log-probability of emitting the reference answer (or any other prompt continuation present in the data), which have the advantage of not relying on specific verifiers and being available at scale. Several recent works have advocated for the use of similar rewards (e.g., VeriFree, JEPO, RLPR, NOVER). We systematically compare variants of likelihood-based rewards with standard baselines, testing performance both on standard mathematical reasoning benchmarks, and on long-form answers where no external verifier is available. We find that using the log-probability of the reference answer as the reward for chain-of-thought (CoT) learning is the only option that performs well in all setups. This reward is also consistent with the next-token log-likelihood loss used during pretraining. In verifiable settings, log-probability rewards bring comparable or better success rates than reinforcing with standard binary rewards, and yield much better perplexity. In non-verifiable settings, they perform on par with SFT. On the other hand, methods based on probability, such as VeriFree, flatline on non-verifiable settings due to vanishing probabilities of getting the correct answer. Overall, this establishes log-probability rewards as a viable method for CoT fine-tuning, bridging the short, verifiable and long, non-verifiable answer settings.

Evaluating the reliability of news sources is a routine task for journalists and organizations committed to acquiring and disseminating accurate information. Recent research has shown that predicting sources' reliability represents an important first-prior step in addressing additional challenges such as fake news detection and fact-checking. In this paper, we introduce a novel approach for source reliability estimation that leverages reinforcement learning strategies for estimating the reliability degree of news sources. Contrary to previous research, our proposed approach models the problem as the estimation of a reliability degree, and not a reliability label, based on how all the news media sources interact with each other on the Web. We validated the effectiveness of our method on a news media reliability dataset that is an order of magnitude larger than comparable existing datasets. Results show that the estimated reliability degrees strongly correlates with journalists-provided scores (Spearman=0.80) and can effectively predict reliability labels (macro-avg. F_1 score=81.05). We release our implementation and dataset, aiming to provide a valuable resource for the NLP community working on information verification.

We present CogEvol, a family of models trained specifically for Learning Environment Generation: turning a course brief into a finished learning artifact (structured-JSON slides or self-contained interactive HTML pages) in a single pass. Across 220k production requests, CogEvol completes a slide in a median of 17 seconds and an interactive page in 59, replacing minutes-long multi-turn agent scaffolding. Reliability is enforced rather than hoped for: a production-grounded data pipeline turns real failures into 53,687 verified SFT samples, and a hybrid rule-plus-VLM reward drives GRPO-based RL, hardened after we caught and fixed a reward-hacking episode that produced visually convincing but unplayable games. CogEvol-27B scores 83.7 on slide quality and 63.7 on a 500-case interactive-HTML benchmark with 26.9x fewer parameters than flagship coding models, and, in collaboration with the OpenMAIC team, serves their live production traffic. CogEvol-4B is released openly under the Apache 2.0 license at https://github.com/CogEvol/CogEvol-4B; external flagships are measured on the same suites under the identical harness. Scaffold editing cuts interactive-page generation cost by a further ~76%, and the full stack runs on domestic Ascend accelerators at application-level parity with A800 GPUs, lowering the unit cost of AI-native education at scale.

Existing defenses against hallucination in retrieval-augmented and multi-agent pipelines remain partial: evidence is trusted despite modality disagreement, debate verifies an aggregate report rather than individual claims, and such verification occurs only after drafting, leaving inter-agent errors undetected until the final text. To close this gap, we present CLAIR-Fin, a nine-agent framework that decomposes each question into atomic claims maintained in a typed Financial Claim Ledger. Each claim is resolved through Asymmetric Evidence Authority, which conditions evidence trust on claim type rather than treating all modalities as equally reliable; Chain-of-Custody Verification, which checks grounding at the hand-off between drafting and adversarial review rather than only at the pipeline's exit; an Adaptive Rebuttal Cycle, which routes contested claims through adversarial debate whose depth scales with what that debate finds; and a terminal entailment audit paired with a continuous Hallucination Risk Index that distinguishes claims that passed scrutiny from claims never contested. We evaluate CLAIR-Fin on BB-FinQA-X, a 500-question cross-modal financial evaluation set built from Bangladesh Bank Annual Report material, stratified by query type, format, and difficulty. Relative to a single-pass retrieval-augmented generation baseline, it raises faithfulness (0.780 rightarrow 0.889) while abstaining on 5.4% of questions when evidence is insufficient rather than forcing an unsupported response, and it exceeds stronger retrieval-strategy baselines such as HyDE and Graph-RAG on faithfulness (leq 0.874).