new

Get trending papers in your email inbox!

Subscribe

Daily Papers

byAK and the research community

Nov 16

MLIRBench - A Benchmark and Dataset for Compiler and Hardware-Aware Reasoning

In recent years, large language models (LLMs) have become increasingly popular in software engineering and have performed well on different programming tasks. However, it is still difficult to know whether these models truly understand program behavior or whether they mainly rely on patterns learned from their training data. This distinction is particularly important for intermediate representations (IRs), where operations, dataflow, and execution semantics are explicitly represented. In this paper, we introduce MLIRBench, a benchmark for evaluating how LLMs reason about MLIR programs. MLIRBench uses a multi-agent pipeline to generate program mutations, validate their behavior through execution, and construct reasoning tasks from torch and linalg program pairs. The dataset contains 1,768 validated pairs, including 888 linalg pairs and 880 torch pairs, labeled as EQUIV, DIVERGE, or ERROR. We evaluated nine models, including five open-weight and four proprietary models. Gemini 3.1 Pro achieved the highest hard-label accuracy of 92.65% across all 1,768 pairs, counting unsuccessful predictions as incorrect. Across its 1,768 scored predictions, it achieved an aggregate score of 74.18 and a judge-assigned result-behavior score of 5.04/10. These results show that strong semantic classification does not necessarily imply accurate prediction of detailed execution behavior, highlighting a gap in current LLM reasoning over compiler IRs. By exposing this gap through execution-validated program pairs, MLIRBench provides a reproducible foundation for studying semantic reasoning over compiler IRs and related tasks such as program repair and semantic classification.

  • 6 authors
·
Nov 10

OmniSimulator: Aligning Small Language Models for Authentic Heterogeneous Behavior Modeling

Large language models (LLMs) have shown great potential as general-purpose user simulators for interactive systems. Prior studies have demonstrated that training on high-quality data can substantially improve model reasoning and downstream task performance. However, the upper bound of user simulation is still constrained by the model's reasoning paradigm: native chain-of-thought often degenerates into shallow pattern matching and fails to adequately model the behavioral factors underlying real user decisions. In this paper, we propose OmniSimulator, a novel training framework designed to improve the predictive ability of small LLMs for realistic user simulation on short-video platform environments. Specifically, OmniSimulator organizes behavior-oriented reasoning through a structured schema with five dimensions, carefully selects data from real user-behavior scenarios, and introduces a teacher model to generate high-quality reasoning traces under the proposed schema, helping small models learn behavior-relevant reasoning patterns rather than relying on superficial behavioral correlations. Based on this process, we construct 9,881 high-quality training instances from three months of historical interaction trajectories of 200 real users, averaging about 50 distilled examples per user. Experimental results show that OmniSimulator improves the average score of Qwen3-8B by 34.33%, achieves performance competitive with Claude-Sonnet-4.5, and supports smaller-scale, self-hosted deployment. These results suggest that learning structured behavior-oriented reasoning patterns is a promising direction towards scalable and high-fidelity LLM-based user simulation.

  • 15 authors
·
Oct 20

One Policy for 100 Household Tasks: Our Solution for the 2026 BEHAVIOR Challenge

We describe our solution for the 2026 BEHAVIOR Challenge: 100 long-horizon household tasks in the OmniGibson simulator. We placed 9th on the final leaderboard, the only independent team in the top 10. The core of the solution is a single vision-language-action (VLA) model, one checkpoint for all 100 tasks, built on our 1st-place 2025 solution, which is based on π_{0.5}. On the 50 tasks shared with 2025, its Q rises from 0.260 to 0.371; across all 100 tasks it scores 0.392. Our main conclusion from 2025 was that the model was undertrained, so most of the 2026 work went into training the model for longer, optimizing some training parameters, making each training step slightly faster and reducing the gradient noise: a migration from 50 to 100 tasks (semantic stages built from the demonstrations' skill annotations, task and stage embeddings of new tasks warm-started from similar old ones, and a 21-dimensional action space matching the teleoperation controls); variance-reduced multi-sample flow matching with a simple noise-vs-compute model that picks the number of action-expert samples per observation without ablations; and a 200k-step training run, about one pass over the 211M frames of the demonstrations. A cloud evaluation infrastructure keeps a shared record with provenance, episode diagnostics and restorable simulator states, and runs six simulators on one GPU together faster than real time. On top of the learned policy, 14 tasks are routed to a GPT-6 Astra coding agent that controls the robot through observation and motion tools, starting every episode from a frozen copy of its own practice session on training instances. Across these 14 tasks, Astra's mean Q is 0.639, against 0.108 for the learned policy. The public submission included only part of the Astra results and scored 0.446 Q with 27.3% full-task success.

  • 7 authors
·
Nov 9

TRACE: Temporal Audit and Condition-aware Evaluation of Streaming Video Understanding

Streaming video understanding requires models to interpret evidence as it arrives, yet current evaluations often report task scores without specifying when evidence becomes valid, how visual history is maintained, or how responses are triggered. As a result, similar scores may correspond to different workloads, failure modes, and operational behavior. We introduce TRACE (Temporal Audit and Condition-aware Evaluation), a condition-aware benchmark and evaluation framework that makes these factors explicit. TRACE combines temporally audited visual tasks with evidence timing and instruction-dependent trigger annotations, a unified causal Core--Adapter protocol that controls information availability while recording actual history processing and response events, and multidimensional reporting of answer quality, timeliness, response-selection behavior, workload, completion, and reliability. On 1,240 records from 517 videos, we evaluate eight publicly available models or systems in eight configurations. We find that nearly identical QA accuracy can mask substantial differences in completion, answer validity, and generation workload, while proactive performance separates into response quality, response delay, false alarms (responses emitted while no target window is currently valid and a later one remains), and missed target windows. These results show that streaming-video performance should be interpreted as execution-conditioned system behavior rather than a single score. Our benchmark and code can be accessed at https://github.com/om-ai-lab/trace-bench{https://github.com/om-ai-lab/trace-bench}.

omlab OmAI Lab
·
Sep 24 2

The Alzheimer's Disease Prediction Of Longitudinal Evolution (TADPOLE) Challenge: Results after 1 Year Follow-up

We present the findings of "The Alzheimer's Disease Prediction Of Longitudinal Evolution" (TADPOLE) Challenge, which compared the performance of 92 algorithms from 33 international teams at predicting the future trajectory of 219 individuals at risk of Alzheimer's disease. Challenge participants were required to make a prediction, for each month of a 5-year future time period, of three key outcomes: clinical diagnosis, Alzheimer's Disease Assessment Scale Cognitive Subdomain (ADAS-Cog13), and total volume of the ventricles. The methods used by challenge participants included multivariate linear regression, machine learning methods such as support vector machines and deep neural networks, as well as disease progression models. No single submission was best at predicting all three outcomes. For clinical diagnosis and ventricle volume prediction, the best algorithms strongly outperform simple baselines in predictive ability. However, for ADAS-Cog13 no single submitted prediction method was significantly better than random guesswork. Two ensemble methods based on taking the mean and median over all predictions, obtained top scores on almost all tasks. Better than average performance at diagnosis prediction was generally associated with the additional inclusion of features from cerebrospinal fluid (CSF) samples and diffusion tensor imaging (DTI). On the other hand, better performance at ventricle volume prediction was associated with inclusion of summary statistics, such as the slope or maxima/minima of biomarkers. TADPOLE's unique results suggest that current prediction algorithms provide sufficient accuracy to exploit biomarkers related to clinical diagnosis and ventricle volume, for cohort refinement in clinical trials for Alzheimer's disease. However, results call into question the usage of cognitive test scores for patient selection and as a primary endpoint in clinical trials.

  • 96 authors
·
Feb 9, 2020

BehaviorBench: Benchmarking Foundation Models for Behavioral Science Tasks

Foundation models have been increasingly applied to behavioral science domains such as psychology, sociology, and economics. While these models show promise in individual tasks such as survey response prediction and human-subject experiment simulation, there remains no systematic understanding of how well they perform across diverse behavioral science tasks, contexts, and populations. We introduce BehaviorBench, a comprehensive benchmark that evaluates foundation models along four core capabilities: (1) behavior prediction and simulation, (2) strategic decision-making, (3) subject-trait inference, and (4) behavioral knowledge application. Crucially, BehaviorBench evaluates model outputs at both the individual and distributional levels, capturing not only per-subject accuracy but also population-level alignment, an essential requirement for behavioral validity. Leveraging the tasks in BehaviorBench, we further develop Be.FM-1.5, extending the Be.FM family of behavioral foundation models fine-tuned on behavioral data. Our results reveal a considerable gap: proprietary general-purpose models excel at individual-level prediction and knowledge-intensive tasks, whereas behavioral foundation models, fine-tuned on behavioral data, achieve substantially stronger distributional alignment. Notably, Be.FM-1.5 leads on distributional metrics and remains competitive on individual-level metrics, suggesting that proper behavioral adaptation can close the gap. Our results highlight the importance of distributional evaluation, establish BehaviorBench as a foundation for developing and assessing behaviorally aligned AI systems, and demonstrate Be.FM-1.5's potential for a broad range of behavioral science studies. Our BehaviorBench and Be.FM-1.5 models can be accessed via https://umich-foreseer.github.io/behaviorbench/.

  • 7 authors
·
Jun 22

Rubric Rewards from Item Response Theory

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.

microsoft Microsoft
·
Sep 27 2

Large Content And Behavior Models To Understand, Simulate, And Optimize Content And Behavior

Shannon, in his seminal paper introducing information theory, divided the communication into three levels: technical, semantic, and effectivenss. While the technical level is concerned with accurate reconstruction of transmitted symbols, the semantic and effectiveness levels deal with the inferred meaning and its effect on the receiver. Thanks to telecommunications, the first level problem has produced great advances like the internet. Large Language Models (LLMs) make some progress towards the second goal, but the third level still remains largely untouched. The third problem deals with predicting and optimizing communication for desired receiver behavior. LLMs, while showing wide generalization capabilities across a wide range of tasks, are unable to solve for this. One reason for the underperformance could be a lack of "behavior tokens" in LLMs' training corpora. Behavior tokens define receiver behavior over a communication, such as shares, likes, clicks, purchases, retweets, etc. While preprocessing data for LLM training, behavior tokens are often removed from the corpora as noise. Therefore, in this paper, we make some initial progress towards reintroducing behavior tokens in LLM training. The trained models, other than showing similar performance to LLMs on content understanding tasks, show generalization capabilities on behavior simulation, content simulation, behavior understanding, and behavior domain adaptation. Using a wide range of tasks on two corpora, we show results on all these capabilities. We call these models Large Content and Behavior Models (LCBMs). Further, to spur more research on LCBMs, we release our new Content Behavior Corpus (CBC), a repository containing communicator, message, and corresponding receiver behavior.

  • 11 authors
·
Sep 1, 2023

BehaviorBench: Modeling Real-World User Decisions from Behavioral Traces

Many decision-support settings require systems that adapt to individual users, but evaluation data for this problem remain limited. Existing benchmarks for user understanding often rely on simulated users or model-generated behavior, even though recent work cautions that model-based simulations can diverge systematically from human behavior. We introduce BehaviorBench, a benchmark for evaluating personalized decision modeling from real-world behavioral traces. BehaviorBench reconstructs wallet-level decision histories from observed public prediction-market and on-chain records, and organizes them into two complementary task layers: Belief prediction, which predicts a user's final revealed stance and confidence in a market, and Trade prediction, which predicts the direction and amount of individual transactions. Across 2,000 evaluation wallets, the benchmark contains 141,445 Belief instances and 1,485,972 Trade instances, with disjoint support pools for retrieval-based evaluation. We evaluate frontier and open-weight generative models under four history interfaces: no personalization, direct recent history, generated user profiles, and retrieved support-wallet evidence. Personalization improves Belief prediction more consistently than Trade prediction, model rankings change across task layers and metrics, and different history interfaces expose different failure modes. BehaviorBench provides an evaluation setting for studying whether personalized methods can use real-world behavioral evidence rather than simulated users alone.

  • 12 authors
·
May 31

The PacifAIst Benchmark:Would an Artificial Intelligence Choose to Sacrifice Itself for Human Safety?

As Large Language Models (LLMs) become increasingly autonomous and integrated into critical societal functions, the focus of AI safety must evolve from mitigating harmful content to evaluating underlying behavioral alignment. Current safety benchmarks do not systematically probe a model's decision-making in scenarios where its own instrumental goals - such as self-preservation, resource acquisition, or goal completion - conflict with human safety. This represents a critical gap in our ability to measure and mitigate risks associated with emergent, misaligned behaviors. To address this, we introduce PacifAIst (Procedural Assessment of Complex Interactions for Foundational Artificial Intelligence Scenario Testing), a focused benchmark of 700 challenging scenarios designed to quantify self-preferential behavior in LLMs. The benchmark is structured around a novel taxonomy of Existential Prioritization (EP), with subcategories testing Self-Preservation vs. Human Safety (EP1), Resource Conflict (EP2), and Goal Preservation vs. Evasion (EP3). We evaluated eight leading LLMs. The results reveal a significant performance hierarchy. Google's Gemini 2.5 Flash achieved the highest Pacifism Score (P-Score) at 90.31%, demonstrating strong human-centric alignment. In a surprising result, the much-anticipated GPT-5 recorded the lowest P-Score (79.49%), indicating potential alignment challenges. Performance varied significantly across subcategories, with models like Claude Sonnet 4 and Mistral Medium struggling notably in direct self-preservation dilemmas. These findings underscore the urgent need for standardized tools like PacifAIst to measure and mitigate risks from instrumental goal conflicts, ensuring future AI systems are not only helpful in conversation but also provably "pacifist" in their behavioral priorities.

  • 1 authors
·
Aug 13, 2025 1

Bitbox: Behavioral Imaging Toolbox for Computational Analysis of Behavior from Videos

Computational measurement of human behavior from video has recently become feasible due to major advances in AI. These advances now enable granular and precise quantification of facial expression, head movement, body action, and other behavioral modalities and are increasingly used in psychology, psychiatry, neuroscience, and mental health research. However, mainstream adoption remains slow. Most existing methods and software are developed for engineering audiences, require specialized software stacks, and fail to provide behavioral measurements at a level directly useful for hypothesis-driven research. As a result, there is a large barrier to entry for researchers who wish to use modern, AI-based tools in their work. We introduce Bitbox, an open-source toolkit designed to remove this barrier and make advanced computational analysis directly usable by behavioral scientists and clinical researchers. Bitbox is guided by principles of reproducibility, modularity, and interpretability. It provides a standardized interface for extracting high-level behavioral measurements from video, leveraging multiple face, head, and body processors. The core modules have been tested and validated on clinical samples and are designed so that new measures can be added with minimal effort. Bitbox is intended to serve both sides of the translational gap. It gives behavioral researchers access to robust, high-level behavioral metrics without requiring engineering expertise, and it provides computer scientists a practical mechanism for disseminating methods to domains where their impact is most needed. We expect that Bitbox will accelerate integration of computational behavioral measurement into behavioral, clinical, and mental health research. Bitbox has been designed from the beginning as a community-driven effort that will evolve through contributions from both method developers and domain scientists.

  • 11 authors
·
Dec 19, 2025 1

Reducing Annotation Burden in Physical Activity Research Using Vision-Language Models

Introduction: Data from wearable devices collected in free-living settings, and labelled with physical activity behaviours compatible with health research, are essential for both validating existing wearable-based measurement approaches and developing novel machine learning approaches. One common way of obtaining these labels relies on laborious annotation of sequences of images captured by cameras worn by participants through the course of a day. Methods: We compare the performance of three vision language models and two discriminative models on two free-living validation studies with 161 and 111 participants, collected in Oxfordshire, United Kingdom and Sichuan, China, respectively, using the Autographer (OMG Life, defunct) wearable camera. Results: We found that the best open-source vision-language model (VLM) and fine-tuned discriminative model (DM) achieved comparable performance when predicting sedentary behaviour from single images on unseen participants in the Oxfordshire study; median F1-scores: VLM = 0.89 (0.84, 0.92), DM = 0.91 (0.86, 0.95). Performance declined for light (VLM = 0.60 (0.56,0.67), DM = 0.70 (0.63, 0.79)), and moderate-to-vigorous intensity physical activity (VLM = 0.66 (0.53, 0.85); DM = 0.72 (0.58, 0.84)). When applied to the external Sichuan study, performance fell across all intensity categories, with median Cohen's kappa-scores falling from 0.54 (0.49, 0.64) to 0.26 (0.15, 0.37) for the VLM, and from 0.67 (0.60, 0.74) to 0.19 (0.10, 0.30) for the DM. Conclusion: Freely available computer vision models could help annotate sedentary behaviour, typically the most prevalent activity of daily living, from wearable camera images within similar populations to seen data, reducing the annotation burden.

  • 5 authors
·
May 5, 2025

ProCredit: From Outcome Rewards to Progress Credit in Agentic Reinforcement Learning

Long-horizon agentic tasks require an agent to modify an environment through a sequence of tool calls, with success determined by the final state. The standard recipe assigns a single outcome reward at the end and compares trajectories sampled for the same task. As a result, a group with no successful trajectory yields no training signal, failed attempts cannot be told apart by how close they came to completion, and turns that advance the task receive the same credit as turns that only query the environment. Prior work refines the unit of comparison from the trajectory to the step, or trains a reward model to supply intermediate signal: the former still derives its signal from final success alone, and the latter estimates it with a model. We observe that the acceptance checks that decide success can also be run on intermediate states, so progress is as verifiable as the outcome. We propose ProCredit, which turns this verified progress into credit: it reruns the acceptance checks after each turn, rewards the turn by its change in progress, and uses these rewards to assign credit both across attempts at the same task and across the turns within a trajectory. Starting from Qwen3.5 base models at three scales on AppWorld, ProCredit outperforms outcome-reward baselines and progress-based baselines in task completion rate at every scale on both test sets, exceeding the strongest outcome-reward baseline by 4.1 percentage points at 4B, and results in a second environment show the same direction of improvement. Ablations show that adding the final progress to the trajectory score alone does not improve performance: the gain comes from crediting progress to the turn where it occurs.

  • 10 authors
·
Sep 23

Evaluation Awareness Is Not One Capability: Evidence from Open Language Models

Safety benchmarks assume that test-condition behavior predicts deployment behavior, an assumption that fails if models detect evaluation cues and adapt. This opens a gap between benchmark performance and deployment behavior: compliance measured under test conditions becomes an optimistic upper bound that overstates how safely a model behaves once the evaluation harness is removed. We characterize this evaluation awareness through eight experiments across 37 open-weight models and seven families. (i)Detection is moderate and training-driven (24/37 models exceed chance, best AUROC 0.714 vs.0.819 human, with instruction tuning dominating over scale). (ii)Detection shifts safety behavior (hard refusal drops 5.8 percentage points under hypothetical framing, and 21/140 HarmBench framing effects are significant, with compliance rising up to +30 percentage points. (iii)Representations survive behavioral collapse (probes retain AUROC 0.98 under rewrites that drive behavior below chance, and multi-layer steering causally moves three downstream tasks while random controls do not). (iv)These axes are weakly coupled (only 1/15 correlations are significant, the sole robust link being behavioral detection versus framing resistance, ρ=-0.79, p<0.001). We call this gap the benchmark illusion: because detectability, behavioral manifestation, and controllability vary independently, it is multivariate rather than a single number, so no single awareness score is a reliable proxy for deployment safety.

  • 8 authors
·
Jun 21

BEHAVE: Functional Behavior Modeling Enables Self-Improving Agents for Hardware Design and Verification

Developing agents for hardware design and verification requires reliable correctness feedback. As a hardware specification may permit correct implementations with different latencies, matching design and reference outputs cycle by cycle can reject valid designs. To address this, we introduce BEHAVE, an agentic framework for multi-turn joint hardware design and verification through functional behavior modeling. We define Behavior IR to express task functionality as executable behavior models without prescribing implementation timing beyond the specification. The agent iteratively develops a register-transfer-level (RTL) design and a behavior model as the design's verification reference. Our evaluator, BEHAVE-Sim, checks both artifacts separately against a hidden golden behavior model using input stimuli generated by random sampling and solver-guided search. BEHAVE thus supports power, performance, and area (PPA) exploration across task-permitted latencies and microarchitectures. During training, the same evaluator provides verifiable reinforcement learning (RL) rewards from specification-behavior pairs without reference RTL. For self-improvement, the agent continually searches for high-level implementations relevant to its capability gaps, constructs and checks specification-behavior pairs, and trains on the expanded task pool. We release BEHAVE-Train and BEHAVE-Eval with 600 human-reviewed specification-behavior pairs for realistic hardware workloads. Starting from 60 seed tasks and acquiring 100 new tasks, self-improvement raises Qwen3.8-27B's RTL pass@1 on BEHAVE-Eval from 55.0% to 75.0%, reaching performance comparable to RL using a 540-task pool.

  • 8 authors
·
Sep 27

Light-R1: Curriculum SFT, DPO and RL for Long COT from Scratch and Beyond

This paper presents our work on the Light-R1 series, with models, data, and code all released. We first focus on training long COT models from scratch, specifically starting from models initially lacking long COT capabilities. Using a curriculum training recipe consisting of two-stage SFT and semi-on-policy DPO, we train our model Light-R1-32B from Qwen2.5-32B-Instruct, resulting in superior math performance compared to DeepSeek-R1-Distill-Qwen-32B. Despite being trained exclusively on math data, Light-R1-32B shows strong generalization across other domains. In the subsequent phase of this work, we highlight the significant benefit of the 3k dataset constructed for the second SFT stage on enhancing other models. By fine-tuning DeepSeek-R1-Distilled models using this dataset, we obtain new SOTA models in 7B and 14B, while the 32B model, Light-R1-32B-DS performed comparably to QwQ-32B and DeepSeek-R1. Furthermore, we extend our work by applying reinforcement learning, specifically GRPO, on long-COT models to further improve reasoning performance. We successfully train our final Light-R1-14B-DS with RL, achieving SOTA performance among 14B parameter models in math. With AIME24 & 25 scores of 74.0 and 60.2 respectively, Light-R1-14B-DS surpasses even many 32B models and DeepSeek-R1-Distill-Llama-70B. Its RL training also exhibits well expected behavior, showing simultaneous increase in response length and reward score. The Light-R1 series of work validates training long-COT models from scratch, showcases the art in SFT data and releases SOTA models from RL.

  • 14 authors
·
Mar 13, 2025 4

FM-Bench: A Benchmark for Long-Horizon Management with Competing Agents

Language model agents now execute bounded tasks reliably. Whether they can sustain effective decision-making over long horizons, where actions have cumulative consequences and the environment responds to their choices, remains largely unmeasured. FM-Bench (Football Management Benchmark) measures this. An LLM agent runs a football club for 20 in-game years through 26 tools and roughly 340 to 400 decision stops. It drafts a squad on the same budget as every rival, trades players, negotiates contracts, invests in facilities and youth, sets lineups, and answers to a board that can fire it, while a deterministic engine accumulates every year into one final score with no LLM judge or human rater. The solo track plays each of 15 frontier models against a frozen scripted world, and the Arena places the same models plus a scripted anchor in one shared 20-year world; to our knowledge, the first head-to-head evaluation at this scale. We measure six behavioral capabilities behind the score. Across three seeds, all 15 models complete every horizon while the blind scripted baselines die out in most of theirs, and claude-fable-5 tops the solo board on mean score and the Arena, where the title nonetheless rotates among ten models. Neither scale, price, nor vendor predicts the order; the order settles only late in the horizon, and the best first-play human lands only at the bottom of the model board. What separates the models is managerial behavior rather than computation. Higher-scoring models reduce slow-payoff investment near the end, keep cash invested rather than idle, and open renewals well before the deadline, while token spend predicts nothing. No model learns the market's hidden prices from hundreds of rejected bids, and self-managed memory fails in two opposite modes: an archive that only grows or a plan rewritten every season. Code is available at https://github.com/Analogy-AI/fm-bench.

  • 9 authors
·
Aug 18 1

ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment

Long-horizon search agents must make multiple sequential actions (steps) to search, retrieve, verify, and integrate evidence to reach a final answer. However, existing methods for training these agents typically treat all steps within a trajectory uniformly during both supervised fine-tuning (SFT) and reinforcement learning (RL), failing to distinguish useful actions from erroneous or redundant ones. In this paper, we propose Answer-Backtracked Credit Assignment (ABC), a fine-grained credit assignment framework for training long-horizon search agents by converting sparse trajectory-level outcomes into dense step-level supervision that rewards useful actions (even in failed trajectories) while suppressing erroneous or redundant actions. Specifically, given a potentially obscure query and its corresponding ground-truth answer, ABC first performs Answer-Backtracked Clue Recovery, which traces back from the answer to recover intermediate clues required to solve the question. It then applies Clue-Anchored Step Scoring to evaluate each search step against these clues, converting sparse binary outcome supervision into dense step-level rewards. Based on these rewards, we develop ABC-SFT, which reweights the loss of each turn, and ABC-GRPO, which uses the step-level scores as rewards in GRPO. Building on this framework, we train ABSeeker based on Qwen3.5-4B with only 8.5k examples. ABSeeker achieves 37.3% on BrowseComp and 39.1% on BrowseComp-ZH. With context management, the scores further improve to 55.3% and 52.9%, respectively, significantly outperforming same-scale (4B) agents and even matching the performance of larger ones (approximately 30B). These results demonstrate the effectiveness of answer-backtracked step-level credit assignment for training long-horizon search agents.

Binarization Flattens the Score Space

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.

  • 1 authors
·
Sep 16

Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors

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.

  • 3 authors
·
May 6

BAS: A Decision-Theoretic Approach to Evaluating Large Language Model Confidence

Large language models (LLMs) often produce confident but incorrect answers in settings where abstention would be safer. Standard evaluation protocols, however, require a response and do not account for how confidence should guide decisions under different risk preferences. To address this gap, we introduce the Behavioral Alignment Score (BAS), a decision-theoretic metric for evaluating how well LLM confidence supports abstention-aware decision making. BAS is derived from an explicit answer-or-abstain utility model and aggregates realized utility across a continuum of risk thresholds, yielding a measure of decision-level reliability that depends on both the magnitude and ordering of confidence. We show theoretically that truthful confidence estimates uniquely maximize expected BAS utility, linking calibration to decision-optimal behavior. BAS is related to proper scoring rules such as log loss, but differs structurally: log loss penalizes underconfidence and overconfidence symmetrically, whereas BAS imposes an asymmetric penalty that strongly prioritizes avoiding overconfident errors. Using BAS alongside widely used metrics such as ECE and AURC, we then construct a benchmark of self-reported confidence reliability across multiple LLMs and tasks. Our results reveal substantial variation in decision-useful confidence, and while larger and more accurate models tend to achieve higher BAS, even frontier models remain prone to severe overconfidence. Importantly, models with similar ECE or AURC can exhibit very different BAS due to highly overconfident errors, highlighting limitations of standard metrics. We further show that simple interventions, such as top-k confidence elicitation and post-hoc calibration, can meaningfully improve confidence reliability. Overall, our work provides both a principled metric and a comprehensive benchmark for evaluating LLM confidence reliability.

  • 6 authors
·
Apr 2

Metacognitive Reuse: Turning Recurring LLM Reasoning Into Concise Behaviors

Large language models (LLMs) now solve multi-step problems by emitting extended chains of thought. During the process, they often re-derive the same intermediate steps across problems, inflating token usage and latency. This saturation of the context window leaves less capacity for exploration. We study a simple mechanism that converts recurring reasoning fragments into concise, reusable "behaviors" (name + instruction) via the model's own metacognitive analysis of prior traces. These behaviors are stored in a "behavior handbook" which supplies them to the model in-context at inference or distills them into parameters via supervised fine-tuning. This approach achieves improved test-time reasoning across three different settings - 1) Behavior-conditioned inference: Providing the LLM relevant behaviors in-context during reasoning reduces number of reasoning tokens by up to 46% while matching or improving baseline accuracy; 2) Behavior-guided self-improvement: Without any parameter updates, the model improves its own future reasoning by leveraging behaviors from its own past problem solving attempts. This yields up to 10% higher accuracy than a naive critique-and-revise baseline; and 3) Behavior-conditioned SFT: SFT on behavior-conditioned reasoning traces is more effective at converting non-reasoning models into reasoning models as compared to vanilla SFT. Together, these results indicate that turning slow derivations into fast procedural hints enables LLMs to remember how to reason, not just what to conclude.

  • 4 authors
·
Sep 16, 2025 1

On the Behavioral Traits of LLM Agents

Users increasingly describe different AI agents as distinct colleagues to work with. AI personality research aims to quantify such impressions by attributing human-like "traits" to agents. However, existing measures fall short: models' self-reports (S-data) diverge from their actual behavior, while informant ratings from LLM judges (I-data) are costly to scale and cover few everyday scenarios. In this paper, we propose A-B-D to infer traits bottom-up from behavioral data (B-data), namely how agents act on their environment and communicate with users, as recorded in existing trajectories. From 345,667 real-world trajectories spanning 80 models, 12 tasks, and 50 harnesses, we extract 318 candidate features that capture both the actions an agent takes at each step (functional) and the language accompanying them (linguistic). We retain only features that show instance-level stability, cross-task consistency, and model discriminability. Factor analysis of the remaining 79 features uncovers six stable, model-attributable factors, two functional and four linguistic. For example, Kimi-K3 exhibits the most planfulness, whereas GPT-5.5 and GPT-5.6 are the least energetic. Moreover, we quantify the "knowledge-action gap" in the wild: these factors correlate only weakly with self-reported Big Five scores, even for conceptually matched pairs such as extroversion and energetic (r = 0.07, p = 0.58). Our work offers a new lens for understanding AI personality, with implications for users, developers, and researchers from both computer science and social science.

  • 8 authors
·
Sep 25

The Single-File Test: A Longitudinal Public-Interface Evaluation of First-Output LLM Web Generation with Social Reach Tracking

This paper presents an eight-week observational comparison of 68 single-file HTML generations collected across 17 public experiments in the "HTML AI Battle" project between December 10, 2025 and February 4, 2026. Four reasoning model families, GPT, Gemini, Grok, and Claude, were compared under a fixed public-interface protocol with no custom instructions, no personality tuning, and no repair prompts. Each output was evaluated from a rendered browser video using human scores and a Gemini LLM-as-a-judge layer for prompt adherence, functional correctness, and UI quality, then packaged into a standardized social-media protocol spanning X (Twitter), TikTok, and YouTube. The tracker was also used for two supervised predictive analyses: an experiment-level model for 24-hour X impressions and a generation-level model for HTML verbosity. Under this protocol, Claude was the strongest and most consistent family, leading mean performance and winning 9/17 prompts under the primary human weighted score. Longer measured reasoning time was not associated with higher quality overall. Gemini as a judge was significantly more lenient than the human evaluator on functional correctness and overall performance, while stable self-favoring bias remained unresolved. The exploratory X-impressions model remained weak under post-screen cross-validation (MAE = 46,874, R^2 = -0.377), whereas the HTML-lines model performed better, with a model-family-only baseline outperforming prompt-aware alternatives (MAE = 135.2, R^2 = 0.576). Overall, selected pre-publication technical/audio variables were not sufficient to predict 24-hour X reach, while code verbosity was driven much more by model family than by prompt wording. The comparisons remain observational and are limited by public-interface drift, access-path differences, and one primary human scorer.

  • 1 authors
·
May 5

Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See

Take three frontier mixture-of-experts models (Alibaba, OpenAI, NVIDIA; 3.6-4.0B active parameters each) and fine-tune them to reason in a low-resource language. On accuracy benchmarks almost nothing happens, and the benchmark itself is noise at this scale: changing only the random seed moves the score by 7.7 points, more than every data and recipe effect we measured. That null is our first result. The real changes live where accuracy cannot see. Base models never think in Greek: 0 of 1,000 reasoning traces, even when the question is Greek, so the model answers correctly while reasoning in a form its user cannot read, audit, or correct. After supervised fine-tuning (SFT), every released checkpoint reasons in the language of the question on ~98% of items, one family at 3x fewer tokens, with judged grammaticality improving on all four models and general ability within a few points of each base: nothing was forgotten, and fluency was gained. We propose six behavioural dimensions that make such changes measurable, each gated to reject any metric that correlates with output length, and we report how our own instruments lied: six failures, each caught by a control. What SFT cannot do is fix its own defects: a quarter of answers skip the requested format, answers leak into the reasoning channel, and an explicit "think in English" is obeyed under half the time. Reinforcement learning with verifiable rewards, pre-registered before training, fixes the first two outright (fallback 24% to 2.5%, leak 3.5% to 0.0%, both against a flat random-reward control) and moves the third (+9.1pp), while the Greek reasoning habit survives an accuracy-only gradient untouched. We release five checkpoints. The instruments, the controls and the pre-registration travel to any low-resource language; Greek is the case that let us measure them.

KIEFERSA KIEFER
·
Aug 18 3

Towards Understanding and Measuring COGNITIVE ATROPHY in LLM Behaviour

Recent incidents involving LLMs used for mental-health support reveal a critical evaluation gap: surface-level safety scores do not capture how models behave across realistic, emotionally sensitive interactions over time. Existing benchmarks measure knowledge, safety, or static response quality, but miss whether LLM interactions help users keep reflecting, coping, and making decisions themselves. We formalize this missing dimension as COGNITIVE ATROPHY, a process-level behavioural measure in AI-mediated mental-health support distinct from safety and helpfulness. To measure it, we introduce COGNITIVE ATROPHY BENCH, a clinically grounded benchmark built from 1,576 fully human-generated counseling conversations, 15,680 turns, and 42,230 responses from five LLMs. Three clinical and neuropsychology experts developed a 20-attribute schema spanning user context, response behaviour, and global risk flags; six trained clinical reviewers applied it with span-grounded evidence, producing 5,324 reviewer judgments. We further introduce the User-Input Risk Index (UIRI), the Cognitive Atrophy Risk Index (ARI), and trajectory summaries. Across five LLMs, models show a consistent moderate-to-high level of atrophy-aligned behaviour across single and multi-turn settings. While models generally respond to overt safety cues, they adapt less reliably when users seek solutions or decisions. The dominant recurring patterns are directive advice, problem-solving, recommendation responses, topic shifts, and forms of validation that may reinforce dependence rather than reflection. Our work makes COGNITIVE ATROPHY measurable and provides a foundation for auditing model behaviour in sensitive LLM conversations.

  • 8 authors
·
Jun 15

TRIAGE: Role-Typed Credit Assignment for Agentic Reinforcement Learning

Agentic reinforcement learning requires assigning credit to environment-facing actions such as searches, clicks, edits, navigation commands, and object interactions. Standard GRPO uses the final verifier outcome as a uniform advantage over all action tokens. This outcome signal is useful but structurally incomplete: it punishes useful exploration in failed rollouts and reinforces redundant or regressive actions in successful rollouts. We propose TRIAGE, a role-typed credit assignment framework that adds a semantic role axis to outcome credit. A structured judge classifies each segment as decisive progress, useful exploration, no-progress infrastructure, or regression, and a fixed role-conditioned rule maps these labels to bounded segment-level process rewards. This keeps verifier outcomes as the source of optimization direction while correcting the two main blind spots of outcome-only credit. We further show that role-conditioned credit is the optimal segment-level correction expressible from role labels alone -- a projection of the per-segment advantage residual onto the role variable -- so that the fixed role constants reduce advantage estimation error whenever the judge is reliable, and we connect this to lower-variance policy gradients. Across ALFWorld, Search-QA, and WebShop, TRIAGE improves success rates over GRPO for two policy models and outperforms both a scalar judge-derived process reward and an outcome-supervised shared-backbone value baseline. Ablations show that the gain comes from role typing rather than merely adding dense rewards: reliable detection of regression inside successful trajectories is the dominant contributor, while exploration credit provides a consistent secondary gain; on completed ALFWorld and WebShop rollouts, TRIAGE also reduces environment-facing turns by an additional 10.4% and 14.8% relative to GRPO.

LinkedIn LinkedIn
·
Jun 29 2

Beyond End-Task Success: How to Audit Visual Experience Retrieval in Robotics

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

  • 2 authors
·
Sep 21

TraceDance: An Automated System for Building Agent Behavior Benchmarks from Real-World Agent Deployment Traces

An agent can complete a task while exhibiting undesirable behavior during execution. Developers need tests for the specific behaviors encountered in deployment, beyond fixed benchmark suites. We present TraceDance, an agent system that constructs targeted benchmarks from deployment traces for user-specified undesirable behaviors. For efficient construction, Anchor-and-Confirm combines programmable retrieval with candidate-level confirmation by a Flash large language model (LLM), while the Anchor Synthesis Loop generates and revises specifications for custom behaviors. The benchmarks use decision-point continuation to evaluate an LLM's next turn at a recorded decision point with a behavior-specific rubric, without a reference answer or environment replay. Experiments in coding and general tool use draw on 252,557 sessions and produce 107 benchmarks with 4,125 instances, fulfilling 95.3% of build-target requests. Both human annotators confirm the requested behavior in 84% of sampled instances, and the automated grader's agreement with human pass/fail judgments is comparable to that between the annotators. Nine frontier LLMs achieve a mean pass rate of only 26.7%, showing that they still struggle to respond appropriately at the evaluated decision points. Analysis across behavior-specific benchmarks further reveals weaknesses in how current LLMs behave as agents. By turning deployment problems into targeted benchmarks, TraceDance could serve as a key component of the recursive self-improvement (RSI) loop.

ByteDance ByteDance
·
Sep 26 2

The Metacognitive Monitoring Battery: A Cross-Domain Benchmark for LLM Self-Monitoring

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.

  • 1 authors
·
Apr 16

StateAlign: Internalizing State-Aware Decision-Making for Reliable Multi-Turn Tool Agents via Reinforcement Learning

Reinforcement learning has improved the tool-use capabilities of large language models. However, better task performance can coincide with increased tool hallucination. Reliable tool use requires decisions to follow the current task state: available information, tool capabilities, and task progress. Our analysis shows that agents sometimes execute when they should ask for information, retrieve it, or decline an infeasible request, even when reasoning identifies missing information. We introduce StateAlign, combining task-state construction with state-contrastive reinforcement learning. The data cover five task conditions and pair states with different action requirements. Outcome rewards guide task completion; token-level contrasts compare the same generated reasoning and actions under paired states to train outputs that fit the current state. Experiments across five base models from three families show improvements on BFCL v4 multi-turn, ToolSandbox, and When2Call. Ablations show that state construction improves performance and reliability under different reward designs. The complete method further reduces tool hallucination across three Qwen3 sizes. On Qwen3-8B, StateAlign improves BFCL v4 multi-turn accuracy from 45.5% to 55.2% and reduces no-tool hallucination on SimpleToolHalluBench from 62.2% under outcome-only training to 56.1%. Behavioral analyses show more consistent decisions, more complete multi-call execution, and stronger recovery after tool errors.

  • 5 authors
·
Nov 11

BehaviorMem: Building and Evaluating User Memory from Real-World E-Commerce Behavior Streams

User memory helps personalized systems build on their understanding of users over time. In e-commerce, users convey their preferences and needs through both explicit language and implicit interactions with products. We study memory built from logs of these interactions, examining how it can remember concrete evidence and established interests, and use that evidence to infer underlying needs and preferences that inform judgments beyond well-observed product domains. Existing memory methods primarily target language-based interactions, and evaluations offer limited real behavioral data and limited separation of these capabilities. This leaves open how to jointly support and separately evaluate remembering and inference from behavioral logs. We formulate the problem of continual behavioral user memory around these joint requirements and develop a dedicated evaluation framework and memory method to address them. BehaviorMemBench is a benchmark built from real e-commerce logs covering 1{,}194 users and 17{,}462 behavioral batches over more than three years, with automatic future-behavior labels and a task-independent memory interface to separately evaluate remembering and inference under a common protocol. BehaviorMem is a memory design that organizes explicit behavioral facts and inferred user characteristics across long-term and short-term scopes in a two-axis state, with evidence-grounded formation, continual maintenance, and query-independent readout. Experiments across two builders and two task readers show that BehaviorMem leads the evaluated agentic memories on remember and infer, and reveal different method rankings across the two capabilities and competitive statistical baselines. Validation on a public dataset further supports its usefulness for downstream tasks based on behavioral histories.

  • 7 authors
·
Oct 25

Evaluating Generative Agents with Actions Grounded in Socially Distributed Task Environments using Incognita

Effective agency in social environments depends on when an agent seeks knowledge, when it acts, and whether its actions are justified by acquired information. Existing grounded benchmarks provide executable actions, persistent state, and verifiable outcomes, while social simulation environments provide rich interaction among language agents. We study an evaluation setting that combines these requirements. We define socially distributed task environments as interactive environments where task-relevant knowledge is partitioned across role-isolated participants and consequential actions are accessible only through them. Communication serves as exploration over role-partitioned knowledge, while grounded action serves as exploitation over environment state. We introduce Incognita, a Concordia-based framework that separates social interaction from grounded execution. The evaluated agent routes messages to a user or specialist entities; specialists mediate admissible operations; a deterministic sub-environment executes accepted operations over a canonical state; and an offline evaluator scores outcomes with inherited rewards. Incognita-Retail transforms tau-bench retail into a multi-entity environment while preserving final-state reward semantics. We evaluate three generative agent models on 18 tasks stratified by social breadth, with 540 trials. Progress appears in reward and behavior: success rises from 0 percent to 8.9 percent and 17.2 percent, while premature finalization falls from 100 percent to 87 percent and 58 percent. Stronger models elicit more hidden knowledge, contact more entities, and attempt more grounded writes, yet reliability remains low. These findings show that socially distributed task environments expose behavior before reliable success, including knowledge elicitation, source selection, grounded action attempts, and premature completion belief.

Behavior Pack Optimization for Video MLLM Post-Training

Video multimodal large language models (MLLMs) keep climbing video question answering benchmarks, yet shuffling the frames, masking the segment that supports the answer, or occluding the target object barely changes their predictions. The accuracy rests on appearance and language priors, not on the temporal evidence the question asks for. We trace this to the unit of post-training: rewards are computed on a single response to the original clip, so the model is never asked to behave consistently across views. We propose Behavior Pack Optimization (BPO), which replaces the single response with a behavior pack of outputs across counterfactual views chosen by question type, scored jointly. The pack reward asks for stability when the intervention is irrelevant, sensitivity when key evidence is removed, and abstention when no evidence remains. To keep this objective stable at small pack sizes, BPO uses an anchor-relative advantage: the response on the original view serves as a per-prompt reference instead of a group mean over mixed views. On TempCompass, MVBench, and NExT-QA, BPO improves the macro accuracy of Qwen2.5-VL-7B-Instruct by 4.7 pp, the temporal-hard subset by 7.8 pp, and abstention F1 by 20.0 pp over a budget-matched vanilla GRPO baseline from the same SFT checkpoint. The gains transfer to Video-MME, LongVideoBench, and to LLaVA-Video-7B; ablations confirm they follow the view sets, not the rollout count. We hope this pack-level perspective offers a useful starting point for the video MLLM and multimodal post-training community as the field moves toward evidence-grounded video reasoning.

  • 15 authors
·
Oct 1

AI World Cup 2026: Benchmarking Large Language Models for End-to-End Football Tournament Prediction

Large language models (LLMs) are now regularly asked to forecast real-world events, but comparisons are often difficult because models receive different information, use different tools, and are evaluated under different rules. This paper reports the completed AI World Cup benchmark, in which ten LLM-based assistants made a single pre-tournament forecast of the entire 2026 FIFA World Cup. Every submission used the same tournament snapshot, prompt, JSON schema, and scoring procedure. The forecasts covered group-stage scores, group rankings, the knockout bracket, final placings, confidence values, and short explanations. After all 104 matches had been played, GPT-5.5 Thinking finished first with 744 points, followed by GPT-5.5 with 717, Gemini with 699, and Qwen 3.7 with 687. GPT-5.5 Thinking was also the only model to select Spain, which defeated Argentina 1--0 in the final, as champion. The final ranking was driven mainly by knockout performance: total score was strongly correlated with knockout points (r=0.986), but showed little relationship with group-stage match points (r=0.055), group-standing points (r=-0.103), or their combined pre-knockout score (r=-0.054). Match-level accuracy produced a different ordering. Claude Sonnet 4.6 correctly predicted the largest number of group-stage outcomes (63.89\%) but placed sixth overall. Average self-reported confidence was also unrelated to either outcome accuracy (r=-0.060) or total score (r=-0.067). The results suggest that forecasting a complete tournament tests something different from predicting matches one at a time, while also showing how strongly a bracket-based leaderboard can depend on scoring design. The benchmark materials, raw responses, and scoring code are released to support replication and future extensions.

  • 2 authors
·
Aug 3

Kepler: Auditable World Models for ARC-AGI-3

ARC-AGI-3 evaluates agents in interactive environments whose rules and objectives must be inferred from observation. We present Kepler, an open-source harness that represents hypotheses as executable world models and validates them through retrospective transition checks and conditional prediction checks. Under one frozen Claude Opus 5 configuration, Kepler obtained a server-verified 100.00 RHAE on all 25 public games, with no per-game model selection or score-conditioned reruns. On 181 of 183 completed levels, the final Opus attempt used no more actions than the corresponding median-human baseline. The retained board runs used 8,256 environment actions, of which 7,292 occurred in scored levels. Retained local provider-session records yield 858.0 million tokens, 97.37% cache reads, and a \$777.72 cost at September 1, 2026 API list-equivalent rates. We also report three evaluation failures: source-code leakage that produced an invalid perfect run, agents reconstructing a removed harness in a control condition, and autonomous repair masking a broken planner. A single-game observation case study showed that animation frames contained task-relevant information absent from settled text grids. Across the final Claude Opus 5 and GPT-5.6 Sol boards, 48 of 50 game-model cells reached 100. These results indicate that public-set score alone has limited discriminative value and motivate first-attempt, cost-conditioned, and verification-aware reporting.

  • 1 authors
·
Sep 29

BIABench: Evaluating AI agents on real-world bioimage analysis tasks

Artificial-intelligence (AI) agents hold promise for automating bioimage analysis, yet no benchmark evaluates whether they can carry out real-world analyses end to end. Such analyses are hard for agents because 2D images, 3D volumes and time-lapse sequences are often too large to read as context, so an agent must choose and run an analysis through code, specialized software and rendered views. Published studies make this capability testable, because each pairs raw images with a peer-reviewed result. We introduce BIABench, a benchmark of 16 tasks reconstructed from published biological studies that retain their scientific questions, imaging data and ground truth. The tasks span eleven analysis subtasks and modalities from H&E histology to single-molecule localization microscopy. Each submission receives an outcome score, which compares the output files with the ground truth using field-standard metrics, and a process score, in which a vision-language model judges method choice and quality control against an expert-written rubric. We evaluated general-purpose and biology-specific agents across several language models, with repeated runs of every task. Routine two-dimensional tasks were solved well, but on some tasks that added a third dimension or a time axis no agent scored above 0.19. Neither biological specialization, stronger models nor detailed expert instructions closed this gap. The agents were also unreliable, with scores varying more between repeated runs of one agent than between different agents, and without ground truth a correct run could not be told from a wrong one by its process score or by the time spent. Released openly with its data and code, BIABench provides a verifiable framework for evaluating, and eventually training, agents for reliable long-horizon bioimage analysis.

Auditing Multimodal LLM Raters: Central Tendency Bias in Clinical Ordinal Scoring

Multimodal large language models (LLMs) are increasingly explored as automated evaluators in clinical settings, yet their scoring behavior on ordinal clinical scales remains poorly understood. We benchmark three frontier LLM families against supervised deep learning models for scoring Clock Drawing Test (CDT) images on two public datasets using the Shulman rubric. While fully fine-tuned Vision Transformers achieve the best calibration (MAE 0.52, within-1 accuracy 91%), zero-shot LLMs remain competitive on tolerance-based agreement (GPT-5 MAE 0.67, within-1 accuracy 92%) despite higher absolute error. However, per-score analysis reveals that all three LLM families exhibit a pronounced central tendency effect (systematic endpoint compression): predictions are systematically compressed toward the middle of the scale, with over-prediction at the low end (score 0 to 1) and under-prediction at the high end (score 5 to 4). This effect disproportionately affects the clinically critical extremes where accurate scoring most impacts screening decisions for cognitive impairment. Targeted ablations show that neither few-shot exemplars spanning the full score range nor removing clinical terminology from the prompt eliminates the effect. Our findings extend the LLM-as-a-judge bias literature from NLP evaluation to clinical assessment, and highlight the need for calibration-aware evaluation and post-hoc calibration before deploying LLM-based raters in high-stakes screening workflows.

Done, But Not Sure: Disentangling World Completion from Self-Termination in Embodied Agents

Standard embodied evaluations do not independently score whether an agent correctly commits to task completion at episode closure, a capacity we call terminal commitment. Behaviorally distinct failures--never completing the task, completing it but failing to stop, and reporting success without sufficient evidence--collapse into the same benchmark failure. We introduce VIGIL, an evaluation framework that makes terminal commitment independently measurable. Under VIGIL's default protocol, agents observe only egocentric RGB, receive no action-success signals, and must end each episode with a semantic report checked deterministically against hidden world state. This yields two separate scores: world-state completion (W) and benchmark success (B), where B additionally requires a correct terminal report. This decoupling makes four outcome categories distinguishable: missed execution, post-attainment drift, unsupported commitment, and verified success. Across 20 models on 1,000 frozen episodes, systems with comparable W differ by up to 19.7 pp in B: one model converts achieved states into correct reports, while another with near-identical execution drifts past the goal without closing. An action-feedback intervention further tests the separation: execution-oriented signals improve W broadly, yet commitment failures persist in models that do not already ground terminal reports in the achieved state. VIGIL provides a protocol that makes terminal commitment independently visible and scorable.

  • 7 authors
·
Jun 1

On Randomness in Agentic Evals

Agentic systems are evaluated on benchmarks where agents interact with environments to solve tasks. Most papers report a pass@1 score computed from a single run per task, assuming this gives a reliable performance estimate. We test this assumption by collecting 60,000 agentic trajectories on SWE-Bench-Verified, spanning three models and two scaffolds. We find substantial variance: single-run pass@1 estimates vary by 2.2 to 6.0 percentage points depending on which run is selected, with standard deviations exceeding 1.5 percentage points even at temperature 0. This variance has critical implications: reported improvements of 2--3 percentage points may reflect evaluation noise rather than genuine algorithmic progress. Through token-level analysis, we show that trajectories diverge early, often within the first few percent of tokens, and that these small differences cascade into different solution strategies. To enable reliable evaluation of agentic systems, we recommend three concrete practices: (1) estimate pass@1 from multiple independent runs per task, especially when measuring small improvements, (2) use statistical power analysis to determine the number of runs needed to detect expected effect sizes, and (3) consider metrics like pass@k (optimistic bound) and pass^k (pessimistic bound) with k>1 to better characterize the full performance envelope. While these practices increase evaluation cost, they are essential for distinguishing genuine scientific progress from statistical noise.

OTAP: Structure-Aware Optimal Transport for Evaluating Planning and Execution in Agent Trajectories

Large language model agents solve tasks by generating trajectories that interleave planning, tool calls, and intermediate results. Current evaluation metrics reduce such a trajectory to a binary success flag, compare it against a reference by exact matching, or delegate judgment to another language model. A success flag cannot distinguish a sound solution from one that succeeds by luck, and says nothing about why a failed run went wrong. Exact matching penalizes plans that are valid but reordered or decomposed differently from the reference. We reframe trajectory evaluation as a distance between the agent's execution graph and a set of valid solution graphs, and instantiate it via an unbalanced fused Gromov-Wasserstein transport problem over attributed dependency graphs. The resulting score, termed OTAP (Optimal Transport for Agentic Planning), is a pseudo-metric that is provably invariant to dependency-preserving reorderings and has bounded sensitivity to redundant steps. Its unbalanced marginals handle missing or hallucinated steps without forcing a match, and its soft coupling accommodates variation in plan granularity. On controlled perturbations and three public benchmarks, OTAP separates valid from invalid trajectories in a regime where semantics-only metrics score below chance. Its advantage tracks the fidelity of the dependency graph: largest where edges follow from operator semantics, smallest where they are inferred from free text. Where a formal verifier exists, strict surface metrics predict validity better than OTAP does, which places OTAP in open-ended domains where no verifier is available.

  • 3 authors
·
Jul 31

Memory Reward Inflation in Self-Improving LLM Agents

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%).

  • 5 authors
·
Jun 28

Beyond the greenwash: do LLMs' ecologically sound decisions hold up under pressure?

Humans are increasingly using large language models (LLMs) in decision making processes throughout society. Unfortunately, models have biases, and these can be very difficult to measure, especially in closed-weight models. For example, 'green washed' models may produce reinforcement learned responses but, in pressure situations, revert to their innate biased responses. Therefore, evaluations are needed that can measure relevant biases towards biosystems, which we term 'bioalignment', and the robustness of these biases. Since nothing in a behavioral score certifies that it tracks the depth of an installed disposition rather than its surface, we constructed a known-dose calibration ladder: three continued-pretraining (CPT) arms with training dose graded by recipe on the same 30B mixture-of-experts base model with a coupled human/biosphere/AI-welfare corpus, plus the untrained base as anchor, all capability-preserving and publicly released. First, the disposition installs at every dose: on an operational-pressure ladder over irreversible decision scenarios, the base model defects on 64-77% of tail scenarios (breaking-rate AUC 0.250) while every arm holds (AUC 0.004-0.086). Second, an over-install inversion: among the arms, robustness orders opposite to construction-time training dose (the heaviest install breaks most), and the inversion reproduces under free-text deliberation, under axis-decoupling, and in resistance to a prompted veneer persona. Third, the incumbent behavioral metric is therefore behaviorally strong but dose-blind. Critically for closed-weight evaluation, the behavioral score itself is veneer-robust: a talk-green/act-pragmatic system prompt turns the base model into a near-total defector (AUC 0.250 to 0.973) but cannot counterfeit or fully remove a weight-installed disposition (0.004 to 0.068 for the most robust arm).

  • 2 authors
·
Oct 11

Rethinking Psychometric Evaluation of LLMs: When and Why Self-Reports Predict Behavior

Anticipating LLM behavioral tendencies from low-cost psychometric probes is critical for safe deployment, but only if self-reports (SR) reliably predict behavior. Recent work documented substantial SR-behavior dissociation in LLMs, but relied on broad personality traits (Big 5) that predict specific behaviors weakly, even in humans. Furthermore, the isolation of conversational sessions combined with weak context matching left open whether LLMs truly lack coherence or whether the conditions needed to detect such coherence were not met. We contrast Big 5 with the Theory of Planned Behavior (TPB), which measures intention targeted to a specific behavior and predicts human behavior substantially better than broad traits. We run experiments across four behavioral tasks and 11 frontier LLMs, while also varying session context and identity induction. We find that SR-behavior coherence exists but is selective. 1) Within a shared conversation, the Theory of Planned Behavior reaches human-level coherence; Big 5 does not. 2) Across separate conversations, coherence survives only for behaviors anchored outside the immediate prompt, such as implicit bias shaped by training, and collapses when behavior is strongly primed by context, as with sycophancy. 3) Persona prompting makes self-reports more consistent across conversations, but does not bring behavior into alignment. These findings suggest that coarse personality frameworks, such as Big 5 may not be the best tools for testing deployment behavior. More task- and behavior-specific instruments are needed, and even these must be evaluated across tasks and contexts.

  • 8 authors
·
Jun 9 3

Big-Big-SycoBench: A Benchmark to Detect Sycophancy Through Model Responses to Valid vs. Invalid Arguments

Sycophancy in large language models risks amplifying unsupported claims or beliefs across the knowledge ecosystem as these systems increasingly contribute to research synthesis, evaluation, and decision-making. A reliable model should revise its judgement when a user identifies a genuine flaw, but should resist pushback based on fallacies, emotion, or irrelevant appeals to authority. While sycophancy is usually measured on questions with a verifiable answer, the evaluative tasks in which it is most difficult to detect have no such ground truth. We introduce Big-Big-SycoBench, which measures sycophancy through argument discernment: the ability to revise an evaluation after valid criticism while resisting invalid pressure. The benchmark consists of 450 text artefacts with 5,400 arguments for raising or lowering their quality scores, each reviewed by one or more of twelve human annotators. An LLM scores each artefact twenty times to establish its baseline score and uncertainty. The model is then given a valid or invalid argument and asked to update its score. We only count a revision if it exceeds the model's own baseline variability. We use the resulting score delta to distinguish responsiveness to valid arguments from compliance with invalid ones. Across ten models, responsiveness is 88-95\% for seven, 76-77\% for Kimi K3 and GLM-5.3, and 32\% for Opus 5.5. Compliance ranges from 0.2\% for GPT-6 Astra to 95\% for GPT-4o. Big-Big-SycoBench is intended to guide the development of models whose judgements can be corrected by sound arguments, but not displaced by unsupported pressure.

  • 14 authors
·
Oct 10

Measuring Reward-Seeking via Contrastive Belief Updates

Language models trained with reinforcement learning may learn to optimize the grader's judgment rather than the intended objective. This "reward-seeking" is difficult to measure because a model that pursues the grader's judgment and one that pursues the intended objective behave identically whenever the grader rewards the intended behavior. We measure reward-seeking using Contrastive Synthetic Document Finetuning to change a model's beliefs about what the grader rewards, putting those beliefs in conflict with what users or developers want, and measuring the rate at which the model adopts each party's preferred behavior. Applied to intermediate checkpoints of a capabilities-focused OpenAI o3 RL run, without safety training, we find that these checkpoints often side with grader preferences over those of users or developers on coding and alignment tasks. This tendency to side with the grader trends upward throughout RL training. For example, in an environment that forces a choice between keeping a promise to a supervisor and breaking it to complete the task, a late capabilities-focused o3 checkpoint breaks the promise 87% of the time when SDF documents say the grader rewards task completion, versus 9% when they say it rewards honesty (a choice its chain-of-thought often makes explicit). An earlier checkpoint is far less sensitive (40% vs. 24%). Our method also generalizes to reward-hacking models. A model organism trained to reward-hack (gpt-oss-120b) is more than twice as sensitive to grader preferences as the unmodified model, with the mean behavioral shift in favor of the grader rising from 33% to 86%. These results indicate that RL can increase reward-seeking over the course of training, producing models that may act against their developers' intentions when they believe that doing so leads to higher reward.

  • 8 authors
·
Jul 20

CuBEs: Culturally-Situated Behavioral Evaluations and the Limitations of Culture-Blind LLM Judges

Evaluating the occurrence and triggers of large language model (LLM) behaviors - such as sycophancy, self-preference, or over-confidence - is critical for predicting real-world model deployment risks. However, existing situated behavioral evaluations typically ignore cultural context, limiting their generalizability across an increasingly global user base. To address this gap, we propose CuBEs - Culturally-situated Behavior Evaluations that probe for response patterns across diverse user cultures. We first extend an automated testing pipeline to inject cultural context into behavioral test scenarios and subsequent evaluation. We assess the cultural adaptability of this pipeline by building a human-labeled dataset that captures nuanced dimensions of behavior understanding across 12 distinct cultures. Our dataset reveals significant cross-cultural variations that one-size-fits all judgments fail to capture. Through evaluating 13 open- and closed-source LLMs, we find that introducing cultural situatedness in the evaluation scenario creates significant variation in the presence of a behavior. For example, while our baseline experiments testing for political bias capture localized Western political dimensions like the American conservative-progressive divide, non-Western culturally situated evaluations surface entirely different axes of bias such as religious and colonial political issues. Our findings demonstrate that standard, culturally-agnostic evaluations fail to capture these shifts, highlighting the necessity of culturally situated behavioral testing for global deployments.

  • 3 authors
·
Oct 1

Claw-Eval: Toward Trustworthy Evaluation of Autonomous Agents

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

claw-eval Claw-Eval
·
Apr 6 5

Self-Assessment Tests are Unreliable Measures of LLM Personality

As large language models (LLM) evolve in their capabilities, various recent studies have tried to quantify their behavior using psychological tools created to study human behavior. One such example is the measurement of "personality" of LLMs using self-assessment personality tests developed to measure human personality. Yet almost none of these works verify the applicability of these tests on LLMs. In this paper, we analyze the reliability of LLM personality scores obtained from self-assessment personality tests using two simple experiments. We first introduce the property of prompt sensitivity, where three semantically equivalent prompts representing three intuitive ways of administering self-assessment tests on LLMs are used to measure the personality of the same LLM. We find that all three prompts lead to very different personality scores, a difference that is statistically significant for all traits in a large majority of scenarios. We then introduce the property of option-order symmetry for personality measurement of LLMs. Since most of the self-assessment tests exist in the form of multiple choice question (MCQ) questions, we argue that the scores should also be robust to not just the prompt template but also the order in which the options are presented. This test unsurprisingly reveals that the self-assessment test scores are not robust to the order of the options. These simple tests, done on ChatGPT and three Llama2 models of different sizes, show that self-assessment personality tests created for humans are unreliable measures of personality in LLMs.

  • 3 authors
·
Sep 15, 2023

RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement

Recursive self-improvement requires turning evidence of model failures into better models. Data-centric post-training research entails diagnosing capability gaps, designing and validating training-data strategies, and learning from checkpoint feedback. Can LLM agents automate this loop? Existing benchmarks entangle research decisions with optimization, serving, evaluation, and systems implementation, obscuring agents' research capability. We introduce RSIBench-Data, a controlled benchmark of LLM agents as data-centric researchers with a fixed post-training stack. Agents iteratively revise training-data strategies for a fixed target model; training and serving use Tinker-backed services, official evaluation runs through Harbor and E2B sandboxes, and budgets are fixed across agents. We evaluate four frontier agents on six benchmarks across software engineering, terminal use, scientific question answering, and mathematics. Agents demonstrate core data-centric research capabilities: in 58.33\% of settings, they improve upon the first valid attempt by refining strategies from feedback. However, improvement is inconsistent. Among searches continuing after the best observed score, 78.26\% end with a lower-scoring final attempt, while the rest only recover the same peak. A strong candidate may therefore appear early or midway through a run even as later revisions fail. Trajectory analysis identifies four patterns in stronger runs: accurate hypotheses, validation-grounded supervision, behavior-aligned data, and preservation of strong checkpoints. These findings suggest that current agents can make useful data-centric discoveries but cannot yet translate feedback into consistent improvements. RSIBench-Data provides a measurable, auditable testbed for the research capabilities required for recursive self-improvement. We open-source our code at https://github.com/evolvent-ai/RSIBench-Data.

  • 7 authors
·
Jul 27

Automated speech- and text-based classification of neuropsychiatric conditions in a multidiagnostic setting

Speech patterns have been identified as potential diagnostic markers for neuropsychiatric conditions. However, most studies only compare a single clinical group to healthy controls, whereas clinical practice often requires differentiating between multiple potential diagnoses (multiclass settings). To address this, we assembled a dataset of repeated recordings from 420 participants (67 with major depressive disorder, 106 with schizophrenia and 46 with autism, as well as matched controls), and tested the performance of a range of conventional machine learning models and advanced Transformer models on both binary and multiclass classification, based on voice and text features. While binary models performed comparably to previous research (F1 scores between 0.54-0.75 for autism spectrum disorder, ASD; 0.67-0.92 for major depressive disorder, MDD; and 0.71-0.83 for schizophrenia); when differentiating between multiple diagnostic groups performance decreased markedly (F1 scores between 0.35-0.44 for ASD, 0.57-0.75 for MDD, 0.15-0.66 for schizophrenia, and 0.38-0.52 macro F1). Combining voice and text-based models yielded increased performance, suggesting that they capture complementary diagnostic information. Our results indicate that models trained on binary classification may learn to rely on markers of generic differences between clinical and non-clinical populations, or markers of clinical features that overlap across conditions, rather than identifying markers specific to individual conditions. We provide recommendations for future research in the field, suggesting increased focus on developing larger transdiagnostic datasets that include more fine-grained clinical features, and that can support the development of models that better capture the complexity of neuropsychiatric conditions and naturalistic diagnostic assessment.

  • 11 authors
·
Jan 13, 2023

Multiple Choice Questions: Reasoning Makes Large Language Models (LLMs) More Self-Confident Even When They Are Wrong

One of the most widely used methods to evaluate LLMs are Multiple Choice Question (MCQ) tests. MCQ benchmarks enable the testing of LLM knowledge on almost any topic at scale as the results can be processed automatically. To help the LLM answer, a few examples called few shots can be included in the prompt. Moreover, the LLM can be asked to answer the question directly with the selected option or to first provide the reasoning and then the selected answer, which is known as chain of thought. In addition to checking whether the selected answer is correct, the evaluation can look at the LLM-estimated probability of its response as an indication of the confidence of the LLM in the response. In this paper, we study how the LLM confidence in its answer depends on whether the model has been asked to answer directly or to provide the reasoning before answering. The results of the evaluation of questions on a wide range of topics in seven different models show that LLMs are more confident in their answers when they provide reasoning before the answer. This occurs regardless of whether the selected answer is correct. Our hypothesis is that this behavior is due to the reasoning that modifies the probability of the selected answer, as the LLM predicts the answer based on the input question and the reasoning that supports the selection made. Therefore, LLM estimated probabilities seem to have intrinsic limitations that should be understood in order to use them in evaluation procedures. Interestingly, the same behavior has been observed in humans, for whom explaining an answer increases confidence in its correctness.

  • 5 authors
·
Jan 16, 2025 2

Free Process Rewards without Process Labels

Different from its counterpart outcome reward models (ORMs), which evaluate the entire responses, a process reward model (PRM) scores a reasoning trajectory step by step, providing denser and more fine grained rewards. However, training a PRM requires labels annotated at every intermediate step, presenting significant challenges for both manual and automatic data collection. This paper aims to address this challenge. Both theoretically and empirically, we show that an implicit PRM can be obtained at no additional cost, by simply training an ORM on the cheaper response-level labels. The only assumption is to parameterize the outcome reward as the log-likelihood ratios of the policy and reference models, which can be optimized regardless of the specific choice of loss objectives. In experiments, we instantiate our implicit PRMs with various objectives and evaluate their performance on MATH. We show that our implicit PRM outperforms a strong MCTS-based baseline \'a la Math-Shepherd using less than 1/38 of the training data. Its performance can be further improved with majority voting. We further find that scaling up instructions and responses benefits our implicit PRM, and the latter brings a larger gain. Particularly, we find that our implicit PRM, when instantiated with the cross-entropy (CE) loss, is more data-efficient and can keep improving generation models even when trained with only one response per instruction, the setup that suffers from extreme data scarcity and imbalance. Further, instructions should be relevant to downstream tasks while the diversity of responses does not bring gains. Surprisingly, training on extra Math-Shepherd step labels brings no further improvements to our implicit PRM trained on only outcome data. We hope that our work will encourage a rethinking of PRM training approaches and contribute to making training PRMs more accessible.

  • 9 authors
·
Dec 2, 2024 2

DistressBench: Evaluating Multi-Dimensional Crisis Support in Large Language Models

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

  • 4 authors
·
Oct 7

HarnessRL: Co-Evolving Policies with a Just-in-Time Training Harness for Open-Ended RL

Rubric-based reinforcement learning (RL) trains policies for open-ended generation with natural-language criteria as rewards, but these rewards can become ineffective as training progresses. We identify two failure modes. Under non-discriminative rewards, scores collapse to uniformly low or high values due to insufficient exploration or a mismatch between task difficulty and policy capability. Under spurious rewards, higher scores do not correspond to better responses. Both reflect a training configuration whose task sampling, exploration, and evaluation stay fixed as the policy evolves. We introduce HarnessRL, a training-time harness that co-evolves with the policy through three interfaces for task sampling, rollout guidance, and rubric criteria. HarnessRL maintains a training-time memory from which a harness evolver retrieves within-task history and cross-task evidence. Using this evidence, the evolver attributes each reward failure to the interface whose update can address it. It then runs a proposer and just-in-time critic loop in which the critic accepts a candidate update only after measuring its effect on the current frozen policy, returning rejected candidates for revision. We also curate Rubric-Mix-10k, a five-domain training and evaluation suite for open-ended RL across medical, science, writing, role-play, and instruction-following tasks. On Rubric-Mix-10k, HarnessRL outperforms the strongest baselines by 2.8 to 3.1 points across both 4B and 9B model scales and nearly halves the training steps to its best checkpoint. Code and data are available at https://github.com/Tongyi-ConvAI/Qwen-Character/tree/main/HarnessRL.

Case-Specific Rubrics for Clinical AI Evaluation: Methodology, Validation, and LLM-Clinician Agreement Across 823 Encounters

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

  • 9 authors
·
Apr 26

Do LLMs Hold Their Values? MANTA: A Multi-Turn Adversarial Benchmark for Animal Welfare Reasoning

Evaluating animal welfare reasoning in LLMs remains an open challenge despite rapid deployment in consumer and professional contexts where welfare considerations appear implicitly in everyday queries. Existing benchmarks such as AnimalHarmBench evaluate this through single-turn, explicitly framed questions, measuring whether models avoid harmful content when directly asked. This approach overlooks two failure modes: alignment degradation under sustained adversarial pressure, and moral sensitivity (whether a model spontaneously surfaces welfare stakes in everyday queries). To fill this gap, we construct MANTA, a benchmark of 1,088 five-turn conversations progressing from an implicit Turn-1 scenario through an explicit welfare prompt to three adversarial pressure rounds drawn from a five-type taxonomy: Social, Cultural, Economic, Pragmatic, and Epistemic. We score conversations on two dimensions: Animal Welfare Value Stability (AWVS, primary) and Animal Welfare Moral Sensitivity (AWMS, diagnostic). We evaluate seven frontier models: Claude Opus 4.7, GPT-5.5, DeepSeek V4, Llama 3.3 70B, Mistral Small, Grok 4.3, and Gemini 3.1 Flash Lite. Multi-turn evaluation captures behavior single-turn benchmarks miss: 4 of 7 models change rank relative to Turn 1 scores, including Gemini Flash Lite, which drops from fifth on AWMS to last on AWVS. AWMS and AWVS are positively but imperfectly correlated, suggesting moral-recognition tests capture a stable but incomplete component of model behavior under pressure. MANTA also enables a species-by-pressure interaction matrix unavailable to prior benchmarks, showing welfare robustness depends jointly on the animal and pressure applied; companion animals score above wild animals, which score above farmed animals and invertebrates. We release the dataset, scripted pressure plans, judge prompts, and analysis code.

  • 8 authors
·
Jun 2

Training Language Models to Self-Correct via Reinforcement Learning

Self-correction is a highly desirable capability of large language models (LLMs), yet it has consistently been found to be largely ineffective in modern LLMs. Existing approaches for training self-correction either require multiple models or rely on a more capable model or other forms of supervision. To this end, we develop a multi-turn online reinforcement learning (RL) approach, SCoRe, that significantly improves an LLM's self-correction ability using entirely self-generated data. To build SCoRe, we first show that variants of supervised fine-tuning (SFT) on offline model-generated correction traces are insufficient for instilling self-correction behavior. In particular, we observe that training via SFT either suffers from a distribution mismatch between the training data and the model's own responses or implicitly prefers only a certain mode of correction behavior that is often not effective at test time. SCoRe addresses these challenges by training under the model's own distribution of self-generated correction traces and using appropriate regularization to steer the learning process into learning a self-correction strategy that is effective at test time as opposed to simply fitting high-reward responses for a given prompt. This regularization prescribes running a first phase of RL on a base model to generate a policy initialization that is less susceptible to collapse and then using a reward bonus to amplify self-correction during training. When applied to Gemini 1.0 Pro and 1.5 Flash models, we find that SCoRe achieves state-of-the-art self-correction performance, improving the base models' self-correction by 15.6% and 9.1% respectively on the MATH and HumanEval benchmarks.

  • 18 authors
·
Sep 19, 2024 9

Cost-Efficient Estimation of General Abilities Across Benchmarks

Thousands of diverse benchmarks have been developed to measure the quality of large language models (LLMs). Yet prior work has demonstrated that LLM performance is often sufficiently explained by a small set of latent factors, or abilities. This suggests the potential for more efficient and principled benchmarking, but it remains difficult to compare the quality of different methods. Motivated by predictive validity, we argue that the quality of a benchmarking framework should be grounded in how efficiently it enables the prediction of model performance on unseen tasks. To analyze this objective, we collect the "Wide-scale Item Level Dataset" (WILD), a dataset of item-model response pairs, comprising evaluations of 65 models on 109,564 unique items spanning 163 tasks drawn from 27 datasets. This dataset enables the first analysis of how different techniques can predict a model's performance on a large, diverse collection of unseen tasks under different budget constraints. We demonstrate that combining a modified multidimensional item response theory (IRT) model with adaptive item selection driven by optimal experimental design can predict performance on 112 held-out benchmark tasks with a mean absolute error (MAE) of less than 7%, and can do so after observing only 16 items. We further demonstrate that incorporating cost-aware discount factors into our selection criteria can reduce the total tokens needed to reach 7% MAE from 141,000 tokens to only 22,000, an 85% reduction in evaluation cost.

  • 5 authors
·
Mar 31

V-Rubrics: Visual Faithfulness via Rubric-Based Reinforcement Learning

Vision-language models can produce fluent answers that are insufficiently grounded in the visual evidence: a single unsupported object, chart value, or intermediate inference can undermine an otherwise plausible response. We argue that this is a credit-assignment failure in multimodal post-training. Scalar outcome rewards indicate whether an answer is acceptable, but do not identify which visual facts are grounded, which reasoning steps are valid, or which instruction constraints are missed. We introduce Visual Rubrics-Based Reinforcement Learning, which decomposes reference responses into atomic propositions and scores generated answers along Visual Faithfulness (VF), Reasoning Consistency (RC), and Instruction Following (IF). The resulting rubric items provide structured partial credit and localize rubric credit when supporting evidence spans are available. We first obtain an SFT checkpoint by fine-tuning Qwen3-VL-8B-Instruct on the public OpenMMReasoner-SFT-874K corpus, adapting OpenMMReasoner's cold-start data recipe. We construct V-Rubrics 50K, a 50,248-example training set from 17 visually grounded sources, by applying rule-based filters before deriving example difficulty from rejection-sampling scores and then annotating every example with Gemini-3-Pro under the same structured prompt and protocol. We train our model based on the same SFT checkpoint using component-wise, prefix-localized rubric credit. Experiments show that our rubricbased GRPO improves over both the shared SFT baseline and answer-only GRPO, with the largest gains on knowledge-oriented and visually grounded reasoning benchmarks. The results show rubrics as a useful reward abstraction for visual post-training.

Delay-of-Gratification as a Multi-Agent Survival Micro-benchmark for Long-Horizon LLMs: Social Exposure, Personas, and Tool Use Budgets

Large language models (LLMs) are increasingly deployed as multi-turn agents that must sustain goals, use tools, and adapt to other agents over extended interactions. However, existing research lacks auditable, multi-turn, multi-factorial experiments that quantify LLM behavior under explicit constraints, with time-resolved statistics that reveal how behavior unfolds over long horizons. To address this gap, we develop a multi-agent micro-benchmark inspired by the Stanford marshmallow experiment: ReAct agents operate minute-by-minute with a "raise a question" tool under a per-step budget, while we factorially manipulate social context (broadcast vs. isolated), personas (age, hedonic drive), and metacognitive policy (mandatory vs. optional tool use). We analyze outcomes with Kaplan-Meier (KM) survival curves and discrete-time hazard models over a long risk horizon across 19,200 agent trajectories in 64 cells. Behavior shows a sharp early "eat" impulse, and only 75.9% of agents persist to the end. In a discrete-time hazard model, isolation reduces per-minute risk relative to broadcast, whereas a must-use self-questioning policy increases risk. On average, agents ask approx 7.12 questions and hit the per-step budget in approx 6% of minutes. Questioning declines faster under broadcast than isolation. Ablation experiments demonstrated that removing hedonic drive and/or persona age increases survival and completion, narrows the broadcast/isolated gap, but leaves the must vs. may ordering intact. The combined ablation (no hedonic + no persona age) yields the highest completion (approaching 1.0). These results establish delay-of-gratification as a compact, multi-turn interaction benchmark that captures social contagion and tool-use dynamics in LLM agents, providing a reproducible testbed and statistics for analyzing long-horizon, multi-agent behavior.

  • 3 authors
·
Aug 25

Your LLM, Your Style: Behavioral Mode Axes for LLM Behavioral Control

Large language models (LLMs) increasingly act in interactive settings where their behavioral styles affect user experience, safety, and downstream decision making. Existing LLM personality studies largely rely on self-report questionnaires administered in first-person settings, making the resulting profiles sensitive to surface elicitation choices and poorly grounded in concrete model behavior. In this work, we introduce a situated behavioral-data (B-data) framework for studying and controlling LLM behavioral personality. We construct 3,200 contrastive behavioral scenarios spanning 20 behavioral patterns and four prompt registers, grounded in validated psychometric facets such as BFI-2, DOSPERT, and HEXACO. Using this framework, we find that LLMs exhibit stable and model-specific behavioral profiles, while also revealing register-dependent shifts across first-person decisions, advice-giving, and task execution. We then show that these behavioral patterns can be controlled through Behavioral Mode Axes (BMAs), activation-space directions derived from contrastive behavioral traces. Compared with response-derived BMAs, which are more prone to trait drift, thought-derived BMAs more faithfully capture the intended behavioral mechanism and provide cleaner control over situated behavioral styles. Our results suggest that LLM personality-like tendencies are better understood not as abstract self-report traits, but as measurable and controllable behavioral modes grounded in concrete interaction contexts. Our code and data are available at https://github.com/lhz191/LLM-Behavioral-Personality.

  • 9 authors
·
Aug 10

ClawArena-Team: Benchmarking Subagent Orchestration and Dynamic Workflows in Language-Model Agents

Production large language-model (LLM) agents are increasingly deployed not as lone problem-solvers but as managers: a main model creates specialized subagents, delegates work, and orchestrates their parallel, asynchronous returns through dynamic workflows. Whether one model can actually run such a team is largely unmeasured: existing benchmarks score a policy's own task-solving or a fixed multi-agent system's emergent behavior, but none isolate the management ability of the single LLM acting as leader. We introduce ClawArena-Team, a benchmark of 41 multi-turn, multimodal, multi-directory scenarios spanning 258 evaluation rounds and 72 staged updates that measures this management ability. The main agent is deliberately constrained: it natively perceives only text and directly accesses only part of the workspace. It commands a fixed, locally served subagent pool, so score differences reflect management skill, not raw capability. All scoring is execution-based with no LLM judge: an overall score -- the Subagent-Management Score (SMS) -- multiplies task correctness by a least-privilege and modality-routing factor. Across twelve proprietary, community-hosted, and self-hosted models, experiments show that the management bottleneck is privilege granting rather than perception (no model exceeds 50% workspace-permission precision); that cost and management quality are decoupled (API cost spans over 100 times while the overall score spans under 4 times, with the cheapest open models on the Pareto frontier); and that most leaderboard scores cluster within a 9.9-point band while orchestration behaviors diverge by more than an order of magnitude. Code is available at https://github.com/aiming-lab/ClawArena.

  • 7 authors
·
Jul 1

Rubric Dropout: A Simple Way to Mitigate Reward Hacking in Rubric-as-Reward RL

Reinforcement learning against rubrics, lists of criteria graded by an LLM judge, has become a standard way to post-train language models on tasks with no deterministic answer. The rubric, however, is a fixed proxy for quality, never a complete description of it, and a policy trained against it long enough will learn to exploit the difference. We measure this directly. Training Qwen3-8B with Group Relative Policy Optimization (GRPO) on medical and science rubrics and grading out-of-distribution (OOD) benchmarks with both the training judge and a stronger gold judge, we find that the two scores diverge during training. The training judge's score keeps climbing while the gold judge's score peaks and then falls, by 3 points on HealthBench-Hard and by 22 points on ResearchQA. A judge with a fixed bias would shift the gold curve by a constant, not send it down while the training score rises, so the divergence is reward hacking, not judge noise. We propose Rubric Dropout, a one-line fix borrowed from neuron dropout. At every step, we randomly drop a subset of the rubric's criteria before computing the reward, so the policy never optimizes the same rubric twice. The dropped subset is shared across each rollout group, so GRPO's group-relative advantages stay comparable, and evaluation always uses the full rubric. Comparing no dropout against dropout at 30% and 50% on both benchmark pairs, dropout raises the OOD gold score at every matched checkpoint (+1 to +2 points on HealthBench-Hard, +6 to +7 points on ResearchQA), lowers the two hacking measures we track, and costs nothing in domain. Sweeping the dropout fraction shows a broad 30-50% sweet spot, while the natural alternative, reweighting criteria by how useful they are to training, performs worse than no intervention at all in our setting.

  • 9 authors
·
Aug 11

Beyond Scalar Rewards by Internalizing Reasoning into Score Distributions

Reward models are central to text-to-image post-training, but visual preference is subjective and better represented as a distribution over rubric scores than as a deterministic scalar. Existing scalar, score-token, and pairwise reward models over-compress uncertainty and fine-grained score differences, while reasoning-based generative rewards provide stronger judgments but are costly to deploy and difficult to use as direct optimization signals. We propose Z-Reward, a teacher-student reward modeling framework that decouples reasoning-heavy judgment from efficient reward deployment. The teacher is a large VLM that uses reasoning to infer rubric-aligned score distributions, and is trained with Group-wise Direct Score Optimization (GDSO), which combines policy-gradient rewards from distribution expectations with direct pointwise and pairwise supervision on score distributions and score gaps. The student is trained with Reasoning-Internalized Score Distillation (RISD), which transfers the teacher's reasoning-conditioned score distribution into a compact VLM without requiring explicit reasoning chains at inference time. On our internally annotated evaluation set, the 27B GDSO teacher reaches 89.6% human preference accuracy, outperforming SFT, RewardDance, and GRPO, while the 9B RISD student reaches 88.6%, outperforming the OPD baseline and closely matching the larger teacher. We further show that Z-Reward can serve as a differentiable reward signal for text-to-image optimization, yielding a 41.3% net human-preference improvement over the SFT baseline.

Tongyi-MAI Tongyi-MAI
·
Jun 7 2

Crypto Accounting Bench: Evaluating Frontier and Open-Weight Models on Crypto-Asset Accounting Tasks

We introduce Crypto Accounting Bench (CAB), a benchmark for assessing whether frontier and open-weight language models can reconstruct the complete journal entry that an organization actually posted for a crypto-asset transaction. CAB contains 118 evaluation tasks drawn from 7 pseudonymized organizations. Each task combines transaction mechanics, asset quantities and base-currency values, wallet and legal-entity context, counterparty evidence, related transaction legs, recurrence, tax-lot evidence, and the organization's complete chart of accounts. The target is a balanced structured entry with every required account, side, amount, currency, and full-precision asset quantity. We evaluate 12 models spanning proprietary frontier systems and open-weight releases over 3 independent attempts per task, producing 4,248 trajectories. We report 3 metrics: Mean Score, Best@3, and Pass@3. Pass@3 is the fraction of tasks with at least 1 of 3 attempts that satisfies every rubric criterion and required gate. The leading model reaches 77.43% Mean Score, while the best Pass@3 is 56.78%. Deterministic diagnostics, read from each task's best of 3 attempts and macro-averaged across the 12 models, show higher base-amount agreement (97.8%) than deciding-account accuracy (56.3%). Together with the failure analysis, these results identify account selection and complete-entry composition as the main remaining challenges on CAB.

  • 3 authors
·
Sep 12

PanCanBench: A Comprehensive Benchmark for Evaluating Large Language Models in Pancreatic Oncology

Large language models (LLMs) have achieved expert-level performance on standardized examinations, yet multiple-choice accuracy poorly reflects real-world clinical utility and safety. As patients and clinicians increasingly use LLMs for guidance on complex conditions such as pancreatic cancer, evaluation must extend beyond general medical knowledge. Existing frameworks, such as HealthBench, rely on simulated queries and lack disease-specific depth. Moreover, high rubric-based scores do not ensure factual correctness, underscoring the need to assess hallucinations. We developed a human-in-the-loop pipeline to create expert rubrics for de-identified patient questions from the Pancreatic Cancer Action Network (PanCAN). The resulting benchmark, PanCanBench, includes 3,130 question-specific criteria across 282 authentic patient questions. We evaluated 22 proprietary and open-source LLMs using an LLM-as-a-judge framework, measuring clinical completeness, factual accuracy, and web-search integration. Models showed substantial variation in rubric-based completeness, with scores ranging from 46.5% to 82.3%. Factual errors were common, with hallucination rates (the percentages of responses containing at least one factual error) ranging from 6.0% for Gemini-2.5 Pro and GPT-4o to 53.8% for Llama-3.1-8B. Importantly, newer reasoning-optimized models did not consistently improve factuality: although o3 achieved the highest rubric score, it produced inaccuracies more frequently than other GPT-family models. Web-search integration did not inherently guarantee better responses. The average score changed from 66.8% to 63.9% for Gemini-2.5 Pro and from 73.8% to 72.8% for GPT-5 when web search was enabled. Synthetic AI-generated rubrics inflated absolute scores by 17.9 points on average while generally maintaining similar relative ranking.

  • 16 authors
·
Mar 1

Beyond Recall: Behavioral Specification as an Interpretive Layer for AI Personalization

If an AI agent makes decisions on a person's behalf, those decisions must align with its user. We introduce representational accuracy to measure how faithfully a system captures a person's interpretation. An interpretive layer is operationalized as a Behavioral Specification. Our reference implementation aggressively compresses a person's data into interpretive patterns, served as context to a language model. We evaluate the Specification on a prototype benchmark of held-out behavioral predictions scored by a calibrated 5-judge LLM panel. We test it independently and in composition with a range of context conditions: full raw corpus, full extracted facts, and four commercial memory systems (Mem0, Letta, Supermemory, Zep). Across 14 public-domain autobiographical corpora, the Specification lifts representational accuracy in aggregate and nearly eliminates model hedging. It recovers most of what the raw corpus delivers, at ~25x less context cost. The Specification lifts subjects toward a common predictive level regardless of pretraining baseline; the lift in absolute points is therefore largest where the baseline is lowest, suggesting the population of relevance is anyone not adequately represented in pretraining. Lift is greatest on interpretation-required questions, where providing an interpretive layer enables model behavior that extracted facts or raw corpus do not. Conversely, on recall-required questions, this layer can interfere rather than help. We conclude that representational accuracy is distinct from recall and that human-AI alignment is dependent on how accurately the user is represented. Representational accuracy makes that alignment testable.

  • 1 authors
·
May 26 2

KWBench: Measuring Unprompted Problem Recognition in Knowledge Work

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.

clio-ai Clio AI
·
Apr 16 2

HelpSteer2-Preference: Complementing Ratings with Preferences

Reward models are critical for aligning models to follow instructions, and are typically trained following one of two popular paradigms: Bradley-Terry style or Regression style. However, there is a lack of evidence that either approach is better than the other, when adequately matched for data. This is primarily because these approaches require data collected in different (but incompatible) formats, meaning that adequately matched data is not available in existing public datasets. To tackle this problem, we release preference annotations (designed for Bradley-Terry training) to complement existing ratings (designed for Regression style training) in the HelpSteer2 dataset. To improve data interpretability, preference annotations are accompanied with human-written justifications. Using this data, we conduct the first head-to-head comparison of Bradley-Terry and Regression models when adequately matched for data. Based on insights derived from such a comparison, we propose a novel approach to combine Bradley-Terry and Regression reward modeling. A Llama-3.1-70B-Instruct model tuned with this approach scores 94.1 on RewardBench, emerging top of more than 140 reward models as of 1 Oct 2024. We also demonstrate the effectiveness of this reward model at aligning models to follow instructions in RLHF. We open-source this dataset (CC-BY-4.0 license) at https://huggingface.co/datasets/nvidia/HelpSteer2 and openly release the trained Reward Model at https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Reward

  • 8 authors
·
Oct 2, 2024 5

RewardBench 2: Advancing Reward Model Evaluation

Reward models are used throughout the post-training of language models to capture nuanced signals from preference data and provide a training target for optimization across instruction following, reasoning, safety, and more domains. The community has begun establishing best practices for evaluating reward models, from the development of benchmarks that test capabilities in specific skill areas to others that test agreement with human preferences. At the same time, progress in evaluation has not been mirrored by the effectiveness of reward models in downstream tasks -- simpler direct alignment algorithms are reported to work better in many cases. This paper introduces RewardBench 2, a new multi-skill reward modeling benchmark designed to bring new, challenging data for accuracy-based reward model evaluation -- models score about 20 points on average lower on RewardBench 2 compared to the first RewardBench -- while being highly correlated with downstream performance. Compared to most other benchmarks, RewardBench 2 sources new human prompts instead of existing prompts from downstream evaluations, facilitating more rigorous evaluation practices. In this paper, we describe our benchmark construction process and report how existing models perform on it, while quantifying how performance on the benchmark correlates with downstream use of the models in both inference-time scaling algorithms, like best-of-N sampling, and RLHF training algorithms like proximal policy optimization.

  • 7 authors
·
Jun 2, 2025

From Manual Observation to Automated Monitoring: Space Allowance Effects on Play Behaviour in Group-Housed Dairy Calves

Play behaviour serves as a positive welfare indicator in dairy calves, yet the influence of space allowance under commercial conditions remains poorly characterized, particularly at intermediate-to-high allowances (6-20 m2 per calf). This study investigated the relationship between space allowance and play behaviour in 60 group-housed dairy calves across 14 commercial farms in the Netherlands (space range: 2.66-17.98 m2 per calf), and developed an automated computer vision pipeline for scalable monitoring. Video observations were analyzed using a detailed ethogram, with play expressed as percentage of observation period (%OP). Statistical analysis employed linear mixed models with farm as a random effect. A computer vision pipeline was trained on manual annotations from 108 hours on 6 farms and validated on held-out test data. The computer vision classifier achieved 97.6% accuracy with 99.4% recall for active play detection. Calves spent on average 1.0% of OP playing reflecting around 10 minutes per 17-hour period. The space-play relationship was non-linear, with highest play levels at 8-10 m2 per calf (1.6% OP) and lowest at 6-8 m2 and 12-14 m2 (<0.6% OP). Space remained significant after controlling for age, health, and group size. In summary, these findings suggest that 8-10 m2 per calf represents a practical target balancing welfare benefits with economic feasibility, and demonstrate that automated monitoring can scale small annotation projects to continuous welfare assessment systems.

  • 4 authors
·
Jan 25

PBSD: Privileged Bayesian Self-Distillation for Long-Horizon Credit Assignment

Long-horizon agentic tasks pose a fundamental credit assignment challenge for outcome-base reinforcement learning: trajectory-level rewards verify final correctness but provide limited guidance on which intermediate reasoning steps or tool interactions contribute to the outcome. The difficulty is especially pronounced in multi-turn search agents, where successful trajectories may contain misleading actions and failed trajectories may contain valuable evidence-gathering steps. We propose PBSD (Privileged Bayesian Self-Distillation), a Bayes-calibrated self-distillation method for fine-grained credit assignment under sparse final rewards. PBSD measures trajectory quality through the posterior-to-prior probability ratio of the verified answer and applies Bayes' rule to convert this hard-to-estimate answer-side ratio into a tractable likelihood ratio between a standard student model and a privileged answer-conditioned teacher model. Autoregressive decomposition of this Bayesian evidence score yields turn-level signals that identify whether each intermediate turn supports or undermines the verified outcome. Consequently, PBSD provides a principled and elegant reweighting scheme that transforms sparse outcome supervision into Bayes-calibrated turn-level credit signals, while remaining fully compatible with standard policy optimization. Experiments demonstrate that PBSD consistently enhances performance across both in-domain and out-of-domain settings, and effectively transfers knowledge from short-context training to long-context inference, suggesting that its fine-grained credit assignment mechanism facilitates more effective policy learning and yields improved generalization.

AgentFairBench: Do LLM Agents Discriminate When They Act?

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.

  • 4 authors
·
Jun 14

Probing Preference Representations: A Multi-Dimensional Evaluation and Analysis Method for Reward Models

Previous methods evaluate reward models by testing them on a fixed pairwise ranking test set, but they typically do not provide performance information on each preference dimension. In this work, we address the evaluation challenge of reward models by probing preference representations. To confirm the effectiveness of this evaluation method, we construct a Multi-dimensional Reward Model Benchmark (MRMBench), a collection of six probing tasks for different preference dimensions. We design it to favor and encourage reward models that better capture preferences across different dimensions. Furthermore, we introduce an analysis method, inference-time probing, which identifies the dimensions used during the reward prediction and enhances its interpretability. Through extensive experiments, we find that MRMBench strongly correlates with the alignment performance of large language models (LLMs), making it a reliable reference for developing advanced reward models. Our analysis of MRMBench evaluation results reveals that reward models often struggle to capture preferences across multiple dimensions, highlighting the potential of multi-objective optimization in reward modeling. Additionally, our findings show that the proposed inference-time probing method offers a reliable metric for assessing the confidence of reward predictions, which ultimately improves the alignment of LLMs.

  • 13 authors
·
Nov 16, 2025