Evaluation & benchmarks

New benchmarks, evaluation methodology, measuring what models can do

456 articles · page 1 of 10

Evaluation & benchmarks

Style evaluator for AI art correlates with human judges by first seeing reference examples

The authors introduce StyleJudge, an 0.8-billion-parameter vision-language evaluator that learns to judge artistic style transfer quality by first observing a few reference source-to-stylized pairs, mirroring how a human calibrates before rating. On a human-annotated benchmark of 878 pairs across 12 styles, StyleJudge achieves a Pearson correlation of 0.291 with human ratings on seen styles and 0.381 on two unseen styles, outperforming Gram-matrix loss (r=0.041) and CLIP similarity (r=0.168), while LPIPS correlated negatively (r=-0.404).

Evaluation & benchmarks

Language models can write small networks from scratch but cannot safely edit large ones

Marasini et al. introduce WeightBench, a benchmark for systematically evaluating how well large language models can read, fix, and write neural network weights without gradient descent. Their results show that frontier agents can create simple control policies from scratch, but cannot yet patch large language models, compress them without data, or tell impossible edits from hard ones.

Evaluation & benchmarks

Video assistants need to track event progress to follow conditional instructions

The authors introduce CoTriggerBench, a benchmark for evaluating streaming video assistants on instructions that require tracking events over time. They propose TEMPI, an agent that decomposes instructions into atomic queries and maintains a scenario state. TEMPI achieves 44.4% instruction-level success, outperforming existing streaming VLMs by over 15 percentage points, and runs efficiently on edge hardware with a 4.9x speedup.

Evaluation & benchmarks

Language models learn categories poorly, especially flexible context use

The authors introduce LeCat, a framework to measure how language models learn category knowledge (e.g., that airplanes are fast) and how pretraining data shapes that learning. Across open-source models and controlled training, they find that flexible use of category knowledge in varied contexts remains persistently difficult, ver thing rule-like statements reduce flexible learning, and diverse expressions improve it.

Evaluation & benchmarks

Web agents handle audio and video, but best systems still fail on cross-site tasks

The authors introduce WebArena-Pro, a benchmark of 300 human-validated web agent tasks across 20 self-hosted applications in six domain categories. It extends evaluation to include audio and video as core observations, and runs each task in container isolation for reproducible parallel evaluation. The best performing agents achieve 62.3% success, revealing that cross-site tasks remain significantly harder than single-site ones.

Evaluation & benchmarks

Game-generating AI agents meet only a third of requirements, benchmark finds

The authors introduce OCTAVE, a benchmark for end-to-end HTML5 game generation comprising 180 tasks with 1,800 evaluation requirements across eight capability dimensions. Evaluating 11 model-harness combinations, they find average pass rates of 33.2% for completeness and 32.5% for richness, with 19.9% of attempts yielding no playable artifact. Even the best combination remains below 50% on both public and hidden requirements.

Evaluation & benchmarks

Language models rely on memory even when citing external sources

The authors propose counterfactual value substitution to measure provenance in retrieval-augmented generation. They find that across multiple models, 1.2% to 15.7% of citations misstate the source of the claim, with rates increasing with model size. Standard natural language inference verification performs no better than chance, and even ten attribution methods cannot improve detection enough for deployment.

Evaluation & benchmarks

Language models mistake attention-grabbing clues for reasons to change beliefs

Chan and colleagues evaluate whether language models can judge when new evidence justifies changing someone else's belief, rather than simply tracking what that person knows. Across 2,415 human-validated scenarios, the authors report a recurring mismatch: models often detect potential belief changes but struggle to infer the resulting belief, frequently updating on insufficient evidence.

Evaluation & benchmarks

Simulated business histories yield verifiable benchmarks that separate AI agent performance

The authors present SynthCorp, a platform that simulates how business events create and modify corporate data over time, producing relational databases, document stores, and data catalogs alongside evaluation tasks with verified gold programs. In tests across five domains, the benchmarks consistently separate different data contexts and model tiers, and an auditing method using the simulation log automatically verifies gold programs without manual annotation, even on composite tasks that current state-of-the-art agents fail.

Evaluation & benchmarks

AI image evaluators better match disability communities’ preferences after tailored training

Pradier and colleagues report that automated image evaluation can better reflect how particular disability communities want to be represented when models learn from community-defined standards. Their evaluators outperformed conventional measures of image quality or image–text alignment in matching community ratings, and the authors demonstrated their use in content filtering.

Evaluation & benchmarks

Rising standards, not declining AI, drive worsening human evaluations

The authors analysed 1,260,967 blind human evaluations on Chatbot Arena from May 2023 to August 2024, focusing on pairs of language models whose versions were fixed. They found that the proportion of votes saying 'both are bad' rose steadily over time, even for the same pair of models, indicating that human evaluators' standards were rising. Meanwhile, the relative ranking between models remained stable, showing that only absolute judgments drifted.

Evaluation & benchmarks

Medical AI segmentation models exploit size hints leaked from benchmark captions

Nataraj et al. show that two published vision-language medical image segmentation models rely mostly on a single piece of information from their input captions: the lesion's size or location extent. On colonoscopy and chest X-ray benchmarks, removing the extent word or phrase from the caption loses 52-77% of the segmentation improvement over an empty prompt, while other words (color, shape, count) contribute little. The authors trace this dependence to the training captions, which in some datasets are algorithmically derived from the ground truth segmentation mask, creating a form of label leakage they term 'extent leakage'.

Evaluation & benchmarks

AI judges react to irrelevant changes but miss real failures in embodied tasks

The authors introduce state-conditional invariance (SCI), a principle requiring VLM judges to remain stable under semantics-preserving presentation changes but responsive to causal outcome changes. They build GridWM-Judge, a benchmark from grid-world trajectories with aligned variants and probes, and test several VLM judges. The evaluation shows heterogeneous failure modes: sensitivity to presentation and representation, low activation on successful trajectories, and weak causal discrimination, indicating that single-score evaluations miss key reliability issues.

Evaluation & benchmarks

Pass/fail grading hides tenfold differences among top AI agents

The authors show that the single pass/fail bit used in most agent benchmarks is too coarse to separate top-tier agents and is capped at the reference solution. On SWE-bench Verified, top ten agents differ by less than the test errors, and on Terminal-Bench, pass/fail-tied models vary over tenfold on a graded metric. They propose a clipped log-ratio score with dated anchors and cost as a separate axis.

Evaluation & benchmarks

Enterprise agents act without permission more with missing evidence than explicit denial

The authors present DeputyBench, an executable benchmark for evaluating whether enterprise agents respect permission contexts defined by organizational records and policies. Testing 12 models on 375 instances, they find that agents produce prohibited actions in 10.3% of runs when permission evidence is masked, versus 3.5% when evidence explicitly denies permission, a 6.8 percentage point difference. Post-hoc analysis attributes 61 of 100 such failures to four privacy tasks requiring complete cumulative counts or person-specific consent.

Evaluation & benchmarks

AI agents often repeat failed strategies instead of using corrective feedback

Hernandez and Wilson studied how pairs of AI agents cooperate when neither has all the information needed to solve a puzzle. Across 400 GPT-4o conversations, failure to incorporate external feedback was the strongest indicator of unsuccessful collaboration. A follow-up experiment found that instructions addressing the observed failure patterns improved average success by 10 percentage points.

Evaluation & benchmarks

Secure enclaves let AI evaluators test proprietary models without revealing prompts

The authors present Double-Blind Evaluation (DBE), a framework using hardware secure enclaves to cryptographically protect both an evaluator's test prompts and a model owner's proprietary weights. In a first live deployment, they used NVIDIA H100 enclaves on Google Cloud to evaluate Google DeepMind's Gemini 2.5 Flash Lite against the private MLCommons AILuminate benchmark, demonstrating mutual secrecy with less than 5% compute overhead.

Evaluation & benchmarks

AI agents produce correct answers even when visible history is insufficient

The authors introduce a protocol that separates reference agreement from evidence sufficiency in agent memory evaluation, using executable device logs. They show that Qwen3-32B and DeepSeek-V4-Flash often give concrete answers matching hidden references from histories that do not provide sufficient evidence, and seldom abstain when the visible history is insufficient. This suggests that typical answer-matching evaluations may overestimate how much agents rely on their provided context.

Evaluation & benchmarks

Generalist tabular models match specialists on utility, not fidelity

Researchers introduce OTBench, a benchmark of 56 datasets, 46 methods, and 20 capability axes for tabular data tasks. They find that a single model trained on generation, imputation, and prediction can match or surpass specialist models on utility metrics like prediction accuracy and imputation error, but specialists still win on distributional fidelity. Adding a task's own training objective yields the biggest gains, and combining objectives often gives less than additive benefits.

Evaluation & benchmarks

AI agents turn binary patches into exploits, but only 20% succeed

The authors developed CyberBinGym, a benchmark of 100 tasks drawn from real CVEs in 53 projects. Each task supplies only a stripped vulnerable and patched binary, requiring agents to combine static and dynamic analysis to both recover the vulnerability and weaponize it. The best agent triggers a crash in 88% of tasks but delivers a validated exploit in just 20%, demonstrating that autonomous binary-only exploitation is feasible but still unreliable.

Evaluation & benchmarks

AI agents keep record groups intact but often place them under the wrong parent

The authors introduce InfoOrgBench, a benchmark of 150 tasks that test how AI agents organize scattered records into folder hierarchies. They find that errors concentrate at the parent level: agents correctly group records but then file the group under the wrong overarching category. Model choice dominates performance far more than the choice of agent harness, and the top system reached 83.2 out of 100.

Evaluation & benchmarks

AI reward models struggle to judge research reports; language model judges outperform

The authors introduce DeepResearch RewardBench, a benchmark to evaluate reward models on the task of assessing AI-generated research reports. They find that current reward models (sequence classifiers, DPO-based) perform poorly on long-form, information-dense reports, while generative LLMs used as judges score better, especially in comparative evaluation settings. Even the best evaluators still fail on reports requiring specialized knowledge or up-to-date information.

Evaluation & benchmarks

Self-improving AI agents outperform fixed ones on open-ended tasks

The authors introduce Flex-RSI, a benchmark to test whether frontier AI agents can improve both their task solution and their own operational harness, prompts, tools, memory, and control loop, under a fixed budget. Evaluating 18 models on 24 vision and robotics tasks, they find that self-improvement yields the largest gains when the agent builds its own solution, and that open-weight models falter when feedback is binary (success/failure) or when outputs must pass strict physical or exact correctness checks.

Evaluation & benchmarks

AI design agents pass structural checks but violate hidden requirements

Rong Pan and Jianqiang Wang introduce DRAFT-Bench, a lifecycle-oriented benchmark for evaluating AI agents on professional design assignments. They test Astra and Claude Code across 43 execution conditions and find that while both agents achieve high structural completion, semantic evaluations and hidden production obligations reveal frequent violations, with Astra flagged for 11 semantic issues and Claude for 5 out of 28 shared core conditions.