Evaluation & benchmarks

New benchmarks, evaluation methodology, measuring what models can do

456 articles · page 7 of 10

Evaluation & benchmarks

Pass@k misses output diversity changes caused by model post-training

Rath and colleagues compared two post-training methods applied to the same Qwen2.5-1.5B-Instruct checkpoint on grade-school mathematics. Across three seeds, the methods moved token entropy, answer entropy and unique-answer counts in opposite directions, yet pass@8 and pass@32 did not consistently distinguish them. Comparisons with the original checkpoint also showed that GRPO’s pass@1 advantage over rejection-sampling fine-tuning reflected better capability retention, not a demonstrated capability gain.

·3 min
Evaluation & benchmarks

Multilingual benchmark exposes deepfake detectors’ reliance on synthetic audio cues

Segna, Tschesche and Rohrbach introduce a consent-based benchmark for testing audio-visual DeepFake detectors across languages, manipulation methods and demographic groups. Their experiments show substantial detection failures when manipulated video retains authentic audio, suggesting that some multimodal systems rely heavily on audio-video inconsistencies rather than robust visual evidence.

·2 min
Evaluation & benchmarks

OOD detection gains mostly reflect dataset identity, not novelty

Lee and Kang measured how much of the reported out-of-distribution detection gain survives when the detector is tested on an OOD dataset it has never seen—not just on held-out samples of the same dataset. They find that most of the gain comes from recognizing which dataset the input belongs to, not from detecting novelty per se, and they derive a closed-form formula that predicts the inflation without needing a separate hold-out set.

·2 min
Evaluation & benchmarks

Random rewards reveal three stages of LLM capability development

The authors explain the spurious-reward paradox—why random rewards improve LLM performance—by introducing reachability, the potential for further training to elicit dormant capabilities. Applying random-reward RL to OLMo checkpoints across training, they find three distinct regimes of responsiveness: early pre-training (dormant), later pre-training (receptive), and mid-training (autodidactic). The pattern reveals that capabilities not yet expressed in current accuracy can be activated by training, even without correctness feedback.

·2 min
Evaluation & benchmarks

Most video benchmarks fail to actually test video understanding

Song et al. introduce the attack pyramid, a five-level hierarchy of shortcut attacks that exploit different information sources (options, text, other items, single frames, or shuffled video). Applying it to 115 video benchmarks, they find that on 35 benchmarks attackers that never see a frame achieve full-video accuracy, and on 51 benchmarks with temporal probes, shuffled frames retain a median of 96% of the full-video accuracy. They screen 505,518 question-answer pairs to produce Video-Index, a meta-benchmark of the 840 hardest verified items, and show that even the best evaluated system (Claude Opus 5 with agent tools) leaves significant room for improvement.

·3 min
Evaluation & benchmarks

Query-language preference emerges strongly under conflicting multilingual evidence

The authors introduce Waldo, a multilingual QA benchmark of 12K instances across five languages, designed to test query-language preference under knowledge gaps and conflicts. Evaluating eight LLMs, they find that models largely ignore query language when a fact is absent from one edition, but under conflicting accounts, responses strongly align with the query-language source. Two mitigation approaches—attention head ablation and LoRA training—reduce the preference gap by up to 61.5%.

·2 min
Evaluation & benchmarks

AI agents can predict accurately but rarely produce scientific insights, benchmark shows.

The authors introduce EurekaBench, a benchmark of 26 long-horizon tasks across six scientific fields, and evaluate seven AI agents on three axes: constraint satisfaction, predictive accuracy, and scientific insights. They find that top agents nearly match human predictive accuracy but score less than half on insight generation, revealing a gap between optimizing for predictions and deepening scientific understanding.

·2 min
Evaluation & benchmarks

Evaluation scores for text-to-3D models shift more with configuration changes than model changes

Lam et al. audit how much the evaluation of text-to-3D scene generators depends on measurement protocol choices rather than scene quality. By fixing all generated scenes and varying render and caption settings in 19 alignment evaluators, they find that configuration variance exceeds between-generator variance for 17 evaluators. Rankings are more stable than scores, but most evaluators change their reported top generator under some configuration.

·2 min
Evaluation & benchmarks

LLM agents struggle to generate physics simulators from descriptions, new benchmark shows

The authors introduce WorldSolver, a benchmark of 168 simulation tasks derived from physical phenomena in 61 classic computer graphics papers across 7 domains. Agents are given a fixed code scaffold describing a simulation environment and must implement the solver. Evaluation uses three criteria: execution checks, visual fidelity of rendered output, and physical plausibility of the state trajectory. The best-performing agents, GPT-5.6-Sol and Claude-Opus-5, achieved overall scores of only 48.7% and 46.7%, respectively.

·3 min
Evaluation & benchmarks

Comment history outperforms explicit personas for simulating human discussion behavior

The authors introduce Wiki-Talkie, a multilingual dataset of real Wikipedia discussions with ground-truth user personas across five languages. Evaluating four LLMs on next-turn generation, they find that conditioning agents on prior comment history consistently outperforms using explicit persona descriptions. Agents systematically underproduce negative sentiments and overproduce references and suggestions, showing a bias toward agreeableness that holds robustly across languages.

·2 min
Evaluation & benchmarks

RAG systems can fail catastrophically when knowledge bases contain semantic errors

The authors introduce RAGStress, a controlled benchmark that stress-tests retrieval-augmented generation under systematic knowledge-base corruption. Evaluating five LLMs across 52,500 model-question-condition evaluations, they find that semantic corruptions like factual errors and contradictions degrade accuracy far more than numeric typos or relevance poisoning, and that clean retrieval can mask underlying fragility.

·2 min
Evaluation & benchmarks

Simple choices outperform complex recipes in iterative reward design benchmarks

Bhamidipaty and colleagues introduce BIRD, a framework that places iterative reward design methods under matched feedback conditions, implementations and policy-training budgets. After comparing six methods and ablating 30 design choices, the authors combine the most reliable components into two algorithms that significantly outperform the evaluated prior methods in their respective settings.

·3 min
Evaluation & benchmarks

AutoSciBench automatically generates scientific agent benchmarks that adapt as agents improve.

Kaplan et al. introduce AutoSciBench, a framework that automatically generates and iteratively refines benchmark tasks for scientific AI agents. By representing tasks as high-level concepts and low-level recipes, then using solver trajectories to close shortcuts and shift tasks toward deeper reasoning, AutoSciBench produced benchmarks that reduced solver accuracy by 22.4 and 25.5 percentage points relative to human-curated benchmarks in computational biology and materials science respectively, while receiving higher average quality ratings across all three tested domains.

·2 min
Evaluation & benchmarks

Robot policies encode changed instructions but still follow scene cues

Jiang and colleagues tested whether language-conditioned robot policies actually use instructions to choose among multiple feasible actions. Across vision-language-action policies and world-action models, robots usually pursued the target associated with the scene even when the instruction validly requested another visible object, exposing a gap between encoding language and letting it control behavior.

·2 min
Evaluation & benchmarks

Current date in system prompts skews LLM evaluation results

Sanz-Guerrero et al. identify a hidden source of non-determinism in LLM evaluation: the automatic injection of the current date into system prompts. Across 9 recent LLMs and 6 datasets covering multiple-choice QA, math reasoning, code generation, and machine translation, the authors find that performance fluctuates solely with the date, with deltas of up to 6% accuracy on MCQA, 14% on math, 7% pass@1 on code, and 2.84 BLEU on translation. This date effect exceeds variability from batch size or numerical precision, and standard techniques like chain-of-thought or few-shot prompting do not reduce it.

·3 min
Evaluation & benchmarks

Framework reveals that LLM judges share perception but differ in prioritization

The authors introduce JudgeProfile, a framework that decomposes LLM evaluation into perception (how a judge compares responses on specific qualities like clarity and correctness) and prioritization (how much each quality influences the final choice). They curate SubjectiveSet, a dataset of 50,013 response pairs from 17 sources evaluated by 21 LLMs across 87 attributes. The key finding is that judges often agree on attribute-level judgments even when their overall choices diverge, and reweighting attribute importance improves agreement with reference labels from 66.48% to 71.97% without updating the model.

·3 min
Evaluation & benchmarks

PDE models falter when sensor patterns change between training and testing

Xu and colleagues introduce PDE-OBS, a benchmark for testing how scientific machine-learning models respond to changes in measurement density and spatial layout. In 3,969 evaluations, every PDE-method pairing had higher average error under mismatched training and test patterns than under matched patterns, and simply providing denser test observations did not reliably improve a fixed model.

·3 min
Evaluation & benchmarks

Real-voice benchmark reveals model-specific demographic bias in Audio-LLMs

The authors introduce HEAR, a large-scale benchmark of 87,105 real human audio samples from 843 demographically diverse participants, designed to measure demographic bias in Audio-LLMs. Evaluating both speech-to-text and real-time speech-to-speech models, they find that voice-conditioned bias varies by model and that personalization instructions exacerbate demographic disparities, establishing voice bias as a controllable characteristic.

·2 min
Evaluation & benchmarks

Answer pooling revives saturated multiple-choice benchmarks by withdrawing elimination credit

The authors propose AnswerPool, a method that converts grouped multiple-choice benchmarks into harder tasks by pooling all answer options from questions sharing a context. Across eight benchmarks and eighteen models, every model scores lower under pooling, and the accuracy loss reveals how much the multiple-choice format helped via elimination. When correct answers are deleted from the pool, most open-weight models answered 87–100% of unanswerable questions, exposing a critical failure to abstain.

·2 min