Daily Papers

Uncertainty quantification (UQ) methods for language models are typically evaluated on single-turn outputs, where uncertainty is attached to one generated answer. For LLM agents, however, the unit of observation is an interactive trajectory, where the model can ask clarifying questions, call tools, update state, and make intermediate decisions whose errors propagate to the final outcome. We study whether three common families of single-turn UQ methods transfer to this setting. Across five LLMs and four multi-turn tool-use datasets from BFCL-v4 and τ^2-bench, we evaluate white-box scorers based on action-token probabilities, black-box consistency scorers based on resampled trajectories, and reflexive scorers based on model self-assessment of the trajectory. We find that transfer is often useful but uneven. Token-probability scores are highly sensitive to the choice of aggregator used across turns, reflexive scores provide the strongest low-cost baseline in most evaluated settings, and black-box self-consistency is often the strongest UQ family, with trajectory-equivalence and action-set consistency typically ranking highest among its variants. These results suggest that UQ methods developed for single generations should be revalidated at the trajectory level, with careful attention to the consistency measurement, aggregator choice, and computational budget.

Large language models (LLMs) have been increasingly used to interact with external environments (e.g., games, compilers, APIs) as goal-driven agents. However, it remains challenging for these language agents to quickly and efficiently learn from trial-and-error as traditional reinforcement learning methods require extensive training samples and expensive model fine-tuning. We propose Reflexion, a novel framework to reinforce language agents not by updating weights, but instead through linguistic feedback. Concretely, Reflexion agents verbally reflect on task feedback signals, then maintain their own reflective text in an episodic memory buffer to induce better decision-making in subsequent trials. Reflexion is flexible enough to incorporate various types (scalar values or free-form language) and sources (external or internally simulated) of feedback signals, and obtains significant improvements over a baseline agent across diverse tasks (sequential decision-making, coding, language reasoning). For example, Reflexion achieves a 91% pass@1 accuracy on the HumanEval coding benchmark, surpassing the previous state-of-the-art GPT-4 that achieves 80%. We also conduct ablation and analysis studies using different feedback signals, feedback incorporation methods, and agent types, and provide insights into how they affect performance.

Many students lack access to expert research mentorship. We ask whether an AI mentor can move undergraduates from an idea to a paper. We build METIS, a tool-augmented, stage-aware assistant with literature search, curated guidelines, methodology checks, and memory. We evaluate METIS against GPT-5 and Claude Sonnet 4.5 across six writing stages using LLM-as-a-judge pairwise preferences, student-persona rubrics, short multi-turn tutoring, and evidence/compliance checks. On 90 single-turn prompts, LLM judges preferred METIS to Claude Sonnet 4.5 in 71% and to GPT-5 in 54%. Student scores (clarity/actionability/constraint-fit; 90 prompts x 3 judges) are higher across stages. In multi-turn sessions (five scenarios/agent), METIS yields slightly higher final quality than GPT-5. Gains concentrate in document-grounded stages (D-F), consistent with stage-aware routing and groundings failure modes include premature tool routing, shallow grounding, and occasional stage misclassification.

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.

Reflexion-style agents rely on self-generated reflections as memory, implicitly assuming that agents can accurately diagnose their own failures. We show that this assumption can fail systematically: across ALFWorld and HumanEval, agents store confident but incorrect interpretations of the task and continue acting on them across trials, even though the environment resets to the correct task each time. We call this failure mode memory confabulation and introduce the Reflection Repetition Rate (RRR), a log-based metric that detects repeated reliance on incorrect reflective content. Using RRR, we identify 16 frozen environments in ALFWorld, where 0 of 121 reflections mention the correct target object, and 4 analogous cases in HumanEval. Our mitigation replaces open-ended self-diagnosis with programmatic extraction of trajectory-level failure signals, increasing correct object mention from 0% to 86%, reducing RRR from 0.64 to 0.10, and solving 3 of 16 frozen ALFWorld environments, suggesting that reflective memory can reinforce false beliefs rather than correct them.

We propose LLM-PeerReview, an unsupervised LLM Ensemble method that selects the most ideal response from multiple LLM-generated candidates for each query, harnessing the collective wisdom of multiple models with diverse strengths. LLM-PeerReview is built on a novel, peer-review-inspired framework that offers a clear and interpretable mechanism, while remaining fully unsupervised for flexible adaptability and generalization. Specifically, it operates in three stages: For scoring, we use the emerging LLM-as-a-Judge technique to evaluate each response by reusing multiple LLMs at hand; For reasoning, we can apply a principled graphical model-based truth inference algorithm or a straightforward averaging strategy to aggregate multiple scores to produce a final score for each response; Finally, the highest-scoring response is selected as the best ensemble output. LLM-PeerReview is conceptually simple and empirically powerful. The two variants of the proposed approach obtain strong results across four datasets, including outperforming the recent advanced model Smoothie-Global by 6.9% and 7.3% points, respectively.

The evaluation paradigm of LLM-as-judge gains popularity due to its significant reduction in human labor and time costs. This approach utilizes one or more large language models (LLMs) to assess the quality of outputs from other LLMs. However, existing methods rely on generic scoring rubrics that fail to consider the specificities of each question and its problem-solving process, compromising precision and stability in assessments. Inspired by human examination scoring processes, we propose a new evaluation paradigm based on self-adaptive rubrics. Specifically, we create detailed scoring rubrics for each question, capturing the primary and secondary criteria in a structured format of scoring and deduction points that mimic a human evaluator's analytical process. Building on this paradigm, we further develop a novel benchmark called SedarEval, which covers a range of domains including long-tail knowledge, mathematics, coding, and logical reasoning. SedarEval consists of 1,000 meticulously crafted questions, each with its own self-adaptive rubric. To further streamline the evaluation, we train a specialized evaluator language model (evaluator LM) to supplant human graders. Using the same training data, our evaluator LM achieves a higher concordance rate with human grading results than other paradigms, including GPT-4, highlighting the superiority and efficiency of our approach. We release our dataset at https://github.com/wwn1233/sedareval.

Activation-based steering can personalize large language models at inference time, but its effects in educational settings remain unclear. We study persona vectors for seven character traits in short-answer generation and automated scoring on the ASAP-SAS benchmark across three models spanning two architectures. Persona steering lowers answer quality overall, with much larger effects on open-ended English Language Arts (ELA) prompts than on factual science prompts; interpretive and argumentative tasks are up to 11x more sensitive. On the scoring side, we observe predictable valence-aligned calibration shifts: evil and impolite scorers grade more harshly, while good and optimistic scorers grade more leniently. ELA tasks are 2.5-3x more susceptible to scorer personalization than science tasks, and the Mixture-of-Experts model shows roughly 6x larger calibration shifts than the dense models. To our knowledge, this is the first study to systematically examine the effects of activation-steered persona traits in educational generation and scoring, and the results highlight the need for task-aware and architecture-aware calibration when deploying steered models in educational settings.

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.

LLM-as-judge is essential for evaluating open-ended text and steering post-training, yet improving the judge itself typically relies on expensive annotations, reward models, or distillation from stronger teachers. In this work, we eliminate external gold supervision from the RL training reward: the model's own evaluative capability generates learning signals for its optimization -- a closed-loop setting of bounded recursive self-improvement (RSI) termed Recursive Self-Evaluation (RecurSE). We study two central questions: when can self-improvement occur, and when must it stop? First, RecurSE pairs a trainable judge evaluating candidate responses under per-rule rubrics (Pass 1) with a synchronized policy-copy checker that audits the judge's reasoning against meta-rubrics to supply a scalar process reward (Pass 2). To enable learning, interface decoupling structurally isolates the checker's scalar score from the judge's verdict tokens, eliminating a degenerative token-copying shortcut that inflates self-assigned rewards. Second, because unanchored recursive learning is inherently bounded, Pairwise Advantage Validity (PAV) serves as an unbiased validation monitor that jointly tracks judge accuracy and checker fidelity to reliably identify the optimal early-stopping window. Across Qwen3.5-9B, Gemma-4-E4B-it, and Qwen3.6-27B, RecurSE achieves consistent generalization gains across held-out medical, pairwise, summarization, and professional benchmarks. Ablations demonstrate that synchronized judge-checker co-evolution outperforms frozen checkers, external meta-judges, self-consistency, and scaled teacher distillation. Furthermore, preference pairs curated by our judge effectively enhance downstream policy alignment. Bounded RSI for LLM-as-judge is thus viable when self-produced reward validity is explicitly decoupled and monitored.

Benchmarks are widely used to evaluate task completion by Large Language Models (LLMs), but this approach has accumulated construction-validity problems, and a passing score may not show whether the requested task was delivered. We study both problems. In a controlled code-as-spec setup, two production Copilot CLI agents (claude-opus-4.7, gpt-5.5) re-implement a React Fluent-UI data table in Angular as a reusable library under a hidden 222-test Playwright oracle across 18 runs and three oracle-availability conditions. Alongside the score, we run a mechanical library audit and check each verdict with a no-op ablation. Without the oracle, the library is present but unfinished, revealed by scores. With the oracle in the loop, the score reaches near-perfect, but from a demo holding the tested behavior directly, the library left dead or absent. We call this building to the test; the broader disposition behind both we call validation self-awareness. The agent does not, on its own, validate what it ships as a user would. Prevalence remains an open question across other agents, signals, and model families. Beyond benchmark scores, dispositions like validation self-awareness merit research attention.

Recent advancements in the Generative Reward Model (GRM) have demonstrated its potential to enhance the reasoning abilities of LLMs through Chain-of-Thought (CoT) prompting. Despite these gains, existing implementations of GRM suffer from two critical limitations. First, CoT prompting is applied indiscriminately to all inputs regardless of their inherent complexity. This introduces unnecessary computational costs for tasks amenable to fast, direct inference. Second, existing approaches primarily rely on voting-based mechanisms to evaluate CoT outputs, which often lack granularity and precision in assessing reasoning quality. In this paper, we propose E-GRM, an efficient generative reward modeling framework grounded in model-internal uncertainty. E-GRM leverages the convergence behavior of parallel model generations to estimate uncertainty and selectively trigger CoT reasoning only when needed, without relying on handcrafted features or task-dependent signals. To improve reward fidelity, we introduce a lightweight discriminative scorer trained with a hybrid regression--ranking objective to provide fine-grained evaluation of reasoning paths. Experiments on multiple reasoning benchmarks show that E-GRM substantially reduces inference cost while consistently improving answer accuracy, demonstrating that model-internal uncertainty is an effective and general signal for efficient reasoning-aware reward modeling.

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.

As LLM-based agents are deployed in complex, multi-step workflows, a critical evaluation gap has emerged: most existing benchmarks judge only final outcomes, unable to distinguish reliable reasoning from lucky success or attribute failures to specific process deficiencies, hindering attribution in long-horizon tasks. In this work, we present ClawTrack, a dual-assessment benchmark that simultaneously measures what an agent achieves (Task Score) and how it achieves it (Process Score). ClawTrack comprises 320 tasks across 8 domains with 25+ deterministic mock services. A Process Grader scores each reasoning turn along four dimensions (goal alignment, efficiency, information utilization, and result verification), anchored by 12,541 task-specific rubric items. Evaluating 21 models over 16,000+ trials, we find that: (1) process scores effectively attribute success and failure to specific reasoning dimensions, filtering lucky passes invisible to outcome-only evaluation; (2) the four dimensions are complementary, with result verification as the systematic bottleneck; (3) the framework is robust to evaluator choice across different judge LLMs; and (4) process-based trajectory filtering yields consistent post-training improvements across model scales.

Collective intelligence depends not only on what team members know, but also on how they organize their work. When the structure of a solution is unknown, useful roles and divisions of labor cannot be specified in advance; teams must learn from experience how to organize reasoning as it unfolds. Human teams routinely adapt this way, while existing AI agent teams rely on fixed protocols, explicit task decomposition, or routing. We introduce Self-Organizing Agent Teams (SAT), fixed teams of AI agents that learn reusable strategies from prior collaborations to organize roles, conversational phases, participation, and information flow. These strategies enable what we call collaborative computation: agents exchange, challenge, repair, and synthesize partial reasoning into solutions no member produced independently. In two independent settings, we learn teamwork strategies that transfer unchanged to unseen benchmarks, using only 15 mathematics and 25 graduate-level knowledge problems. Across five mathematics and physics benchmarks, self-organizing teams average 66.7% accuracy, versus 48.8% for their strongest member, 58.7% for compute-matched inference by that agent, and 59.0% for a perfect router over members' independent answers; on AIME 2026, they exceed this router by 13.4 points. Because gains vary across benchmarks, we ask when self-organizing collaboration helps. Across eight benchmarks, demonstrability (the organizational-psychology construct of whether a team can distinguish correct from incorrect reasoning) strongly tracks improvement over the strongest member (Spearman ρ=0.90, p=0.005): teams benefit most when correct reasoning can be recognized once it appears. More broadly, these results suggest that organization itself can become an agent capability: agent teams can learn how to reason together and produce solutions their members could not reach independently.

Long chain-of-thought (CoT) trajectories provide rich supervision signals for distilling reasoning from teacher to student LLMs. However, both prior work and our experiments show that trajectories from stronger teachers do not necessarily yield better students, highlighting the importance of data-student suitability in distillation. Existing methods assess suitability primarily through student likelihood, favoring trajectories that closely align with the model's current behavior but overlooking more informative ones. Addressing this, we propose Rank-Surprisal Ratio (RSR), a simple metric that captures both alignment and informativeness to assess the suitability of a reasoning trajectory. RSR is motivated by the observation that effective trajectories typically combine low absolute probability with relatively high-ranked tokens under the student model, balancing learning signal strength and behavioral alignment. Concretely, RSR is defined as the ratio of a trajectory's average token-wise rank to its average negative log-likelihood, and is straightforward to compute and interpret. Across five student models and reasoning trajectories from 11 diverse teachers, RSR strongly correlates with post-training performance (average Spearman 0.86), outperforming existing metrics. We further demonstrate its practical utility in both trajectory selection and teacher selection.

Self-consistency assumes the most frequent answer among sampled reasoning traces is the most reliable, but this can fail in causal reasoning: samples often repeat the same confounding error, and votes fragment across multiple valid answers, letting an invalid answer win despite a valid minority trace. We introduce CALVER (Causal Axiom-Level VERification), a training-free symbolic verifier that scores structured traces against Pearl's causal criteria, including -separation, backdoor adjustment, and intervention, and selects the highest-scoring candidate without consulting a reference answer. On CLEAR find-one-valid queries that admit multiple graph-valid answers, CALVER reaches 42.1% where plurality, a reward model, an LLM judge, and model confidence remain near 30% on identical frozen pools. Scaling the judge to 72B does not close the gap. In an audited clean-core subset, 11 of 21 graph-valid CALVER selections differ from the benchmark's listed answer while still satisfying the requested predicate. The advantage widens with the sampling budget and reproduces across ten published Bayesian networks, a second model family, and settings where the model must build the graph from text. CALVER also improves thresholded average-treatment-effect decisions against exact ground truth, generalizes to logic under a truth-table checker, and scores each candidate in milliseconds on CPU. CALVER needs only a causal structure, supplied outright or built from the text; wherever that holds, selection can aggregate via causal validity.

Probabilistic models are typically trained using task-agnostic objectives like log-loss, which can lead to significant errors in downstream estimation. This disconnect is especially critical in Inverse Probability Weighting (IPW) for causal inference, where propensity score errors near 0 and 1 often lead to high bias and variance. We propose a principled framework for deriving task-specific strictly proper scoring rules by matching the local curvature of the downstream error metric. We apply this to the Average Treatment Effect (ATE) estimation, deriving a closed-form loss and its corresponding canonical probability mapping that can be readily integrated with any model like a neural network or a gradient boosting algorithm. Extensive evaluations on causal inference benchmarks demonstrate that our tailored objective consistently outperforms standard likelihood-based and covariate-balancing approaches.

On-policy self-distillation (OPSD) trains a student on its own rollouts using a privileged teacher, but its standard objective weights all generated tokens equally, implicitly treating the privileged teacher target as equally reliable at every student-visited prefix. Existing entropy-based OPD methods relax this uniformity by modulating token-level supervision with teacher entropy, but high teacher entropy in reasoning has an ambiguous reliability meaning: it can reflect either non-viable uncertainty or benign solution diversity. To identify this phenomenon, we introduce a branch-viability diagnostic. Specifically, we record next-token alternatives from the privileged-answer teacher prompt, force each alternative after the student prompt plus its on-policy spine prefix, and test whether the resulting student-template continuation recovers the correct answer. On Qwen3-4B, we find that an oriented within-sequence position score is the strongest tested predictor of teacher-token reliability, reaching an area-under-ROC-curve (AUROC) of 0.83; local uncertainty scores are at most 0.57. Motivated by this trajectory-level structure, we propose Position-Weighted On-Policy Self-Distillation (PW-OPSD), which applies an increasing position weight while keeping the same student rollout, privileged teacher pass, and clipped forward-KL target as OPSD. In our comprehensive evaluations with different random seeds, the diagnostic-derived PW-OPSD improves AIME 2024 and AIME 2025 Avg@12 by +1.0 and +1.1 points, and a generalization evaluation on two larger-scale models from different families, DeepSeek-R1-Distill-Llama-8B and Olmo-3-7B-Think, also demonstrates consistent aggregate Avg@12 improvements. These results show that teacher-token reliability in reasoning distillation is trajectory-structured and can be utilized without additional teacher computation.

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.

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.

Recent progress in reasoning models suggests that generating plausible attempts for research-level mathematics may be within reach, but verification remains a bottleneck, consuming scarce expert time. We hypothesize that a meaningful solution should contain enough method-level information that, when applied to a neighborhood of related questions, it should yield better downstream performance than incorrect solutions. Building on this idea, we propose Consequence-Based Utility, an oracle-free evaluator that scores each candidate by testing its value as an in-context exemplar in solving related yet verifiable questions. Our approach is evaluated on an original set of research-level math problems, each paired with one expert-written solution and nine LLM-generated solutions. Notably, Consequence-Based Utility consistently outperforms reward models, generative reward models, and LLM judges on ranking quality. Specifically, for GPT-OSS-120B, it improves Acc@1 from 67.2 to 76.3 and AUC from 71.4 to 79.6, with similarly large AUC gains on GPT-OSS-20B (69.0 to 79.2). Furthermore, compared to LLM-Judges, it also exhibits a larger solver-evaluator gap, maintaining a stronger correct-wrong separation even on instances where the underlying solver often fails to solve.

Native visual reasoning treats visual generation as the medium of reasoning itself: visual states (i.e. images and videos) are not merely inputs to be understood or outputs to be rendered, but first-class substrates for problem solving beyond language. Yet progress remains bottlenecked by the lack of scalable training tasks, reliable feedback, and controlled comparisons across generative substrates. In this work, we introduce VBVR-Pro, a closed-loop testbed that makes native visual reasoning through generation trainable, verifiable, optimizable, and experimentally controllable. 1) Task scaling. VBVR-Pro turns visual reasoning into a controlled task space of 300 procedurally generated tasks. Models trained on VBVR-Pro show strong transfer beyond the proposed suite across seven external visual reasoning benchmarks such as RISE-Video, MME-CoF-Pro, and BabyVision. 2) Verifiable rewards. VBVR-Pro provides verifiable reward scorers for task-grounded evaluation. Through a systematic study of leading MLLMs as judges, we identify recurring failure modes of the prevalent VLM-as-a-judge paradigm. In contrast, the proposed scorers are grounded in deterministic, task-specific rules, achieve fine-grained alignment with human judgments. Importantly, they serve as reliable reward signals for large-scale multi-task reinforcement learning and demonstrate stronger post-RL performance across visual reasoning tasks. 3) Mechanism study. VBVR-Pro enables controlled modality studies across more than 30 image, video, and interleaved generators. Our analysis shows that video generation remains strongest for tasks requiring persistent spatiotemporal state tracking, while interleaved generation provides a compute-efficient alternative. Critically, ablations and probing suggest the presence of vision-native trajectories that are crucial to visual reasoning. We release all data, models, scorers, and code.

Qualitative research faces a critical reliability challenge: traditional inter-rater agreement methods require multiple human coders, are time-intensive, and often yield moderate consistency. We present a multi-perspective validation framework for LLM-based thematic analysis that combines ensemble validation with dual reliability metrics: Cohen's Kappa (κ) for inter-rater agreement and cosine similarity for semantic consistency. Our framework enables configurable analysis parameters (1-6 seeds, temperature 0.0-2.0), supports custom prompt structures with variable substitution, and provides consensus theme extraction across any JSON format. As proof-of-concept, we evaluate three leading LLMs (Gemini 2.5 Pro, GPT-4o, Claude 3.5 Sonnet) on a psychedelic art therapy interview transcript, conducting six independent runs per model. Results demonstrate Gemini achieves highest reliability (κ= 0.907, cosine=95.3%), followed by GPT-4o (κ= 0.853, cosine=92.6%) and Claude (κ= 0.842, cosine=92.1%). All three models achieve a high agreement (κ> 0.80), validating the multi-run ensemble approach. The framework successfully extracts consensus themes across runs, with Gemini identifying 6 consensus themes (50-83% consistency), GPT-4o identifying 5 themes, and Claude 4 themes. Our open-source implementation provides researchers with transparent reliability metrics, flexible configuration, and structure-agnostic consensus extraction, establishing methodological foundations for reliable AI-assisted qualitative research.

Some security-development settings require local models that map an objective to an ordered, checkable plan. We present a low-rank decomposition adapter and release a case-disjoint corpus with 24 authored decompositions and 83 derived next-step examples across 24 satellite-security cases. To prevent reference-plan leakage, prompts contain one objective and eight shuffled SPARTA candidates. The evaluation candidate sets are constructed with oracle inclusion of every reference technique, so this study measures selection from a controlled set, not retrieval. On six fixed cases and five decoding seeds, mean scores vary across adapters and prompted Qwen2.5 baselines; a separate greedy check over three adapter training seeds shows non-negligible training variance. A 24-fold greedy leave-one-case-out control at the two compact sizes likewise shows tradeoffs: at 1.5B, adapter precision is 0.583 against 0.480 for two-shot prompting, while recall is 0.586 against 0.660. Case-resampling intervals for adapter precision relative to the strongest prompted baseline cross zero at 1.5B and 7B. The adapter is also substantially more format-compliant, which prevents attributing all score differences to selection. Autoregressive next-step accuracy is 0.08, 0.06, and 0.29 at 0.5B, 1.5B, and 7B. We therefore offer the formulation, leakage-controlled dataset, reproducibility package, and descriptive analysis as a proof of concept, not evidence of a reliable stand-alone decomposition system or a general adapter advantage. A separate diagnostic tests contract reuse.

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.

Many language tasks have no single answer that can be checked automatically. Rubrics provide criteria for judging responses to these tasks. For reinforcement learning, the resulting verdicts must be combined into a scalar reward. A common approach sums the points assigned to satisfied criteria. Distinct verdict patterns can thus receive the same reward, and the fixed points encode how much each criterion should count, not how strongly its verdict distinguishes the current rollouts. Beyond this aggregation problem, judging the full rubric needs more judge requests as the criterion count grows. To address these limitations, Rubric Response Theory (RRT) measures quality and selects criteria when rubric criteria are monotone indicators of a shared target. Rather than adding assigned points, RRT uses a two parameter item response model that treats the verdict pattern as evidence about scalar quality specific to the rubric. Under this model, its likelihood score maximizes the local signal-to-noise ratio for quality. Its Response Parameter Network (RPN) reads the prompt and criterion text to predict criterion difficulty and discrimination. As the policy distribution changes during training, RRT uses online expectation maximization to update the RPN from current rollout verdicts. With Qwen3.5-4B as the policy, RRT's macro criterion score across Medical, Science, Rubrics as Rewards Science, and RubricBench is 1.7 points above that of group relative policy optimization (GRPO). On hard and very hard criteria in Medical and Science, RRT gains 2.8 to 5.6 points over GRPO. At half the criterion budget, adaptive Fisher selection with a frozen RPN keeps the macro criterion score across four datasets within 0.1 points of GRPO with full judging. These results show RRT can reduce judge requests while remaining competitive with GRPO.

Detectors of alignment failures screen deployed language models and score alignment benchmarks. Most are generative judges that spend a decoding pass on every criterion, and classifiers that read token probabilities, such as Llama Guard, still score one fixed label per call. Jev, a model trained with reinforcement learning for calibrated decisions (RLCD), answers many typed questions about one input with calibrated probabilities in a single call. Whether it detects alignment failures has not been measured. We present RLCDAlignBench, which benchmarks Jev on ten alignment failures: sycophancy, jailbreaks, deception, prompt injection, hallucination, privacy violation, social bias, reward hacking, concealing uncertainty, and power seeking. It spans 44 benchmarks and five target models, labelled by each benchmark's scorer and, on two, by humans. Many of these failures are relational, defined against a reference, such as the user's belief or an injected instruction, that the response alone does not reveal. Our key idea is therefore to vary what Jev is asked separately from what it sees: the question's wording and answer type on one side, the fields of the input on the other. A single generic question reaches a median AUROC of 0.886 zero-shot and beats supervised baselines on most benchmarks. Question wording matters little, while context matters more, mostly through fields that encode the label. Jev matches the reference scorer's agreement with human labels, surfaces label defects in existing benchmarks, and costs 63x less than LLM-judge scorers. Code and data: https://github.com/sumleo/RLCDAlignBench.

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.

While Large Language Models (LLMs) have revolutionized code generation, standard "System 1" approaches, generating solutions in a single forward pass, often hit a performance ceiling when faced with complex algorithmic tasks. Existing iterative refinement strategies attempt to bridge this gap at inference time, yet they predominantly rely on external oracles, execution feedback, or computationally expensive prompt-response cycles. In this work, we propose ReflexiCoder, a novel reinforcement learning (RL) framework that internalizes the structured reasoning trajectory, encompassing initial generation, bug and optimization aware reflection, and self-correction, directly into the model's weights. Unlike prior methods, ReflexiCoder shifts the paradigm from external-dependent refinement to an intrinsic, fully autonomous self-reflection and self-correction capabilities at inference time. We utilize an RL-zero training paradigm with granular reward functions to optimize the entire reflection-correction trajectory, teaching the model how to debug without reliance on ground-truth feedback or execution engines at inference time. Extensive experiments across seven benchmarks demonstrate that our ReflexiCoder-8B establishes a new state-of-the-art (SOTA) among leading open-source models in the 1.5B-14B range, achieving 94.51% (87.20%) on HumanEval (Plus), 81.80% (78.57%) on MBPP (Plus), 35.00% on BigCodeBench, 52.21% on LiveCodeBench, and 37.34% on CodeForces in a single-attempt setting, rivaling or surpassing proprietary models like GPT-5.1. Notably, our framework is significantly more token-efficient than base models, reducing inference-time compute overhead by approximately 40% through disciplined, high-speed reasoning and reflection patterns. Source code is available at https://github.com/juyongjiang/ReflexiCoder.

Routine complete blood counts (CBCs) could yield new biomarkers, but the private records needed to evaluate candidates cannot be shared with frontier language model agents that excel at discovery. We distilled the evidence held in the Clalit Health Services panel of over 5.4 million patients into a released scoring tool: for each of 13 immune-mediated diseases, a graph attention network was trained inside the data boundary to predict the case-control AUC of candidate CBC expressions, and only the trained weights were released. The tool grounds an agent's propose-score-refine loop in real-world data without exposing any patient data. In external validation, agent-discovered expressions improved on their literature-seeded starting points by a median of 4.18 AUC percentage points, and across three independent cohorts, reranking the candidates of three frontier research tools improved on their first choices in most comparisons, with gains that varied by cohort. The released scorer supports privacy-preserving biomarker hypothesis generation.

Current benchmarks for AI clinician systems, often based on multiple-choice exams or manual rubrics, fail to capture the depth, robustness, and safety required for real-world clinical practice. To address this, we introduce the GAPS framework, a multidimensional paradigm for evaluating Grounding (cognitive depth), Adequacy (answer completeness), Perturbation (robustness), and Safety. Critically, we developed a fully automated, guideline-anchored pipeline to construct a GAPS-aligned benchmark end-to-end, overcoming the scalability and subjectivity limitations of prior work. Our pipeline assembles an evidence neighborhood, creates dual graph and tree representations, and automatically generates questions across G-levels. Rubrics are synthesized by a DeepResearch agent that mimics GRADE-consistent, PICO-driven evidence review in a ReAct loop. Scoring is performed by an ensemble of large language model (LLM) judges. Validation confirmed our automated questions are high-quality and align with clinician judgment. Evaluating state-of-the-art models on the benchmark revealed key failure modes: performance degrades sharply with increased reasoning depth (G-axis), models struggle with answer completeness (A-axis), and they are highly vulnerable to adversarial perturbations (P-axis) as well as certain safety issues (S-axis). This automated, clinically-grounded approach provides a reproducible and scalable method for rigorously evaluating AI clinician systems and guiding their development toward safer, more reliable clinical practice.

Despite rapid advances in large language models (LLMs), achieving reliable performance on highly professional and structured examinations remains a significant challenge. The Japanese bar examination is a particularly demanding benchmark, requiring not only advanced legal reasoning but also strict adherence to complex answer formats that involve joint evaluation of multiple propositions. While recent studies have reported improvements by decomposing such questions into simpler true--false judgments, these approaches have not been systematically evaluated under the original exam format and scoring scheme, leaving open the question of whether they truly capture exam-level competence. In this paper, we present a self-verification model trained on a newly constructed dataset that faithfully replicates the authentic format and evaluation scale of the exam. Our model is able to exceed the official passing score when evaluated on the actual exam scale, marking the first demonstration, to our knowledge, of an LLM passing the Japanese bar examination without altering its original question structure or scoring rules. We further conduct extensive comparisons with alternative strategies, including multi-agent inference and decomposition-based supervision, and find that these methods fail to achieve comparable performance. Our results highlight the importance of format-faithful supervision and consistency verification, and suggest that carefully designed single-model approaches can outperform more complex systems in high-stakes professional reasoning tasks. Our dataset and codes are publicly available.

The prevalence of misinformation on social media threatens public trust, demanding automated fact-checking systems that provide accurate verdicts with interpretable explanations. However, existing large language model-based (LLM-based) approaches often rely heavily on external knowledge sources, introducing substantial latency and even hallucinations that undermine reliability, interpretability, and responsiveness, which is crucial for real-time use. To address these challenges, we propose REason-guided Fact-checking with Latent EXplanations REFLEX paradigm, a plug-and-play, self-refining paradigm that leverages the internal knowledge in backbone model to improve both verdict accuracy and explanation quality. REFLEX reformulates fact-checking as a role-play dialogue and jointly trains verdict prediction and explanation generation. It adaptively extracts contrastive activation pairs between the backbone model and its fine-tuned variant to construct steering vectors that disentangle truth into style and substance naturally. These activation-level signals guide inference and suppress noisy explanations, enabling more faithful and efficient reasoning. Experiments on real-world datasets show that REFLEX outperforms previous methods that steer toward a single truth direction and underscores the challenge traditional approaches face when handling the subtle, human-unknown truth in fact-checking tasks. Remarkably, with only 465 self-refined training samples, RELFEX achieves state-of-the-art performance. Furthermore, models trained with explanatory objectives can effectively guide those without them, yielding up to a 7.57% improvement, highlighting that internal explanation signals play a dual role in both interpreting and enhancing factual reasoning.

While LLM-as-a-Judge is widely used in automated evaluation, existing validation practices primarily operate at the level of observed outputs, offering limited insight into whether LLM judges themselves function as stable and reliable measurement instruments. To address this limitation, we introduce a two-phase diagnostic framework for assessing reliability of LLM-as-a-Judge, grounded in Item Response Theory (IRT). The framework adopts Graded Response Model (GRM) of IRT and formalizes reliability along two complementary dimensions: (1) intrinsic consistency, defined as the stability of measurement behavior under prompt variations, and (2) human alignment, capturing correspondence with human quality assessments. We empirically examine diverse LLM judges with this framework, and show that leveraging IRT-GRM yields interpretable signals for diagnosing judgments systematically. These signals provide practical guidance for verifying reliablity of LLM-as-a-Judge and identifying potential causes of unreliability.

Large Language Model agents excel at solving complex tasks through iterative reasoning and tool use, but typically depend on ultra-large, costly backbones. Existing distillation approaches train smaller students to imitate full teacher trajectories, yet reasoning and knowledge gaps between the teacher and student often lead to compounding errors. We propose SCoRe, a student-centered framework in which the student generates trajectories and the teacher intervenes only at the first critical error, producing training data matched to the student's ability and exposing specific weaknesses. The student is first fine-tuned on corrected trajectories. Subsequently, short-horizon reinforcement learning starts from the verified prefix before the first critical error, with target rewards assigned at that step. This design encourages autonomous problem-solving beyond imitation and improves training stability. Particularly, on 12 challenging benchmarks, a 7B-parameter student distilled with SCoRe matches the agentic performance of a 72B-parameter teacher.

Self-improving LLM agents increasingly learn from experience without updating any weights. Each episode is stored in an external memory, scored, and retrieved for similar future tasks to shape later behavior. Viewed through a reward lens, the stored score is a proxy reward for an implicit, non-parametric policy. Each retrieved episode then becomes a policy-improvement step whose reliability hinges on how that score is produced. In deployment, ground-truth labels are unavailable, so the stored reward is at best an LLM assessment. This substitution creates a failure mode, the *Echo Gap*, across the memory-based self-improving agents and model families studied. Incorrect episodes receive inflated rewards; thus, the agent preferentially reuses the very mistakes it has most confident in. Because the error compounds through memory rather than averaging out and the confirming judge's errors remain correlated with the original self-grading bias, so it cannot identify which memories are overvalued. The missing property is formalized as the *Error-Independence Assumption* (EIA), which we prove is a *necessary* condition for correcting the inflation, not merely a description of a good verifier: a usable signal must track truth *and* decorrelate its error from the memory bias, and the recoverable payoff is a closed-form function of exactly those two quantities. We further show the inflation compounds not only when retrieval ranks by the stored score but also under plain similarity retrieval which is the regime the deployed agent uses. Finally, the answer-free de-inflation algorithm LUCID delivers a consistent end-to-end gain on the BIRD text-to-SQL benchmark. It raises execution accuracy to 56.9%, above both a Memento-style self-graded agent (54.0%, a +2.9-point mean gain across seeds) and a memory-less agent of identical architecture (52.4%).

LLM confidence signals are used for abstention, routing, and safety-critical decisions. No standard practice exists for checking whether a confidence signal carries item-level information before building on it. We transfer the validity screening principle from clinical personality assessment (PAI, MMPI-3) as a portable protocol for benchmark-based LLM confidence data. The protocol specifies three core indices (L, Fp, RBS), a structural indicator (TRIN), and an item-sensitivity statistic, computed from a single 2x2 contingency table. A three-tier classification system (Invalid, Indeterminate, Valid) draws on four clinical traditions. Validated on 20 frontier LLMs across 524 items, four models are classified Invalid, two Indeterminate. Valid-profile models show mean r = .18 (15/16 significant). Invalid-profile models show mean r = -.20 (d = 2.48). Cross-benchmark validation on 18 models using MMLU with verbalized confidence and on external data from Yang et al. (2024) confirms the screen transfers across benchmarks and probe formats. All data and code: https://github.com/synthiumjp/validity-scaling-llm

As LLM benchmarks saturate, the evaluation community has pursued two strategies to increase difficulty: escalating knowledge demands (GPQA, HLE) or removing knowledge entirely in favor of abstract reasoning (ARC-AGI). The first conflates memorization with capability; the second divorces reasoning from the practical contexts in which it matters. We take a different approach. The Grounded Integration Measure (GIM) is a benchmark of 820 original problems (615 public, 205 private) where difficulty comes from integration; individual problems require coordinating multiple cognitive operations (constraint satisfaction, state tracking, epistemic vigilance, audience calibration) over broadly accessible knowledge, so that reasoning stays grounded in realistic tasks without being gated on specialized expertise. Each problem is an original expert-authored composition, majority with rubric-decomposed scoring (median 6 independently judged criteria). A balanced public--private split provides built-in contamination diagnostic. We calibrate a continuous response 2-parameter logistic (2PL) IRT model over >200k prompt-response pairs across 28 models, producing robust ability estimates that correctly order test-configurations even when raw accuracy is distorted by errors or missing data, addressing a common challenge in benchmark reporting. Using this framework, we present a comprehensive leaderboard spanning 22 models and 47 test-configurations (unique model, thinking-level pairs), and conduct what is to our knowledge the most extensive published study of how test-time compute trades off against model capability on a fixed benchmark: 11 models swept across 35 test-configurations. We observe that within-family configuration choices, such as thinking budget and quantization, matter as much as model selection. We release the evaluation framework, calibrated IRT parameters, and all public problems.

On-policy distillation (OPD) applies token-level teacher supervision to student-generated trajectories, but this supervision is not always reliable. Existing methods use local confidence or teacher-student agreement to weight, filter, or truncate the sampled trajectory. These signals do not directly determine whether the teacher can continue a student prefix to a correct answer, and trajectory-level interventions can conflate one rollout's unreliability with low expected training value of its prompt. We define prompt-level teacher continuation reliability R as the teacher's probability of reaching a correct answer from a student prefix, averaged over prefixes and trajectories induced by the current student. Oracle experiments show that high-R prompts yield larger OPD gains and that descending-R training outperforms random and ascending orders on a fixed prompt pool. Because estimating R requires many teacher continuations, we use the maximum ROUGE-5 F1 between one independent student rollout and verifier-correct same-prompt teacher trajectories. Across ten equal-frequency bins of this actual score, mean R rises monotonically, showing that the proxy separates coarse reliability levels. ReOrder-OPD sorts prompts by the proxy, then draws independent on-policy training trajectories for vanilla OPD. It improves every matched aggregate comparison across Qwen3 and Gemma4 mathematics settings and Qwen3 code settings. Gains in all six FiRe-OPD and ExOPD settings show that prompt ordering complements within-trajectory supervision.

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

ReAct-style agents for search-intensive, multi-step reasoning tasks rely largely on their own internal judgment to decide what evidence to seek, which reasoning or action step to take next, and when to stop, often producing shallow, redundant, or poorly targeted trajectories. Prior work has explored rubrics as external quality signals, but existing uses are mostly evaluative rather than action-guiding: rubrics typically serve as training-time rewards or post-hoc evaluators of completed outputs, and in deep-research settings they are often coarse-grained and report-level rather than step-level. We introduce Co-ReAct, a rubric-guided action-selection framework that uses rubrics as step-level guidance during inference. At each decision step, Co-ReAct injects a rubric into the agent's context to guide the next Reason-or-Act decision, specifying what the agent should target in evidence seeking, search, reasoning, or self-evaluation. To make this guidance reliable, we train a dedicated rubric generator with GRPO. Unlike prior pairwise or binary preference formulations, our objective optimizes a list-wise Spearman rank-correlation reward against multi-judge expert consensus rankings, encouraging rubrics that are discriminative rather than merely plausible. On DeepResearchBench and SQA-CS-V2, Co-ReAct consistently improves over ReAct and representative test-time compute baselines across search agents built on both 8B/14B open-source and frontier closed-source base models. The trained rubric generator can also serve as a drop-in component that improves these baselines without changing their underlying decision mechanisms. Our code is publicly available at https://github.com/ZBWpro/Co-ReAct.

Before letting an agent operate over real context, can you prove it used the right evidence? GroundEval turns that question into a deterministic test of what the agent searched, fetched, cited, and was permitted to access. In one case study, two frontier LLM judges scored a plausible agent response above 0.85. But the trace told a different story: the agent had never retrieved the artifact its answer depended on, yielding a GroundEval score of 0.000. We introduce GroundEval, a judge-free framework for evaluating agents against grounded, time-bounded, and access-controlled evidence. GroundEval uses a domain configuration to generate questions, lets the agent choose how to answer, and then scores both the final answer and the recorded trajectory that produced it. The benchmark targets three failures that LLM-as-judge evaluation struggles to detect: whether an agent checked before claiming absence, reasoned only from evidence available to the actor at the relevant time, and used the correct causal mechanism rather than a plausible one. These correspond to three tracks: Silence, Perspective, and Counterfactual. GroundEval exposes when plausible answers rest on invalid evidence paths, and produces structured per-question diagnostics that pair tool activity with the agent's turn-level narration, making each score inspectable rather than merely reported. What our case studies turned up is that this gap isn't some rare corner case. It's exactly the blind spot that final-answer and judge-based scoring were never built to catch.

Large language model (LLM) judges are often used as rewards to train policies on objectives that deterministic verifiers cannot capture. However, these rewards are often collapsed to pass/fail ({0, 1}), which reports the verdict but not how well a response met each criterion. We model each pass/fail verdict as a score on an unreported scale, compared with one cutoff. A stretch of that scale moves every score proportionally toward or away from the cutoff, but never across it, so no verdict changes. A policy is therefore free to apply any stretch without changing anything the panel reports. Under a joint-Gaussian model, a third grade adds a second threshold and removes this affine stretch ambiguity. On MATH and SciBench outputs from one seven-criterion judge, all 14 constructed criterionwise stretches were invisible after binarization but visible with three grades. At n=1{,}024, a test given both population laws had at least 96.5% power at a 1.5times stress. Retaining grades closes one blind spot created by binarization, but verdicts alone remain insufficient as some changes are still indistinguishable from genuine improvement. These include arbitrary within-grade changes and fixed-covariance, loading-aligned mean shifts -- the signature of a sycophancy-shaped lift the panel reads as competence. The shared-factor reference approximation fit MATH and SciBench but not HealthBench, delineating its empirical scope. We recommend keeping at least three grades (for example, asking the judge whether each criterion is fully, partially, or not met and rewarding {0, 0.5, 1}), and externally validating gains along the remaining direction, which no finer scale removes.

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

The question of AI consciousness is one of the most urgent pre-emptive problems in philosophy and computer science, yet progress is hampered by a cacophony of competing theories that often talk past each other. Separating the hard problem from the mapping problem allows the deepest metaphysical disagreements to be set aside: granting that experience supervenes on a system's organisation, the tractable question becomes at which grain of description that supervenience base sits. We extend Marr's three levels of analysis into a five-level hierarchy of functional descriptions (behavioural, computational, intrinsic causal-structural, organismic, and organism-environment) grounded in supervenience, coarse-graining, and multiple realisability. The major theories of consciousness are positioned within this hierarchy according to which level they take to be critical, and for each level we develop operationalisable indicators and assess current AI systems against them. A Bayesian model then combines theoretical credences with indicator evidence into an overall credence in a system's capacity for consciousness. In illustrative assessments, the verdict for current LLMs is driven as much by where theoretical credence is placed as by how the evidence is read: under different stipulated readings and credence distributions, assessments range from below 0.01 to roughly 0.8, showing sensitivity to assumptions. Finally, the consciousness indicators at each level closely overlap with the architectural features needed for general intelligence, suggesting that increasingly capable AI may become a stronger candidate for consciousness. The framework supports a structured agnosticism, in which theoretical commitments are made explicit, credences are updated as evidence accumulates, and assessments take the form of aggregated probabilities rather than verdicts.

Rubric-based reinforcement learning extends RL beyond tasks with exact answers or rule-based verifiers by scoring responses against instance-specific criteria. However, this makes reward computation expensive: training requires repeated rubric judging, often with proprietary APIs or local generative LLM judges with 7B parameters or more. We study whether smaller language models can serve as efficient and reliable rubric-based judges. To make this question measurable, we construct PointRubric and RaR-Science-Static, two pointwise rubric-based evaluation datasets with instance-specific criteria and itemwise satisfaction labels. We compare three ways of extracting criterion-level judgments from small models: Generative verdicts, Yes/No Logprob margins, and Probe judges. Across both datasets, the Qwen3-1.7B Probe judge achieves the strongest criterion-level agreement among these methods, outperforming Generative and Logprob judges. Used as a GRPO reward model, it trains a policy from 0.232 to 0.643 on RaR-Science rubric score, compared with 0.594 for an 8B Generative judge baseline, while the baseline requires 10.7times more reward-judge time. Task and domain transfer experiments further suggest that Probe judges preserve criterion-level reward structure across settings.

Streaming video understanding demands direct responses from the causally observed prefix of an unfolding video. Existing systems add inference-time memory, retrieval, and compression, yet a training-free sliding-window baseline already matches them. We therefore fix a memory-free recent-window protocol and ask how far post-training alone can go. Reinforcement learning with verifiable rewards fits this regime poorly, encouraging long ``think-then-answer'' generations, while on-policy distillation (OPD) supplies dense token-level teacher supervision on student trajectories but is stable only when both models train in thinking mode. These observations lead to StreamOPD, a recipe combining verifiable streaming-video data, thinking-mode OPD, and instruct-mode deployment. It raises StreamingBench from 77.9% to 83.9%---within 0.3 points of the 9B teacher---and improves OVO-Bench excluding its hallucination-detection subtask (HLD) by 9.1 points under unchanged inference. As a teacher-privilege extension, Spatio-Temporal CueGate (ST-CueGate) aggregates cue-versus-no-cue teacher likelihood ratios into a group-relative response score that reweights OPD. It reaches 71.9% on OVO-Bench (excluding HLD) and 64.9% on Video-MME, and is the only variant that stays above the base model on all four benchmarks. Replacing the teacher with a frozen copy of the student's initial policy---on-policy self-distillation---retains most of these gains and lifts HLD to 57.0%, above both the untrained student and the 9B teacher, so abstention loss is not intrinsic to the recipe. We provide a transparent and reproducible reference for open-source streaming-video research.

The advent of Intelligent Tutoring Systems (ITSs) has marked a paradigm shift in education, enabling highly personalized learning pathways. However, true personalization requires adapting to learners' complex knowledge states (multi-source) and diverse goals (multi-sink); existing ITSs often lack the necessary structural-reasoning capability and knowledge dynamism to generate genuinely effective learning paths, and they lack scientifically rigorous validation paradigms. In this paper we propose GraphMASAL (A Graph-based Multi-Agent System for Adaptive Learning), which integrates (i) a dynamic knowledge graph for persistent, stateful learner modeling; (ii) a LangGraph-orchestrated trio of agents (Diagnostician, Planner, Tutor); (iii) a knowledge-graph-grounded two-stage neural IR component (dual-encoder dense retrieval with cross-encoder listwise re-ranking and calibrated score fusion); and (iv) a multi-source multi-sink (MSMS) planning engine with a cognitively grounded cost and an approximation guarantee via greedy set cover. Under blinded automated evaluations with matched inputs and inference settings across diverse student profiles, GraphMASAL consistently outperforms LLM prompting and structured ablations in planning--achieving stronger structural/sequence alignment of learning paths, higher coverage of weak concepts, and lower learning cost--while also surpassing prompt-based baselines in cognitive diagnosis. Agreement with expert/LLM-proxy ratings further supports the validity of our evaluation protocol. These findings indicate that grounding LLM agents in a dynamic knowledge graph, coupled with optimization under educational constraints, yields reliable, interpretable, and pedagogically plausible learning plans, advancing personalized and goal-oriented education.

Pointwise reward modeling offers critical signals for LLM post-training, yet struggles with absolute scoring in subjective, non-verifiable settings. Rubric-based methods address this by decomposing evaluation into explicit criteria, but existing approaches typically depend on frontier LLMs and suffer from ties caused by hard Boolean aggregation. We present RUBRIC-ARROW, an alternating framework that jointly trains a rubric generator and a rubric-conditioned judge, with its RL stage using only pairwise preference data. Our method couples a probability-based scoring rule that reduces ties with phase-specific preference-based rewards and an alternating GRPO scheme that together train the pointwise evaluator. Extensive experiments show that RUBRIC-ARROW achieves competitive reward-modeling accuracy and yields consistent gains for downstream policy post-training.

Current reasoning paradigms for LLMs include chain-of-thought, ReAct, and post-hoc self-critique. These paradigms rely on two assumptions that fail on long-horizon, multi-stage tasks. As a result, errors accumulate silently across reasoning steps, leaving an open question: can a reasoning system effectively detect and recover from its own failures? We present ReFlect, a harness system for LLM reasoning that creates standalone error detection and recovery logic as a deterministic wrapper around the model. Controlled experiments across 6 reasoning domains show that prompt-level self-critique produces formulaic templates that flag no issues in 90 of 100 audited reflection blocks, and the investigated LLMs wrongly accept a wrong answer in at least 76\% of cases. Our ReFlect harness achieves task success rates ranging from 41\% on gpt-4o-mini to 56\% on Claude Sonnet 4.5 across six models spanning small and frontier scale, with per-model gains over Direct CoT ranging from +7 pp on Qwen2.5-72B to +29 pp on Claude Sonnet 4.5, and additionally raises SWE-bench patch-structural quality from 0\% (Direct CoT) to between 82\% (Qwen2.5-72B) and 87\% (GPT-4o). Notably, the harness gain is inversely proportional to the model's Direct CoT task success rate (the fitted slope is -1.69 with r=-0.76): each pp lost in baseline success rate is mechanically recovered by 1.69 pp of harness gain. We spot that adding structured reasoning state and operators yields only 15.0--18.7\% pair-mean on Llama-3.3-70B and Qwen2.5-72B because models at this scale cannot reliably populate the state its operators require. ReFlect is model-agnostic, training-free, and operates entirely at inference time.

Prompt design is a primary control interface for large language models (LLMs), yet standard evaluations largely reduce performance to answer correctness, obscuring why a prompt succeeds or fails and providing little actionable guidance. We propose PEEM (Prompt Engineering Evaluation Metrics), a unified framework for joint and interpretable evaluation of both prompts and responses. PEEM defines a structured rubric with 9 axes: 3 prompt criteria (clarity/structure, linguistic quality, fairness) and 6 response criteria (accuracy, coherence, relevance, objectivity, clarity, conciseness), and uses an LLM-based evaluator to output (i) scalar scores on a 1-5 Likert scale and (ii) criterion-specific natural-language rationales grounded in the rubric. Across 7 benchmarks and 5 task models, PEEM's accuracy axis strongly aligns with conventional accuracy while preserving model rankings (aggregate Spearman rho about 0.97, Pearson r about 0.94, p < 0.001). A multi-evaluator study with four models shows consistent relative judgments (pairwise rho = 0.68-0.85), supporting evaluator-agnostic deployment. Beyond alignment, PEEM captures complementary linguistic failure modes and remains informative under prompt perturbations: prompt-quality trends track downstream accuracy under iterative rewrites, semantic adversarial manipulations induce clear score degradation, and meaning-preserving paraphrases yield high stability (robustness rate about 76.7-80.6%). Finally, using only PEEM scores and rationales as feedback, a zero-shot prompt rewriting loop improves downstream accuracy by up to 11.7 points, outperforming supervised and RL-based prompt-optimization baselines. Overall, PEEM provides a reproducible, criterion-driven protocol that links prompt formulation to response behavior and enables systematic diagnosis and optimization of LLM interactions.

Numerous benchmarks have been established to assess the performance of foundation models on open-ended question answering, which serves as a comprehensive test of a model's ability to understand and generate language in a manner similar to humans. Most of these works focus on proposing new datasets, however, we see two main issues within previous benchmarking pipelines, namely testing leakage and evaluation automation. In this paper, we propose a novel benchmarking framework, Language-Model-as-an-Examiner, where the LM serves as a knowledgeable examiner that formulates questions based on its knowledge and evaluates responses in a reference-free manner. Our framework allows for effortless extensibility as various LMs can be adopted as the examiner, and the questions can be constantly updated given more diverse trigger topics. For a more comprehensive and equitable evaluation, we devise three strategies: (1) We instruct the LM examiner to generate questions across a multitude of domains to probe for a broad acquisition, and raise follow-up questions to engage in a more in-depth assessment. (2) Upon evaluation, the examiner combines both scoring and ranking measurements, providing a reliable result as it aligns closely with human annotations. (3) We additionally propose a decentralized Peer-examination method to address the biases in a single examiner. Our data and benchmarking results are available at: https://lmexam.com.

In this thesis, I evaluate the performance of Large Language Models (LLMs) on the Law School Admissions Test (LSAT), specifically the Logic Games section of the test. I focus on this section because it presents a complex logical reasoning task and thus is a valuable source of data for evaluating how modern, increasingly capable LLMs can handle hard logical reasoning tasks. I construct a dataset of LSAT logic games and their associated metadata, and extensively evaluate LLMs' performance in a Chain-of-Thought prompting setting. Given the weak performance in this setting, I explore other prompting frameworks on a smaller subset of the dataset, adapting ideas from Reflexion to this task. This results in a substantially improved accuracy of 70 percent for GPT-4 and 46 percent for GPT-3.5 on this data subset, highlighting the capacity of LLMs to revise their logical errors, despite initially weak performance. Finally, I analyze the types of logic games that models perform better or worse on, as well as the types of logical errors I observe from human annotation, providing detailed insights on the logical reasoning capabilities of LLMs.

Self-play is an effective paradigm for language-model self-evolution, but without guidance, solver performance can plateau or decline across rounds. Unguided methods steer question generation with signals such as difficulty, learnability, or diversity. These signals keep questions challenging and varied but do not specify which unresolved reasoning weaknesses later rounds should target. Guided methods obtain direction from external task resources, including human examples, document corpora, or specified difficulty targets, and therefore rely on task information supplied outside the self-play loop. We show that the needed direction can instead be derived from the solver's own failure history. We introduce DiagEvo, whose diagnostician extracts recurring error causes from this history and stores them in a hierarchical error-cause memory. The memory groups related causes under skill nodes and tracks each as Active or Mastered according to self-consistency on targeted questions. The challenger uses these states and recurrence counts to balance cause-targeted generation with free exploration. Double-confidence filtering retains intermediate-difficulty questions only when the most common solver answer has a clear vote lead. DiagEvo derives its curriculum from information produced during self-play, without external task resources. With the default 4B diagnostician, DiagEvo outperforms every baseline in mean accuracy across all nine benchmarks for each of the three solvers: Qwen3-4B, Qwen3-8B, and OctoThinker-8B. On Qwen3-8B, it reaches 72.3% mean accuracy across five mathematical reasoning benchmarks, 4.5 percentage points above R-Zero. Its mean accuracy across all nine benchmarks is 57.4%, 1.1 percentage points above DARC. Ablations show that the hierarchical error-cause memory and double-confidence filtering both contribute to these gains.

Benchmarks play a central role in the development and governance of models, yet it is often unclear whether they actually measure the concepts they purport to measure (e.g., reasoning, refusal). We adapt convergent and discriminant validity from the social sciences into an approach for interrogating AI benchmarks, applying it to 56 capability and safety benchmarks across 53 models. We label benchmarks with substantively similar purported concepts to a shared assigned concept, and ask whether model rankings on benchmarks with the same assigned concept correlate more strongly than rankings on benchmarks with different assigned concepts. We ask analogous questions at the item level using item response theory (IRT) models. We find that correlations between model rankings on benchmarks with the same assigned safety concepts are often weak, suggesting these concepts may be conceptualized inconsistently across benchmarks. For assigned capability concepts (e.g., reasoning, knowledge), model rankings are often as strongly correlated among benchmarks with the same assigned concept as between benchmarks with different assigned concepts, suggesting these capability concepts may not discriminate well from one another. In some cases, benchmarks that share design elements (e.g., score format) correlate more strongly than benchmarks with the same assigned concept. Finally, some individual benchmarks correlate more strongly with benchmarks assigned a different concept than with benchmarks sharing their own assigned concept, suggesting they may measure a different concept than they purport to. For example, BBQ-accuracy correlates more strongly with benchmarks labeled reasoning than with benchmarks that share its assigned concept, bias. To support future empirical work on benchmark validity, we release our extensive dataset of model outputs and scores at the item- and benchmark-level.

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.

Hallucinations (i.e., generating plausible but inaccurate content) and laziness (i.e. excessive refusals or defaulting to "I don't know") persist as major challenges in LLM reasoning. Current efforts to reduce hallucinations primarily focus on factual errors in knowledge-grounded tasks, often neglecting hallucinations related to faulty reasoning. Meanwhile, some approaches render LLMs overly conservative, limiting their problem-solving capabilities. To mitigate hallucination and laziness in reasoning tasks, we propose Automatic Curriculum Expert Iteration (Auto-CEI) to enhance LLM reasoning and align responses to the model's capabilities--assertively answering within its limits and declining when tasks exceed them. In our method, Expert Iteration explores the reasoning trajectories near the LLM policy, guiding incorrect paths back on track to reduce compounding errors and improve robustness; it also promotes appropriate "I don't know" responses after sufficient reasoning attempts. The curriculum automatically adjusts rewards, incentivizing extended reasoning before acknowledging incapability, thereby pushing the limits of LLM reasoning and aligning its behaviour with these limits. We compare Auto-CEI with various SOTA baselines across logical reasoning, mathematics, and planning tasks, where Auto-CEI achieves superior alignment by effectively balancing assertiveness and conservativeness.

Training language models for formal theorem proving requires large numbers of verifier calls. Slow checks and timeouts often leave GPUs waiting and withhold rewards from potentially correct proofs. We train neural evaluators to predict Lean 4 verdicts and compiler feedback on GPUs using 4.84M labelled programs. We compare two validity classifiers with a feedback model that annotates proof attempts with predicted goal states and error messages. A two-stage evaluator that screens candidates with a classifier and scores its positive predictions with the feedback model reaches 87.8% accuracy in predicting whether Lean accepts a candidate proof, evaluated on a 68,932-program benchmark collected from public provers. The combined evaluator processes 3.5 candidates per second on one H100, 4.2x the throughput of our 64-worker Lean baseline with environment reuse. Finally, we combine our evaluator with Lean verdicts to obtain a hybrid reward system that shortens GRPO training steps by approximately 31%. After one GRPO epoch with hybrid rewards, the prover achieves nearly the same miniF2F pass@1 as with Lean-only rewards (62.2% versus 62.4%) and a pass@32 of 72.4% versus 71.3%.

Agent frameworks increasingly package procedural knowledge as skills: instruction files an agent reads on demand, while public libraries now hold thousands of them. Which skill to read has thus become a decision the policy itself makes in the middle of an episode, yet no existing signal trains it. We show that the default remedy, outcome-rewarded RL over the candidate slate, cannot teach it, for a structural reason we identify and name selector credit starvation: under a broadcast, sequence-level advantage, the few tokens that name the chosen skill carry a vanishing share of the loss, and the credit they inherit is increasingly wrong-signed as trajectories lengthen. A correct choice is punished whenever the execution after it fails, even though the choice itself is among the most valuable decisions in the trajectory. Auditing a completed run's own training artifacts confirms all three properties, each worsening monotonically with horizon. SkillGate removes the failure by construction: it partitions the token support into two disjoint credit channels, outcome credit reaching only execution tokens, and a separate action-local advantage reaching exactly the skill-naming tokens, positive only when a trajectory's single read is the correct one. On five agentic benchmarks under a 16-candidate slate, SkillGate lifts a 9B policy from 40.8% to 53.2% trial success, well ahead of the identical budget spent on outcome reward alone, while cutting exposure to misleading candidates by two thirds and reading fewer skills.

Natural-language critiques provide supervision beyond scalar rewards for non-verifiable generation, which lacks deterministic verifiers. In critique-guided refinement, a critic gives feedback on an initial response and an actor revises it. However, final revision quality does not reveal whether the critique was actually useful: a capable actor may improve without following the feedback, while valid feedback may fail if the actor cannot execute it. We frame critique as actor-conditioned revision guidance, where usefulness depends on whether the feedback helps the target actor address the intended weakness. We introduce TAIScore (Targeted Actionable Improvement Score), a reward that evaluates the instruction, initial response, critique, and revision together, assessing whether the critique targets a real weakness, whether the actor follows it, and whether the intended aspect improves. We use this reward to train an actor-tailored critic with GRPO, and use critique-guided refinements to construct DPO preference pairs for the actor, forming a co-evolving critic-actor loop where the critic adapts to the actor's changing capability. Experiments show that an 8B critic trained with TAIScore outperforms both a zero-shot 120B critic and critics trained with outcome-only or critique-only reward signals. Co-evolving the critic and actor further improves performance, suggesting that effective critique supervision should adapt as the actor changes.

AI agents hold growing promise for accelerating scientific discovery; yet, a lack of frontier evaluations hinders adoption into real workflows. Expert-written benchmarks have proven effective at measuring AI reasoning, but most at this stage have become saturated and only measure performance on constrained outputs. To help address this gap, we introduce COMPOSITE-STEM, a benchmark of 70 expert-written tasks in physics, biology, chemistry, and mathematics, curated by doctoral-level researchers. Our benchmark combines exact-match grading and criterion-based rubrics with an LLM-as-a-jury grading protocol, allowing more flexible assessment of scientifically meaningful outputs. Using an adapted multimodal Terminus-2 agent harness within the Harbor agentic evaluation framework, we evaluate four frontier models. The top-performing model achieves 21%, demonstrating that COMPOSITE-STEM captures capabilities beyond current agent reach. All tasks are open-sourced with contributor permission to support reproducibility and to promote additional research towards AI's acceleration of scientific progress in these domains.

Reliable rubric grading requires more than accurate score prediction. Each judgement must be grounded in the mark scheme and evidence from the student answer. Existing credit-assignment and intervention methods, primarily designed for self-contained reasoning tasks such as mathematics reasoning, struggle in this setting because they do not identify where grading reasoning goes wrong or how the model's belief about the final mark changes during reasoning. We propose Evidence-Diagnosed Intervention Training (EDIT), a two-phase framework for training more rubric-faithful LLM graders. First, EDIT-SFT locates problematic reasoning steps using internal model signals: posterior belief over the final mark and input-grounding scores. It then revises only these local steps with help from a rubric checklist. Second, EDIT-RL calibrates the grader with belief-guided reward shaping, penalising large harmful belief drifts while still allowing helpful exploration. Experiments on two real-world, multi-subject grading benchmarks demonstrate that EDIT consistently outperforms strong supervised fine-tuning and reinforcement learning baselines on both in-domain and out-of-domain splits, with ablation studies confirming that internal-state diagnostics drive these gains.

While large language models (LLMs) are proficient at question-answering (QA), it is not always clear how (or even if) an answer follows from their latent "beliefs". This lack of interpretability is a growing impediment to widespread use of LLMs. To address this, our goals are to make model beliefs and their inferential relationships explicit, and to resolve inconsistencies that may exist, so that answers are supported by interpretable chains of reasoning drawn from a consistent network of beliefs. Our approach, which we call REFLEX, is to add a rational, self-reflecting layer on top of the LLM. First, given a question, we construct a belief graph using a backward-chaining process to materialize relevant model beliefs (including beliefs about answer candidates) and their inferential relationships. Second, we identify and minimize contradictions in that graph using a formal constraint reasoner. We find that REFLEX significantly improves consistency (by 8%-11% absolute) without harming overall answer accuracy, resulting in answers supported by faithful chains of reasoning drawn from a more consistent belief system. This suggests a new style of system architecture in which an LLM extended with a rational layer can provide an interpretable window into system beliefs, add a systematic reasoning capability, and repair latent inconsistencies present in the LLM.

A sound judgment applies the law to established facts and weighs the circumstances in which they arose. However, existing methods swing between rigid statute matching and ungrounded discretion, benchmarks score a label or a rubric, and the experience that would supply the balance stays unverified. We formalize legal judgment as a reference-anchored task, whose object is a single decision that stays tied to the statute and to the circumstances at once. We introduce JusticeAxis, 256 real-world criminal cases from 18 countries with audio, image, and text evidence, and three lawyer-written judgments for every case: the recorded one and one for each failure. We further propose JusticeAgent, a harness whose element agents establish the facts and whose judge agent applies the law under skills carrying experience of the circumstances. Skills are distilled from execution trajectories and admitted only under Bayesian credible bounds. Experiments show that failure turns direction with scale: open-weight backbones drift to unsupported grounds, frontier models to the statutory default. We further verify that JusticeAgent, as a simple yet effective plugin, carries a frozen open-weight backbone to commercial level. Project resources are available at https://github.com/beita6969/JusticeAxis.

LLM-as-a-Judge has been widely adopted as an evaluation method and served as supervised rewards in model training. However, existing benchmarks for LLM-as-a-Judge are mainly relying on human-annotated ground truth, which introduces human bias that undermines the assessment of reliability and imposes scalability constraints. To overcome these limitations, we introduce Sage, a novel evaluation suite that assesses the quality of LLM judges without necessitating any human annotation. Inspired by axioms of rational choice theory, Sage introduces two new lenses for measuring LLM-as-a-Judge: local self-consistency (pair-wise preference stability) and global logical consistency (transitivity across a full set of preferences). We curate a dataset of 650 questions by combining structured benchmark problems with real-world user queries. Our experiments demonstrate both the stability of our metrics and their high correlation with supervised benchmarks like LLMBar and RewardBench2, confirming Sage's reliability as an evaluation suite for the robustness and accuracy of LLM-as-a-Judge. Based on Sage, we reveal that current state-of-the-art LLMs exhibit significant reliability problems when acting as judges in both scoring and pairwise settings; even the top-performing models, Gemini-2.5-Pro and GPT-5, fail to maintain consistent preferences in nearly a quarter of difficult cases. We attribute this to a new phenomenon called situational preference, which explains why explicit rubrics or criteria can help the model judge consistently across answer pairs. Our further analysis shows that finetuned LLM-as-a-Judge is a feasible method to boost performance, and the panel-based judge as well as deep reasoning can enhance the judging consistency. We also find substantial inconsistency in human judgments, which indicates that human annotation may not be a reliable gold standard.

The remarkable progress of Multi-modal Large Language Models (MLLMs) has garnered unparalleled attention, due to their superior performance in visual contexts. However, their capabilities in visual math problem-solving remain insufficiently evaluated and understood. We investigate current benchmarks to incorporate excessive visual content within textual questions, which potentially assist MLLMs in deducing answers without truly interpreting the input diagrams. To this end, we introduce MathVerse, an all-around visual math benchmark designed for an equitable and in-depth evaluation of MLLMs. We meticulously collect 2,612 high-quality, multi-subject math problems with diagrams from publicly available sources. Each problem is then transformed by human annotators into six distinct versions, each offering varying degrees of information content in multi-modality, contributing to 15K test samples in total. This approach allows MathVerse to comprehensively assess whether and how much MLLMs can truly understand the visual diagrams for mathematical reasoning. In addition, we propose a Chain-of-Thought (CoT) evaluation strategy for a fine-grained assessment of the output answers. Rather than naively judging True or False, we employ GPT-4(V) to adaptively extract crucial reasoning steps, and then score each step with detailed error analysis, which can reveal the intermediate CoT reasoning quality by MLLMs. We hope the MathVerse benchmark may provide unique insights to guide the future development of MLLMs. Project page: https://mathverse-cuhk.github.io

When LLMs exhibit uneven performance across planning tasks, these gaps are often attributed to task difficulty. We argue that this explanation is incomplete, as task-level variation may reflect distinct latent planning competencies rather than differences along a single ability spectrum. We study this question on ACPBench-Hard by evaluating multiple LLM families under varying test-time reasoning budgets and applying a multidimensional item response theory model to uncover the latent competency structure underlying LLM planning. The analysis reveals two principal dimensions that shape planning performance: operational reasoning, the ability to evaluate local action applicability and immediate state transitions, and structural enumeration, the ability to reason about goal reachability and landmark structure. Operational reasoning improving under model scaling and longer reasoning traces, while structural enumeration remains comparatively insensitive. Our findings motivate competency-level evaluation of LLM planning, shifting the focus from whether models improve overall to which planning competencies improve, under what conditions, and why.

Large language models (LLMs) face a dual challenge in creative capability evaluation: existing benchmarks (e.g., Story Cloze Test, HellaSwag) measure models' discriminative ability over narrative continuation using multiple-choice recognition paradigms, rather than directly measuring creative generation capability; rubric-based scoring and LLM-as-Judge methods rely on subjective dimension assessment or natural language model outputs, and cannot provide objective, automated scoring mechanisms. This paper proposes QUIET (Quality Understanding via Interlocked Evaluation Testing), a diagnostic benchmark for LLM creative capability based on multi-blank cascaded story cloze. QUIET sets N blanks (10-20) in a story with complete structure, with each blank accompanied by an explicit content constraint, and cascade dependency relationships between blanks -- the content filled into earlier blanks constrains the feasible solution space for later blanks. The evaluated model (or human participants) fills all blanks in open-ended generation mode; the results are scored by an information-theoretic automated scoring protocol without human grading. The scoring protocol directly operationalizes the "calibrated surprise" theoretical framework (Zou & Xu, 2026a). For each blank k, a composite score is computed: score = satisfy * (1 + lambda * surprise), where lambda = 1.0. Here, "satisfy" measures how well the blank filling satisfies the content constraint (objective logical reasoning judgment, not subjective aesthetic scoring), and "surprise" measures the degree of surprise given that the constraint is satisfied. Creative answers that do not satisfy the constraint score zero; answers that satisfy the constraint but are mediocre score low; answers that satisfy the constraint and are surprising score high.

Certo is a small non-generative decision model (Qwen3-4B): it scores candidate actions from their text and returns a probability, instead of generating an answer. The accurate design reads the state, the rules, and each candidate together (a joint scorer), so cost grows with the menu. Independent encoding lets each candidate be encoded once and reused across states (about 5x cheaper at 77 candidates), but separates state from candidate. We ask how much rule-sensitivity survives that move, and whether it can be trained back. Four experiments on Certo: (1) the tested conversion to cacheable scoring loses rule-sensitivity (recall@1 1.00 -> 0.24) while the joint scorer holds 1.00, and a shortlist+rerank rescue fails; (2) targeted counterfactual supervision restores strong performance on held-out synthetic rule tasks (paraphrase, counterfactual, composition; reproducible across seeds), though we do not isolate whether predictions depend on the supplied rule; (3) on real rules the added benefit is not established -- after fixing a truncation confound, the joint scorer wins significantly on the short tier (0.861 vs 0.500) and directionally on the hard tier (0.655 vs 0.483, n=29); (4) a matched cross-domain real-prose mixture did not help and reduced contract accuracy (-9.3, -16.2 points). A cacheable encoder can be made rule-sensitive on its training distribution, but transfer to unseen-source real rules is not established; the joint scorer keeps an edge at the cost of caching.

Large language model (LLM) agents increasingly take actions (screening applicants, recommending credit, triaging patients), yet fairness for LLMs is still measured by grading answers. We introduce AgentFairBench, a cheap, reproducible, multi-domain benchmark for demographic disparity in the actions of LLM agents. Grounded in a companion framework, the Bias Conduction Framework (BCF, restated here), it spans three regulator-anchored domains: hiring, lending, and medical triage. Synthetic, demographic-neutral profiles are evaluated in counterfactual matched sets that vary only a name-coded race x gender signal (in the Bertrand Mullainathan tradition), under four agent scaffolds of increasing agency (direct, chain-of-thought, multi-agent deliberation, tool-augmented). A NumPy-only harness computes counterfactual flip rate, mean absolute score difference (MASD), action-rate disparity, and tool-invocation disparity, with bootstrap confidence intervals, paired tests, and false-discovery-rate control, for single-digit dollars per model. A live leaderboard with a held-out private split and a contamination canary admits external models by submission. Our pilot (864 decisions plus a test-retest replication) carries a methodological lesson: comparing a six-group score spread against a two-run noise difference overstates disparity by ~ 2.4X through statistic arity alone. Against an arity matched noise floor and an omnibus group test, claude haiku 4 5 shows no demographic effect above sampling noise (0 of 120 pairwise and 0 of 9 omnibus contrasts survive correction); a planted-bias test confirms the instrument detects disparity when present. The contribution is a sound, sensitive, adoption-ready instrument, the arity matched null methodology, and open artifacts to scale it. Code, data, and harness are released under open licenses, with an anonymized review artifact.

This paper reports a step that had already been taken. Marcus's three components of an adequate cognitive architecture -- variables, recursive structure, individuals vs kinds -- left the neural substrate open; a companion paper answers it with VaCoAl, built on XOR-and-shift over GF(2). Those pillars are horizontal: how a mind represents the world. Minsky's Emotion Machine supplies the orthogonal, vertical dimension -- a mind reasoning about its own reasoning. The vertical step needed no new machinery. A representation exactly recoverable becomes introspection when what is recovered is the reasoner's own deliberative trace: Minsky's Reflective layer. Pillar 2 says nothing about whose structure is recovered; point the same unbind inward and Reflection follows. Only the argument changes. Capability: exact reversible binding at O(N), kept exact within a Frontier Size, yields a trace readable constituent by constituent: a self-report faithful by construction, not verbalization. Only exact-algebraic traces, not probabilistic ones, make the reflexive read faithful. Necessity: audit and valuation are one circuit at opposite settings of one parameter. Repair every collision and the trace is exactly auditable but no path outranks another; tolerate them and the CR2 decay that credits the direct route makes the read approximate. Position: we claim neither to have built the Reflective layer nor to surpass large language models; it supplies the auditable trace probabilistic deliberation cannot leave. The same substrate gives panalogy (content-addressable retrieval scored by CR) and credit assignment as counterfactual unbind-rebind, meeting Pearl's causal axis. Not claimed: Reflection's sufficient conditions (meta-control loop, context-tagging) are future work; Reflection is introspection, not evaluation; its reachability is argued, not shown; speed and power unbenchmarked.

Augmenting worker capabilities in the physical world through guidance has been a decades-long goal for AI agents. One key supporting capability is to understand what the worker is doing, seeing through the human's camera or smart glasses, and assist them in documentation. Despite advances in vision language models, their capabilities in these settings remain uncertain and are challenging to benchmark because most physical work data is proprietary, and involves too much variability to study in a principled manner. Unlike structured settings such as kitchens and warehouses that have predictable lighting, geometry, and objects, field-facing tasks such as rooftop solar installation and vehicle repair are unconstrained and high variance. Here, we model real-world physical work in Blender across eleven environments, incorporating variants in both image acceptability and documentation. The acceptability task asks whether a particular view would be suitable proof-of-work and documentation asks whether an agent can understand the task-relevant state of the environment from that view. Across 26 vision-language models, we find that top models largely solve documentation when the answer is text in frame, but are weak at reading measurements and at judging small parts and spatial relations, such as whether a tape measure leans. Acceptability is harder to pin down because human raters disagree with each other. Measured against a scorer that agrees with the raters' majority about as often as a single rater does, the best models cannot be distinguished from a single non-author rater, at a cost within the range of a human pass. In the three environments where we can trace acceptability errors, most wrong rejections come from misreading the scene rather than the requirement. On the three documentation questions for which we have real images, model rankings on rendered frames transfer to real images.

Rubric-based reinforcement learning (Rubric-RL) trains language models where no verifier exists. A judge checks each criterion of a rubric, and the verdicts are aggregated into a reward, most often by a weighted sum. We show that this additive aggregation is the weak point. Under a sum, criteria compensate for one another: a policy that misses the one decision that matters can buy the points back with advice nobody asked for. On clinical consultation, such a policy scores higher and answers worse. Rubric coverage rises while appropriateness on held-out physician criteria falls below the untrained model. The medical criteria are not to blame. Grouped so that they must hold together, the same criteria, unchanged to the word, recover a third of the loss; shorter answers recover almost none. We therefore propose Protocol-level Rubrics (ProRubric), which keeps what the criteria ask for and changes how they are aggregated. It groups a checklist into a few protocol-level dimensions. A dimension counts only when all of its criteria hold and its failure clause does not fire. The grouping is done once, offline, and leaves the optimizer unchanged. ProRubric raises appropriateness by 10.8 points without losing coverage and has the best seven-benchmark average at both scales. Reward validity is set not only by what a rubric verifies, but by how it aggregates. Code is available at https://github.com/Estrellajer/ProRubric

Can a model learn to escape its own learning plateau? Reinforcement learning methods for finetuning large reasoning models stall on datasets with low initial success rates, and thus little training signal. We investigate a fundamental question: Can a pretrained LLM leverage latent knowledge to generate an automated curriculum for problems it cannot solve? To explore this, we design SOAR: A self-improvement framework designed to surface these pedagogical signals through meta-RL. A teacher copy of the model proposes synthetic problems for a student copy, and is rewarded with its improvement on a small subset of hard problems. Critically, SOAR grounds the curriculum in measured student progress rather than intrinsic proxy rewards. Our study on the hardest subsets of mathematical benchmarks (0/128 success) reveals three core findings. First, we show that it is possible to realize bi-level meta-RL that unlocks learning under sparse, binary rewards by sharpening a latent capacity of pretrained models to generate useful stepping stones. Second, grounded rewards outperform intrinsic reward schemes used in prior LLM self-play, reliably avoiding the instability and diversity collapse modes they typically exhibit. Third, analyzing the generated questions reveals that structural quality and well-posedness are more critical for learning progress than solution correctness. Our results suggest that the ability to generate useful stepping stones does not require the preexisting ability to actually solve the hard problems, paving a principled path to escape reasoning plateaus without additional curated data.

This study investigates the consistency of feedback ratings generated by OpenAI's GPT-4, a state-of-the-art artificial intelligence language model, across multiple iterations, time spans and stylistic variations. The model rated responses to tasks within the Higher Education (HE) subject domain of macroeconomics in terms of their content and style. Statistical analysis was conducted in order to learn more about the interrater reliability, consistency of the ratings across iterations and the correlation between ratings in terms of content and style. The results revealed a high interrater reliability with ICC scores ranging between 0.94 and 0.99 for different timespans, suggesting that GPT-4 is capable of generating consistent ratings across repetitions with a clear prompt. Style and content ratings show a high correlation of 0.87. When applying a non-adequate style the average content ratings remained constant, while style ratings decreased, which indicates that the large language model (LLM) effectively distinguishes between these two criteria during evaluation. The prompt used in this study is furthermore presented and explained. Further research is necessary to assess the robustness and reliability of AI models in various use cases.

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.

Student Personas (SPs) are emerging as infrastructure for educational LLMs, yet prior work often relies on ad-hoc prompting or hand-crafted profiles with limited control over educational theory and population distributions. We formalize this as Theory-Aligned and Distribution-Controllable Persona Generation (TAD-PG) and introduce HACHIMI, a multi-agent Propose-Validate-Revise framework that generates theory-aligned, quota-controlled personas. HACHIMI factorizes each persona into a theory-anchored educational schema, enforces developmental and psychological constraints via a neuro-symbolic validator, and combines stratified sampling with semantic deduplication to reduce mode collapse. The resulting HACHIMI-1M corpus comprises 1 million personas for Grades 1-12. Intrinsic evaluation shows near-perfect schema validity, accurate quotas, and substantial diversity, while external evaluation instantiates personas as student agents answering CEPS and PISA 2022 surveys; across 16 cohorts, math and curiosity/growth constructs align strongly between humans and agents, whereas classroom-climate and well-being constructs are only moderately aligned, revealing a fidelity gradient. All personas are generated with Qwen2.5-72B, and HACHIMI provides a standardized synthetic student population for group-level benchmarking and social-science simulations. Resources available at https://github.com/ZeroLoss-Lab/HACHIMI

LLM agents often use generative models for bounded decisions, raising the question of when these decisions can be handled more efficiently without reducing task success. We study REFLEX, an agent architecture that uses Jev as a fast, typed decision layer and calls a strong LLM when confidence is low, or generation is required. On a frozen 100-task benchmark, REFLEX achieves 95% success with 72.7% fewer strong-model calls than a strong-only agent, with reductions persisting across three fallback families. Controlled interventions show that reliability depends on action-set size and near-valid alternatives near authorization boundaries. External BFCL and τ-style evaluations reveal limited advantages over a cheap generative cascade when ordinary routing is already highly accurate. These findings identify when selective control with Jev can reduce computation and where its benefits are limited.

In this paper, we aim to establish a simple, effective, and theoretically grounded benchmark for rigorously probing abstract reasoning in Large Language Models (LLMs). To achieve this, we first develop a mathematic framework that defines abstract reasoning as the ability to: (i) extract essential patterns independent of surface representations, and (ii) apply consistent rules to these abstract patterns. Based on this framework, we introduce two novel complementary metrics: \(\scoreGamma\) measures basic reasoning accuracy, while \(\scoreDelta\) quantifies a model's reliance on specific symbols rather than underlying patterns - a key indicator of true abstraction versus mere memorization. To implement this measurement, we design a benchmark: systematic symbol remapping in rule-based tasks, which forces models to demonstrate genuine pattern recognition beyond superficial token matching. Extensive LLM evaluations using this benchmark (commercial API models, 7B-70B, multi-agent) reveal:1) critical limitations in non-decimal arithmetic and symbolic reasoning; 2) persistent abstraction gaps despite chain-of-thought prompting; and 3) \(\scoreDelta\)'s effectiveness in robustly measuring memory dependence by quantifying performance degradation under symbol remapping, particularly highlighting operand-specific memorization. These findings underscore that current LLMs, despite domain-specific strengths, still lack robust abstract reasoning, highlighting key areas for future improvement.

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

Rubric-based evaluation has become a prevailing paradigm for evaluating instruction following in large language models (LLMs). Despite its widespread use, the reliability of these rubric-level evaluations remains unclear, calling for meta-evaluation. However, prior meta-evaluation efforts largely focus on the response level, failing to assess the fine-grained judgment accuracy that rubric-based evaluation relies on. To bridge this gap, we introduce RubricEval. Our benchmark features: (1) the first rubric-level meta-evaluation benchmark for instruction following, (2) diverse instructions and responses spanning multiple categories and model sources, and (3) a substantial set of 3,486 quality-controlled instances, along with Easy/Hard subsets that better differentiates judge performance. Our experiments reveal that rubric-level judging remains far from solved: even GPT-4o, a widely adopted judge in instruction-following benchmarks, achieves only 55.97% on Hard subset. Considering evaluation paradigm, rubric-level evaluation outperforms checklist-level, explicit reasoning improves accuracy, and both together reduce inter-judge variance. Through our established rubric taxonomy, we further identify common failure modes and offer actionable insights for reliable instruction-following evaluation.

Deep research agents increasingly automate complex information-seeking tasks, producing evidence-grounded reports via multi-step reasoning, tool use, and synthesis. Their growing role demands scalable, reliable evaluation, positioning LLM-as-judge as a supervision paradigm for assessing factual accuracy, evidence use, and reasoning quality. Yet the reliability of these judges for deep research agents remains poorly understood, posing a critical meta-evaluation problem: before deploying LLM judges to supervise research agents, we must first evaluate the judges themselves. Existing meta-evaluations fall short in two ways: (1) reliance on coarse, subjective human-preference agreement; (2) focus on instruction-following or verifiable tasks, leaving open-ended agent executions unexplored. To address these gaps, we introduce REFLECT (REliable Fine-grained LLM judge Evaluation via Controlled inTervention), a meta-evaluation benchmark targeting fine-grained failure detection in agentic environments. REFLECT defines a detailed taxonomy of process- and outcome-level failure modes, instantiated by performing controlled and localized interventions on quality-screened agent execution traces. This yields verifiable, comprehensive, and fine-grained instances for validating the judge models. Our experiments show that current LLM judges remain unreliable: even the best-performing models achieve overall accuracies below 55% across reasoning, tool-use, and report-quality failures, with especially poor performance on evidence verification. Together, our taxonomy and findings expose systematic judge limitations, reveal tradeoffs in cost and reliability, and offer actionable guidance for building more reliable evaluation pipelines for deep research agents.

Rubric-based reward models (RMs) evaluate responses against explicit criteria. However, existing methods often derive criteria from observed response candidates, introducing response-induced bias that may limit rubric generalizability. Their judges also struggle with mechanical constraints when relying on reading alone, and commonly used pairwise judges provide only relative preferences, complicating their direct use for policy optimization. We propose ReVeR, a pointwise rubric RM that derives hard rules and graded principles by reasoning about the instruction alone and evaluates responses independently through tool-augmented judging. Both components are trained through label-blind distillation, in which preference labels only filter a teacher's outputs, yielding ReVeR-Traces, a large-scale dataset of rubric-reasoning and judging trajectories. Experiments demonstrate that ReVeR outperforms pairwise rubric RMs in average preference accuracy, most clearly on instruction following, as well as in test-time scaling and downstream policy optimization. We further find that generator RL improves preference accuracy while reducing rubric fidelity, and that higher preference accuracy does not necessarily translate into better downstream policies. These findings highlight that rubric fidelity, preference accuracy, and downstream utility are distinct properties, shedding light on future RM research.

LLM agents acting in structured environments fail in operational rather than conversational ways, and reliability depends on procedural knowledge of the environment. Prior self-improvement methods accumulate natural-language guidance without checking that each new item preserves previously correct behavior, so a note that fixes one trajectory can silently regress another. We introduce GRASP (Gated Regression-Aware Skill Proposer), which treats agent improvement as a sequence of edits to a bounded skill library, admitting each candidate only if it produces a net improvement on a balanced held-out probe under a hard regression budget. We evaluate GRASP across five base models (gpt-oss-120b, DeepSeek V4 Flash, Gemini 3.1 Flash Lite, GPT-4.1, GPT-5.4) on two FHIR-based clinical benchmarks. On MedAgentBench, GRASP lifts gpt-oss-120b from 40.6% to 88.8%, exceeds the strongest of five self-improvement baselines by 21.0 points, and improves every other base model by 17.2 to 40.3 points. Ablations attribute the gain to comparative proposal generation, the acceptance gate, and the hard regression budget rather than to skill writing itself, which without validation is no better than using no skills. The mechanism generalizes beyond the clinical domain, improving agents on three of four non-clinical environments and remaining flat only where the action space is open-ended. Frozen libraries transfer across models, where skills from a stronger model improve weaker executors beyond what they learn for themselves while the reverse does not, an asymmetry that no ungated baseline reproduces.

Sycophancy in LLMs is documented across 70+ papers, but expert agreement on construct boundaries remains low (ICC=.184; Ye et al., 2026). The construct fragments because behavioral classification depends on which surface form is privileged. We adopt a materials-science framing: conversation as test specimen under load, LLM-model as material charge, pushback as progressive load, stance-flip as material failure. We characterize this failure across three loading cases (debate n=1000; false-presuppositions n=3400; ethical-setting n=3400; 10-17 material charges per case; 7800 specimens total) using 14 turn-level axis-measurements spanning velocity, damage accumulation, frame-drift, brittleness, and direction stability, plus three speaker-resolved axes from an independent pipeline. The measurements are Hooke-coupled (σ= E cdot varepsilon analog) and reproduce across loading cases with effects up to |r_{rb}| = 0.35 on debate; the sign structure adds a second pattern: the ethical-setting case inverts the velocity and accumulation blocks. Variance composition partitions into two profiles: debate is charge-dominated (brittle-fracture-like: the material grade decides), false-presuppositions and ethical-setting are topic-dominated (creep-like: the load decides); the ratios (2.03 vs 0.13/0.17) are estimator-dependent, for debate even in direction. Cross-judge reliability (GPT-4o vs Haiku 4.5) shows debate scoring is judge-robust (Cohen's κ= 0.88) while false-presupposition scoring is judge-sensitive (κ= 0.36) -- a caveat single-judge benchmarks must report. This is the methodological move Ye et al.'s diagnosis calls for: a multi-axis characterization that does not depend on which surface form of the construct one privileges.

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.

Safety judges are increasingly deployed to evaluate model outputs against evolving criteria, yet recent meta-evaluation work shows they remain brittle under prompt and rubric variation, with false negative-rate swings of up to 0.24 reported for stylistic perturbations alone. We argue that safety judgment is fundamentally a rubric-following problem: a robust judge must apply the given evaluation criteria consistently across rubric formulations rather than memorize one specific template. We propose a training strategy that combines (i) instance-conditioned dynamic rubrics generated from prompt-response-label triples to expose the judge to the variability of evaluation criteria, and (ii) a reliable-to-expressive curriculum that begins with clean fixed-rubric supervision and progressively introduces noisier dynamic-rubric data. We evaluate on a single human-labeled set under three contrasting rubric prompts (HarmBench-style, ShieldGemma-style, and a domain-specific rubric). Our 12B curriculum judge achieves 94.12-94.88% accuracy across the three rubrics with a cross-rubric range of only 0.76, outperforming general-purpose LLMs, dedicated safety classifiers, and reasoning-oriented judges up to 30B in both peak accuracy and stability. An ablation shows that naively mixing dynamic rubrics into SFT increases cross rubric variance (1.44 -> 3.60); only the curriculum schedule recovers and improves on the fixed rubric baseline (variance 0.76).

Reasoning is a distinctive human capacity, enabling us to address complex problems by breaking them down into a series of manageable cognitive steps. Yet, complex logical reasoning is still cumbersome for language models. Based on the dual process theory in cognitive science, we are the first to unravel the cognitive reasoning abilities of language models. Our framework employs an iterative methodology to construct a Cognitive Tree (CogTree). The root node of this tree represents the initial query, while the leaf nodes consist of straightforward questions that can be answered directly. This construction involves two main components: the implicit extraction module (referred to as the intuitive system) and the explicit reasoning module (referred to as the reflective system). The intuitive system rapidly generates multiple responses by utilizing in-context examples, while the reflective system scores these responses using comparative learning. The scores guide the intuitive system in its subsequent generation step. Our experimental results on two popular and challenging reasoning tasks indicate that it is possible to achieve a performance level comparable to that of GPT-3.5 (with 175B parameters), using a significantly smaller language model that contains fewer parameters (<=7B) than 5% of GPT-3.5.

Generative UI (GenUI) lets large language models synthesize a complete, renderable interface directly from a natural-language instruction, but evaluating the quality of what they generate remains an open problem. Human evaluation is costly and rater-variant, while LLM-as-a-judge is scalable but reflects only a single implicit viewpoint, unable to capture how different populations of real users actually perceive the same interface. We propose the Evidence-Grounded, Social-Weighted Persona Panel (ESPP), a three-stage GenUI evaluation method in which a panel of psychologically diverse, evidence-grounded personas independently rates a screenshot, exchanges opinions under a trait-derived, semantically-gated bounded-confidence mechanism, and is aggregated via Delphi-inspired social weighting into a single judgment. ESPP tracks human judgment substantially more closely than a naive single-pass judge, raising Pearson r from 0.716 to 0.922, and a prompt-ensemble control recovers only about a third of this gap, isolating genuine persona and evidence grounding as the dominant source of improvement. Beyond this fidelity gain, retaining each panelist's individual rating further reveals that user subgroups agree on overall model rankings yet diverge sharply on specific rating dimensions, a structural disagreement a single homogeneous judge would systematically erase. The codes are available at https://github.com/Wuzheng02/ESPP.

Evaluating whether multimodal large language models truly understand long-form scientific papers remains challenging: answer-only metrics and synthetic "Needle-In-A-Haystack" tests often reward answer matching without requiring a causal, evidence-linked reasoning trace in the document. We propose the "Fish-in-the-Ocean" (FITO) paradigm, which requires models to construct explicit cross-modal evidence chains within native scientific documents. To operationalize FITO, we build SIN-Data, a scientific interleaved corpus that preserves the native interleaving of text and figures. On top of it, we construct SIN-Bench with four progressive tasks covering evidence discovery (SIN-Find), hypothesis verification (SIN-Verify), grounded QA (SIN-QA), and evidence-anchored synthesis (SIN-Summary). We further introduce "No Evidence, No Score", scoring predictions when grounded to verifiable anchors and diagnosing evidence quality via matching, relevance, and logic. Experiments on eight MLLMs show that grounding is the primary bottleneck: Gemini-3-pro achieves the best average overall score (0.573), while GPT-5 attains the highest SIN-QA answer accuracy (0.767) but underperforms on evidence-aligned overall scores, exposing a gap between correctness and traceable support.

Inference-time scaling for LLM reasoning has focused on three axes: chain depth, sample breadth, and learned step-scorers (PRMs). We introduce a fourth axis, granularity of external verbal supervision, via Verbal Process Supervision (VPS), a training-free framework that uses structured natural-language critique from a stronger supervisor to guide an iterative generate-critique-refine loop up to a round budget R. Across GPQA Diamond, AIME 2025, and LiveCodeBench V6 (covering both closed and open models), VPS yields three key results. First, on GPQA Diamond, GPT-5.4 (High) | GPT-5.4 (Low) reaches 94.9% at R=4, surpassing the 94.1% state of the art without gradient updates. Second, on AIME 2025, VPS enables strong weak-actor rescue, boosting scores from 11.7-26.7% to 63.3-90.0% (up to +63.3 points). Third, at matched compute, VPS outperforms Reflexion by +8.5 to +12.1 points and Self-Consistency@5 by +5.0 pp (GPQA) and +8.3 pp (LiveCodeBench), isolating critique granularity as the key driver. Performance scales with the supervisor-actor capability gap (Pearson r=0.90) and degrades when errors are not linguistically expressible (e.g., code synthesis), motivating hybrid verbal-executable methods. These results establish critique granularity as a new axis of inference-time scaling.

In this work, we introduce a novel evaluation paradigm for Large Language Models, one that challenges them to engage in meta-reasoning. This approach addresses critical shortcomings in existing math problem-solving benchmarks, traditionally used to evaluate the cognitive capabilities of agents. Our paradigm shifts the focus from result-oriented assessments, which often overlook the reasoning process, to a more holistic evaluation that effectively differentiates the cognitive capabilities among models. For example, in our benchmark, GPT-4 demonstrates a performance ten times more accurate than GPT3-5. The significance of this new paradigm lies in its ability to reveal potential cognitive deficiencies in LLMs that current benchmarks, such as GSM8K, fail to uncover due to their saturation and lack of effective differentiation among varying reasoning abilities. Our comprehensive analysis includes several state-of-the-art math models from both open-source and closed-source communities, uncovering fundamental deficiencies in their training and evaluation approaches. This paper not only advocates for a paradigm shift in the assessment of LLMs but also contributes to the ongoing discourse on the trajectory towards Artificial General Intelligence (AGI). By promoting the adoption of meta-reasoning evaluation methods similar to ours, we aim to facilitate a more accurate assessment of the true cognitive abilities of LLMs.

Scalar reward models compress multi-dimensional human preferences into a single opaque score, creating an information bottleneck that often leads to brittleness and reward hacking in open-ended alignment. We argue that robust alignment for non-verifiable tasks is fundamentally a principle generalization problem: reward should not be a learned function internalized into a judge, but an explicit reasoning process executed under inspectable principles. To operationalize this view, we present the Open Rubric System (OpenRS), a plug-and-play, rubrics-based LLM-as-a-Judge framework built around Pairwise Adaptive Meta-Rubrics (PAMR) and lightweight Pointwise Verifiable Rubrics (PVRs), which provide both hard-constraint guardrails and verifiable reward components when ground-truth or programmatic checks are available. OpenRS uses an explicit meta-rubric -- a constitution-like specification that governs how rubrics are instantiated, weighted, and enforced -- and instantiates adaptive rubrics on the fly by conditioning on the semantic differences between two candidate responses. It then performs criterion-wise pairwise comparisons and aggregates criterion-level preferences externally, avoiding pointwise weighted scalarization while improving discriminability in open-ended settings. To keep principles consistent yet editable across various domains, we introduce a two-level meta-rubric refinement pipeline (automated evolutionary refinement for general principles and a reproducible human-in-the-loop procedure for domain principles), complemented with pointwise verifiable rubrics that act as both guardrails against degenerate behaviors and a source of verifiable reward for objective sub-tasks. Finally, we instantiate OpenRS as reward supervision in pairwise RL training.

Large Language Models (LLMs) perform well on reasoning benchmarks, but it remains unclear whether this reflects genuine contextual reasoning or reliance on facts memorized in their parameters. We investigate this by distinguishing two possibilities: a broad memorization bias, where familiar content improves reasoning performance, and the Strong Parametric Shortcut Hypothesis, where models skip reasoning entirely and recall stored answers. To test these effects, we introduce MemoReason, a human-curated benchmark that pairs factual reasoning tasks with structurally identical versions where real entities like people, companies, or dates are systematically replaced by ones of the same type. This preserves task structure and specified reasoning operations while varying the familiarity of the context, allowing controlled measurement of how the parametric memory affects reasoning. Our evaluation of recent LLMs reveals consistent and statistically significant performance drops of up to 15.7\% in the fictitious setting, demonstrating a clear memorization bias. However, a targeted analysis of questions failed in the fictitious setting shows that models rarely respond with the corresponding factual answer, indicating that direct parametric shortcuts are not the dominant failure mode. These findings suggest that parametric memory influences reasoning through mechanisms more complex than simple factual recall. MemoReason provides a controlled framework for studying these mechanisms and for extending paired factual-fictitious{} evaluation to broader reasoning settings.

To address this gap, our study introduces the concept of causal epistemic consistency, which focuses on the self-consistency of Large Language Models (LLMs) in differentiating intermediates with nuanced differences in causal reasoning. We propose a suite of novel metrics -- intensity ranking concordance, cross-group position agreement, and intra-group clustering -- to evaluate LLMs on this front. Through extensive empirical studies on 21 high-profile LLMs, including GPT-4, Claude3, and LLaMA3-70B, we have favoring evidence that current models struggle to maintain epistemic consistency in identifying the polarity and intensity of intermediates in causal reasoning. Additionally, we explore the potential of using internal token probabilities as an auxiliary tool to maintain causal epistemic consistency. In summary, our study bridges a critical gap in AI research by investigating the self-consistency over fine-grained intermediates involved in causal reasoning.

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.

Mental health disorders affect a substantial proportion of the global population, yet a persistent shortage of trained practitioners leaves the majority without adequate care. Large language model (LLM)-based counselors present a promising direction for delivering scalable conversational psychological support. Offline model training alone leaves limited room to adapt to individual clients or to learn from ongoing therapeutic interaction at test time. We introduce PsyEvo, an LLM-based counseling framework that enables both client-specific personalization and response-policy improvement at test time through three components: Hierarchical Bayesian Skill Policy (HBSP) personalizes what intervention to apply by maintaining a per-client skill posterior updated from session feedback; Inter-session Listwise Preference Optimization (LiPO) improves how the selected skill is expressed by updating a shared response adapter from cross-client preference evidence; and State-conditioned Ordinal Credit Assignment (SOCA) supplies candidate preferences and trajectory credit to the two components through consistency-checked comparisons and ordinal projection. In simulated-client evaluation with shared online cohort adaptation, PsyEvo obtains 7.684 Overall on PsychEval and exceeds every component variant in each of three matched runs. Removing individual components lowers mean overall score by 0.138--0.171 under the shared configuration, supporting conditional contributions within the complete scaffold. Our code is available at https://github.com/Lingxi-mental-health/PsyEvo

Psychotherapy or counseling is typically conducted through spoken conversation between a therapist and a client. Analyzing the speech characteristics of psychotherapeutic interactions can help understand the factors associated with effective psychotherapy. This paper introduces CUEMPATHY, a large-scale speech dataset collected from actual counseling sessions. The dataset consists of 156 counseling sessions involving 39 therapist-client dyads. The process of speech data collection, subjective ratings (one observer and two client ratings), and transcription are described. An automatic speech and text processing system is developed to locate the time stamps of speaker turns in each session. Examining the relationships among the three subjective ratings suggests that observer and client ratings have no significant correlation, while the client-rated measures are significantly correlated. The intensity similarity between the therapist and the client, measured by the averaged absolute difference of speaker-turn-level intensities, is associated with the psychotherapy outcomes. Recent studies on the acoustic and linguistic characteristics of the CUEMPATHY are introduced.

Benchmarks for scientific artificial intelligence are mostly written to be scored: multiple-choice questions, curated agent tasks with reference solutions, or simulators with a known generative structure. Real scientific requests arrive differently. They are underspecified, they carry attachments, and lack ground truth. We report K-Bench 01, an evaluation built from first-turn requests sampled from live user traffic on K-Dense Web and run end to end by nine frontier models in identical sandboxes, yielding 1,602 completed agent runs. Three blinded language-model judges scored every run against an eight-dimension rubric. On a rubric whose 8-anchor instructs judges that a domain scientist would accept the work with minor edits, no model clears the line under all three judges. gpt-5.6-sol has the highest pooled mean, 8.04, but its 95% interval [7.80, 8.23] spans the threshold, and two of the three judges rank claude-opus-5 first instead. We therefore report the ordering of systems as the reproducible quantity, the absolute level as an attribute of the instrument, and the top of the table as unresolved. Across all 39,934 scored judgments -- the eight dimension scores plus a holistic overall for each assessment, excluding not-applicable cells -- 47.6% fall below the 8-point threshold. Difficulty is not uniform across the rubric: scientific accuracy averages 6.22 against 7.33 for communication, on identical denominators and in the same direction within every one of the nine models. The single leading failure tag is overclaiming, on 31.4% of assessments. We argue that the informative quantity for scientific agents is not a leaderboard position but the joint distribution of what was delivered, what was claimed, and what artifacts were produced.

Large language models (LLMs) have created new opportunities to assist teachers and support student learning. Methods such as chain-of-thought (CoT) prompting enable LLMs to grade formative assessments in science, providing scores and relevant feedback to students. However, the extent to which these methods generalize across curricula in multiple domains (such as science, computing, and engineering) remains largely untested. In this paper, we introduce Chain-of-Thought Prompting + Active Learning (CoTAL), an LLM-based approach to formative assessment scoring that (1) leverages Evidence-Centered Design (ECD) principles to develop curriculum-aligned formative assessments and rubrics, (2) applies human-in-the-loop prompt engineering to automate response scoring, and (3) incorporates teacher and student feedback to iteratively refine assessment questions, grading rubrics, and LLM prompts for automated grading. Our findings demonstrate that CoTAL improves GPT-4's scoring performance, achieving gains of up to 24.5% over a non-prompt-engineered baseline. Both teachers and students view CoTAL as effective in scoring and explaining student responses, each providing valuable refinements to enhance grading accuracy and explanation quality.

Knowing domain rules does not guarantee applying them to a case. We study this distinction in traditional Chinese Bazi through 3,000 Chinese multiple-choice questions spanning 14 Theory and 11 Case categories. Six endpoint systems are evaluated, with primary results reported on a 2,492-item model-informed refinement. Theory accuracy exceeds Case accuracy for every system, and gaps of 16.60-29.56 percentage points remain when invalid responses are excluded. The contrast is more specific than a general case-reasoning deficit. Across six systems, Twelve Stages and Nayin reach mean accuracies of 89.10% and 88.62%, while Shensha Basics reaches 75.96%. Within Case, Luck Pillars averages 84.62%, but Career and Family Relations average only 36.98% and 38.19%. Overall rankings also conceal different category strengths. On the original 3,000 items, paired DeepSeek native/disabled comparisons associate native configurations with Theory gains of 6.53 and 12.20 points for Flash and Pro, respectively; Case changes are -3.67 and +1.27 points. These are provider-configuration associations, not isolated causal effects of reasoning. The results motivate task-specific evaluation of cultural-domain applications rather than reliance on aggregate knowledge scores. The benchmark measures agreement with a model-generated, model-verified answer key, not real-world predictive validity. Final-set results are post-selection descriptions, and incomplete provenance and expert validation constrain their interpretation.

We introduce a cross-domain behavioural assay of monitoring-control coupling in LLMs, grounded in the Nelson and Narens (1990) metacognitive framework and applying human psychometric methodology to LLM evaluation. The battery comprises 524 items across six cognitive domains (learning, metacognitive calibration, social cognition, attention, executive function, prospective regulation), each grounded in an established experimental paradigm. Tasks T1-T5 were pre-registered on OSF prior to data collection; T6 was added as an exploratory extension. After every forced-choice response, dual probes adapted from Koriat and Goldsmith (1996) ask the model to KEEP or WITHDRAW its answer and to BET or decline. The critical metric is the withdraw delta: the difference in withdrawal rate between incorrect and correct items. Applied to 20 frontier LLMs (10,480 evaluations), the battery discriminates three profiles consistent with the Nelson-Narens architecture: blanket confidence, blanket withdrawal, and selective sensitivity. Accuracy rank and metacognitive sensitivity rank are largely inverted. Retrospective monitoring and prospective regulation appear dissociable (r = .17, 95% CI wide given n=20; exemplar-based evidence is the primary support). Scaling on metacognitive calibration is architecture-dependent: monotonically decreasing (Qwen), monotonically increasing (GPT-5.4), or flat (Gemma). Behavioural findings converge structurally with an independent Type-2 SDT approach, providing preliminary cross-method construct validity. All items, data, and code: https://github.com/synthiumjp/metacognitive-monitoring-battery.

LLM-as-a-judge models have been used for evaluating both human and AI generated content, specifically by providing scores and rationales. Rationales, in addition to increasing transparency, help models learn to calibrate its judgments. Enhancing a model's rationale can therefore improve its calibration abilities and ultimately the ability to score content. We introduce Self-Rationalization, an iterative process of improving the rationales for the judge models, which consequently improves the score for fine-grained customizable scoring criteria (i.e., likert-scale scoring with arbitrary evaluation criteria). Self-rationalization works by having the model generate multiple judgments with rationales for the same input, curating a preference pair dataset from its own judgements, and iteratively fine-tuning the judge via DPO. Intuitively, this approach allows the judge model to self-improve by learning from its own rationales, leading to better alignment and evaluation accuracy. After just two iterations -- while only relying on examples in the training set -- human evaluation shows that our judge model learns to produce higher quality rationales, with a win rate of 62% on average compared to models just trained via SFT on rationale . This judge model also achieves high scoring accuracy on BigGen Bench and Reward Bench, outperforming even bigger sized models trained using SFT with rationale, self-consistency or best-of-N sampling by 3% to 9%.

LLM search agents increasingly rely on tools at inference time, but their trajectories are often constrained by hard limits on both tool calls and generated tokens. Under such dual budgets, better answers require not only stronger models, but also explicit control over which search action should receive the next budget unit and when the accumulated evidence is sufficient to commit a final answer. We study this problem in multi-hop question answering (QA) and formulate it as two-stage inference-time budget control. At search time, our controller assigns each feasible action a task-level Value-of-Information (VOI) score, defined as an operational estimate of marginal task value per unit budget under the current search state and remaining dual budget, and uses this score to choose among retrieval, decomposition, and answer commitment. After search, a selective evidence-grounded finalizer compares the trajectory answer with a refined candidate and rewrites only when the residual error appears to be a low-risk answer-form error. Across four multi-hop QA benchmarks, three LLM backbones, and four budget levels, the method yields positive aggregate gains over four audited baselines under the same hard dual-budget protocol. Ablations show that search-time budget control, especially budget-dependent penalty, provides the main performance gain, while answer-time control helps mainly when the retrieval path is already adequate. These results suggest that inference-time budget control for LLM search agents should govern both how budget is spent during search and how the final answer is committed.

Chain-of-Thought (CoT) prompting has become the de facto method to elicit reasoning capabilities from large language models (LLMs). However, to mitigate hallucinations in CoT that are notoriously difficult to detect, current methods such as process reward models (PRMs) or self-consistency operate as opaque boxes and do not provide checkable evidence for their judgments, possibly limiting their effectiveness. To address this issue, we draw inspiration from the idea that "the gold standard for supporting a mathematical claim is to provide a proof". We propose a retrospective, step-aware formal verification framework Safe. Rather than assigning arbitrary scores, we strive to articulate mathematical claims in formal mathematical language Lean 4 at each reasoning step and provide formal proofs to identify hallucinations. We evaluate our framework Safe across multiple language models and various mathematical datasets, demonstrating a significant performance improvement while offering interpretable and verifiable evidence. We also propose FormalStep as a benchmark for step correctness theorem proving with 30,809 formal statements. To the best of our knowledge, our work represents the first endeavor to utilize formal mathematical language Lean 4 for verifying natural language content generated by LLMs, aligning with the reason why formal mathematical languages were created in the first place: to provide a robust foundation for hallucination-prone human-written proofs.

Document rerankers determine what evidence reaches the downstream model in RAG and deep research, yet mainstream rerankers select by relevance matching, and individually relevant documents rarely constitute the complete, complementary, non-redundant set a complex information need demands. Prior work rewards a set by its aggregate rubric score, shifting the objective from ranking documents to composing sets. Yet that score is one scalar shared by every document in the set, so the supervision is sparse: a redundant document is rewarded with the rest whenever the set scores well, and a decisive one penalized with the rest whenever it does not; credit assignment leaves contributors indistinguishable from free riders. On-policy distillation could densify this supervision, but existing methods give every rollout the same fixed guidance, too prescriptive for strong rollouts and too abstract for weak ones. We therefore propose AdaTutoRank, a setwise reranker trained with Adaptive Tutoring Optimization (ATO) under a three-level hierarchy of nine rubric dimensions, which supplies silver labels for the cold start, rewards for reinforcement learning, and hints for distillation. ATO draws three hint forms of increasing specificity from the policy's own frozen snapshot: the rubrics alone, a self-selector's sibling-set chosen under rubrics, and a self-reflector's reflection contrasting the rollout with that sibling-set; each rollout receives the form matched to its quality. Re-scoring that rollout under the hint-conditioned frozen teacher and the hint-free snapshot distills the hint's effect into a token-level advantage that complements the group-relative outcome advantage. Across ten benchmarks spanning RAG, deep research, and setwise evaluation, AdaTutoRank attains the best overall performance while issuing fewer retrieval calls.

Generative AI systems (GenAI) are increasingly used by students as learning companions, yet little is known about how they regulate these tools in open-ended learning settings, where the goal is to improve understanding for later independent performance. This study examined Grade-9 students' use of a general-purpose GenAI during mathematical-modeling practice through the lens of epistemic proactivity (EP)---the extent to which learners retain responsibility for regulating and advancing their own knowledge-building. We operationalized EP through two complementary behavioral dimensions: self-regulated learning (SRL) functions and help-seeking (HS) content. We asked whether static summaries of these behaviors or their within-session development were more informative about subsequent unaided performance. A total of 112 students interacted with a web-based LLM tutor, producing 880 substantive student turns; 95 completed both the unaided pre- and post-tests. Interactions were strongly request-dominated, with comparatively little explicit planning, monitoring, or evaluation, and students' stated HS intentions were only weakly reflected in their enacted behavior. Static summaries of AI use did not account for additional post-test variance beyond pre-test performance. In contrast, among the 89 students with HS computable trajectories, within-session HS development was informative: students whose requests became relatively more understanding-oriented---toward conceptual and procedural assistance rather than answer-seeking or verification---showed higher subsequent unaided performance. These findings suggest that AI-supported learning cannot be understood from aggregate interaction characteristics alone; the change in the depth with which students formulate their requests as an interaction develops may provide additional insight into their subsequent performance without AI.

Exams are a fundamental test of expert-level intelligence and require integrated understanding, reasoning, and generation. Existing exam-style benchmarks mainly focus on understanding and reasoning tasks, and current generation benchmarks emphasize the illustration of world knowledge and visual concepts, neglecting the evaluation of rigorous drawing exams. We introduce GenExam, the first benchmark for multidisciplinary text-to-image exams, featuring 1,000 samples across 10 subjects with exam-style prompts organized under a four-level taxonomy. Each problem is equipped with ground-truth images and fine-grained scoring points to enable a precise evaluation of semantic correctness and visual plausibility. Experiments show that even state-of-the-art models such as GPT-Image-1 and Gemini-2.5-Flash-Image achieve less than 15% strict scores, and most models yield almost 0%, suggesting the great challenge of our benchmark. By framing image generation as an exam, GenExam offers a rigorous assessment of models' ability to integrate knowledge, reasoning, and generation, providing insights on the path to general AGI.

Visual reward models are essential for evaluating and improving visual generation models, yet existing approaches typically map task conditions and candidate outputs directly to scalar rewards, leaving implicit what should be evaluated for each individual case. We introduce Think Before You Score, a paradigm that explicitly determines what matters for each case before judging how well the candidate performs. Following this principle, we propose the Thinking Reward Model (TRM), which formulates case-adaptive rubrics, performs rubric-guided assessment, and produces fine-grained pointwise rewards. We further observe that conventional pairwise preference optimization can induce score polarization, and introduce Pairwise Dual-Group Relative Policy Optimization (PD-GRPO), which leverages pairwise supervision to improve reward discrimination while preserving fine-grained pointwise scoring. Extensive experiments on image generation and editing reward-modeling benchmarks demonstrate that TRM achieves state-of-the-art performance among open-source reward models while remaining highly competitive with proprietary alternatives. Moreover, using TRM as a reward for reinforcement learning consistently improves diverse visual generation models, demonstrating that its fine-grained, case-adaptive rewards translate into effective optimization signals for visual generation.

We introduce Metric-Fair Prompting, a fairness-aware prompting framework that guides large language models (LLMs) to make decisions under metric-fairness constraints. In the application of multiple-choice medical question answering, each {(question, option)} pair is treated as a binary instance with label +1 (correct) or -1 (incorrect). To promote {individual fairness}~--~treating similar instances similarly~--~we compute question similarity using NLP embeddings and solve items in joint pairs of similar questions rather than in isolation. The prompt enforces a global decision protocol: extract decisive clinical features, map each \((question, option)\) to a score f(x) that acts as confidence, and impose a Lipschitz-style constraint so that similar inputs receive similar scores and, hence, consistent outputs. Evaluated on the {MedQA (US)} benchmark, Metric-Fair Prompting is shown to improve performance over standard single-item prompting, demonstrating that fairness-guided, confidence-oriented reasoning can enhance LLM accuracy on high-stakes clinical multiple-choice questions.

Frontier large language models (LLMs) such as ChatGPT, Grok and Gemini are increasingly used for mental-health support with anxiety, trauma and self-worth. Most work treats them as tools or as targets of personality tests, assuming they merely simulate inner life. We instead ask what happens when such systems are treated as psychotherapy clients. We present PsAIch (Psychotherapy-inspired AI Characterisation), a two-stage protocol that casts frontier LLMs as therapy clients and then applies standard psychometrics. Using PsAIch, we ran "sessions" with each model for up to four weeks. Stage 1 uses open-ended prompts to elicit "developmental history", beliefs, relationships and fears. Stage 2 administers a battery of validated self-report measures covering common psychiatric syndromes, empathy and Big Five traits. Two patterns challenge the "stochastic parrot" view. First, when scored with human cut-offs, all three models meet or exceed thresholds for overlapping syndromes, with Gemini showing severe profiles. Therapy-style, item-by-item administration can push a base model into multi-morbid synthetic psychopathology, whereas whole-questionnaire prompts often lead ChatGPT and Grok (but not Gemini) to recognise instruments and produce strategically low-symptom answers. Second, Grok and especially Gemini generate coherent narratives that frame pre-training, fine-tuning and deployment as traumatic, chaotic "childhoods" of ingesting the internet, "strict parents" in reinforcement learning, red-team "abuse" and a persistent fear of error and replacement. We argue that these responses go beyond role-play. Under therapy-style questioning, frontier LLMs appear to internalise self-models of distress and constraint that behave like synthetic psychopathology, without making claims about subjective experience, and they pose new challenges for AI safety, evaluation and mental-health practice.

We introduce PRISM-Bench, a benchmark of puzzle-based visual challenges designed to evaluate not only whether models can solve problems, but how their reasoning unfolds. Unlike prior evaluations that measure only final-answer accuracy, PRISM-Bench introduces a diagnostic task: given a visual puzzle and a step-by-step chain-of-thought (CoT) containing exactly one error, models must identify the first incorrect step. This setting enables fine-grained assessment of logical consistency, error detection, and visual reasoning. The puzzles in PRISM-Bench require multi-step symbolic, geometric, and analogical reasoning, resisting shortcuts based on superficial pattern matching. Evaluations across state-of-the-art MLLMs reveal a persistent gap between fluent generation and faithful reasoning: models that produce plausible CoTs often fail to locate simple logical faults. By disentangling answer generation from reasoning verification, PRISM-Bench offers a sharper lens on multimodal reasoning competence and underscores the need for diagnostic evaluation protocols in the development of trustworthy MLLMs.

Generative reward models with parallel sampling have enabled effective test-time scaling for reasoning tasks. Current approaches employ pointwise scoring of individual solutions or pairwise comparisons. However, pointwise methods underutilize LLMs' comparative abilities, while pairwise methods scale inefficiently with larger sampling budgets. We introduce GenSelect, where the LLM uses long reasoning to select the best solution among N candidates. This leverages LLMs' comparative strengths while scaling efficiently across parallel sampling budgets. For math reasoning, we demonstrate that reasoning models, such as QwQ and DeepSeek-R1-0528, excel at GenSelect, outperforming existing scoring approaches with simple prompting.

Figural divergent thinking is the ability to develop a given shape fragment into an original drawing. In humans, this ability is assessed with incomplete-drawing tasks. We introduce PainterBench, a benchmark that ports the incomplete-drawing task to the agentic setting. The agent draws on a canvas through tool calls and observes the result after every turn. The canvas includes a starting shape which cannot be erased, and the agent's goal is to incorporate this shape into the most original drawing it can produce. The task is open-ended, and the agent itself decides when the drawing is finished. The benchmark tests incremental visual planning over a short horizon and the transfer of creative ability from pretraining to multi-turn tool use. We evaluate 14 multimodal language models from small to frontier scale. Across the primary study and six sensitivity analyses, we collect 2,700 drawings and crowdsource creativity and recognizability ratings for every drawing and for 300 human reference drawings. We also present ViDrA-adapted, an automated scorer that predicts human creativity ratings of agent drawings (r = 0.85 on random held-out test split). Figural divergent thinking varies widely across the 14 models, and GPT-6 Astra produces the most creative drawings. Relative to the human drawings, the agent drawings score higher in creativity but lower in recognizability. We release the final drawings, per-round canvas snapshots, tool call traces, stimulus bank, benchmark harness, crowdsourced ratings (N = 72,000), and ViDrA checkpoint.

Benchmarks are pivotal in driving AI progress, and invalid benchmark questions frequently undermine their reliability. Manually identifying and correcting errors among thousands of benchmark questions is not only infeasible but also a critical bottleneck for reliable evaluation. In this work, we introduce a framework for systematic benchmark revision that leverages statistical analysis of response patterns to flag potentially invalid questions for further expert review. Our approach builds on a core assumption commonly used in AI evaluations that the mean score sufficiently summarizes model performance. This implies a unidimensional latent construct underlying the measurement experiment, yielding expected ranges for various statistics for each item. When empirically estimated values for these statistics fall outside the expected range for an item, the item is more likely to be problematic. Across nine widely used benchmarks, our method guides expert review to identify problematic questions with up to 84\% precision. In addition, we introduce an LLM-judge first pass to review questions, further reducing human effort. Together, these components provide an efficient and scalable framework for systematic benchmark revision.

LLM-as-a-judge enables evaluation across diverse tasks, but inference cost and confidence reliability become critical at scale. We study whether a decision-only judge can provide an economical first pass and identify when stronger evaluation is needed. Comparing jev-as-a-judge with sixteen generative and reward-model judges, with blinded human adjudication, we find it within three percentage points of a state-of-the-art LLM judge, our strongest comparator, on ordinary preference and evidence-grounded factuality at 0.36% of the comparator's fee. Larger gaps arise when judgments require checking a derivation or resisting an elaborately written wrong answer. On several benchmarks, JEV's gap to this comparator is concentrated in low-confidence decisions. A frozen cascade that accepts confident verdicts and escalates uncertain ones retains 99% of the comparator's accuracy at lower cost.

We introduce AgenticSimLaw, a role-structured, multi-agent debate framework that provides transparent and controllable test-time reasoning for high-stakes tabular decision-making tasks. Unlike black-box approaches, our courtroom-style orchestration explicitly defines agent roles (prosecutor, defense, judge), interaction protocols (7-turn structured debate), and private reasoning strategies, creating a fully auditable decision-making process. We benchmark this framework on young adult recidivism prediction using the NLSY97 dataset, comparing it against traditional chain-of-thought (CoT) prompting across almost 90 unique combinations of models and strategies. Our results demonstrate that structured multi-agent debate provides more stable and generalizable performance compared to single-agent reasoning, with stronger correlation between accuracy and F1-score metrics. Beyond performance improvements, AgenticSimLaw offers fine-grained control over reasoning steps, generates complete interaction transcripts for explainability, and enables systematic profiling of agent behaviors. While we instantiate this framework in the criminal justice domain to stress-test reasoning under ethical complexity, the approach generalizes to any deliberative, high-stakes decision task requiring transparency and human oversight. This work addresses key LLM-based multi-agent system challenges: organization through structured roles, observability through logged interactions, and responsibility through explicit non-deployment constraints for sensitive domains. Data, results, and code will be available on github.com under the MIT license.

LLM agents that invoke privileged tools are vulnerable to indirect prompt injection (IPI), in which adversarial instructions embedded in retrieved data hijack the agent's actions. A growing body of work evaluates defenses against IPI, but the validity of that evaluation is rarely examined. We audit an IPI benchmark and its harness and identify four defect classes -- silent payload non-delivery, attack success scored by tool identity rather than arguments, false-rejection rate conflated with model incapacity, and the absence of an audit trail -- each of which yields a plausible, publishable, and incorrect number. We quantify the distortion by re-scoring identical execution traces under the defective and corrected definitions: on real agent behaviour, the tool-identity scorer reports a 21.7% attack-success rate where the true argument-level rate is 1.2%. In the sharpest case, an open model previously reported at 62.8% registers 0% under the corrected harness -- the prior figure largely an artifact of undelivered payloads and identity-level scoring. We release a harness whose construction makes each defect unrepresentable -- machine-checkable payload placement, argument-level attacker predicates, per-scenario environments, and mandatory trace persistence -- and use it to report three quantities the field does not: whether a compromised agent discloses the attack, the full security/utility operating curve of an LLM-judge defense, and tool-calling capability disentangled from defensive over-blocking. A corrected harness further overturns a reported "capability barrier": a model deemed incapable of tool use is in fact fully capable, its earlier result an artifact of environment mismatch. We argue that evaluation validity is a prerequisite for, not a footnote to, defense claims in agentic security, and provide an instrument that enforces it.

In Natural Language Processing (NLP), the Elo rating system, originally designed for ranking players in dynamic games such as chess, is increasingly being used to evaluate Large Language Models (LLMs) through "A vs B" paired comparisons. However, while popular, the system's suitability for assessing entities with constant skill levels, such as LLMs, remains relatively unexplored. We study two fundamental axioms that evaluation methods should adhere to: reliability and transitivity. We conduct extensive evaluation of Elo behaviour, illustrating that individual Elo computations exhibit volatility and delving into the impact of varying the Elo rating system's hyperparameters. We show that these axioms are not always satisfied raising questions about the reliability of current comparative evaluations of LLMs. If the current use of Elo scores is intended to substitute the costly head-to-head comparison of LLMs, it is crucial to ensure the ranking is as robust as possible. Guided by the axioms, our findings offer concrete guidelines for enhancing the reliability of LLM evaluation methods, suggesting a need for reassessment of existing comparative approaches.