Evaluation & benchmarks

New benchmarks, evaluation methodology, measuring what models can do

456 articles · page 6 of 10

Evaluation & benchmarks

Protocol risks in multi-agent systems depend on downstream agent decisions, not just local exploitability

The authors introduce Agent-Canary-Mas, an end-to-end system for measuring protocol risk in long-horizon multi-agent tasks. Using 975 matched malicious-benign task pairs across five business domains and three agent-to-agent protocols, they separate local attack success from final task harm. Local success often exceeds 90%, while consequence conversion ranges from 7.1% to 100.0%, and cross-agent evidence lifts attribution F1 from 29.7% to 95.5%.

·2 min
Evaluation & benchmarks

Skill-based benchmark reveals stark agent reliability gaps hidden by lenient scoring

The authors introduce SkillForgeBench, a benchmark that uses human-authored agent skills as procedural blueprints to automatically generate 842 sandboxed tasks spanning five professional domains. Evaluating eight mainstream agents reveals that strict, all-constraint satisfaction yields rates as low as 12.6% for weaker models and only 40.9% for the best (Opus 4.8), contrasting sharply with loose scoring rates of 70–85.7%. The work demonstrates that process rubric adherence is far rarer than completion rubric success, and that explicit skill guidance disproportionately benefits frontier models.

·3 min
Evaluation & benchmarks

MLLMs struggle to perceive human physical states quantitatively from video

The authors introduce HumanPhys, a benchmark and training corpus for evaluating multimodal large language models (MLLMs) on physically grounded human perception. The benchmark spans 13 sub-tasks covering spatial state, kinematics, blink events, deformation, illumination, and camera pose, with ground truth derived programmatically from sensor calibration or renderer state rather than human annotation. They find that all tested MLLMs perform poorly, with overall scores ranging from 19.74 to 46.94, and that training on the corpus improves performance on held-out benchmarks such as SPAR-Bench.

·3 min
Evaluation & benchmarks

New benchmark tests whether LLM agents can actively investigate causal systems

Chen et al. introduce ACED-Bench, a neuro-symbolic benchmark comprising 180 tasks in which LLM agents must actively collect evidence from a hidden probabilistic graphical model and return causal answers. The best contemporary models approach ceiling on a qualification check but leave substantial room for improvement on the core benchmark, with evidence-ledger analysis isolating failures in experiment design, evidence retention, and answer commitment.

·2 min
Evaluation & benchmarks

Benchmark evaluates AI agents' sustained progress on long-horizon tasks with process rewards

Li et al. present LHTB 2.0, an extended benchmark for evaluating AI agents on long-horizon terminal-based workflows. The authors add 31 tasks (30 via a systematic reconstruction pipeline) to the original 46, correct defects, and record checkpoints every 30 minutes over three-hour trials. Across 1,145 scored runs from 15 frontier models, they find that dense progress ladders capture substantial partial progress that binary pass/fail verifiers miss: on reconstruction tasks, binary scoring averaged 0.221 while dense ladder scoring averaged 0.653.

·3 min
Evaluation & benchmarks

Mean benchmark accuracy misleads model release decisions; recipe cards certify tail risk

The authors compare standard mean-accuracy reporting against the actual tail failure rate of fine-tuned model reruns. Across 2,124 models from 36 (model, training-budget) cells with 59 independent reruns each, mean-based release rules violate a 10% deployment failure budget 48.0% of the time. Their proposed recipe cards, which use exact finite-sample binomial bounds, reduce held-out budget violations to 0.6%.

·3 min
Evaluation & benchmarks

LLM authorization gates frequently block benign actions when safety alignment overrides explicit user permission

The authors present AuthGateBench, the first policy-guided benchmark for evaluating LLMs as authorization gates in coding agents. By constructing contrast pairs with identical actions but different requests, they show that models block most risky actions even when explicitly authorized, revealing unstable decision making. They also show that retrieving environment signals after the fact substantially improves accuracy.

·2 min
Evaluation & benchmarks

Grounding rewards must check source independence and conflict handling in search agents

Zhao et al. argue that evaluating search agents by final answer quality or citation support misses how answers use evidence across sources. They propose Cross-Source Evidence Grounding (CSEG), a six-layer decision tree over the full search trajectory, and show that a 4B reward model trained on it raises an agent's independently judged grounding score from 2.83 to 4.19 on a 1 to 7 scale.

·2 min
Evaluation & benchmarks

New benchmark tests LLM agents on exploiting cross-protocol vulnerabilities

The authors introduce ProtoArena, a benchmark of 80 controlled environments designed to evaluate LLM agents' ability to exploit vulnerabilities arising from interactions among network protocol components. Agents interact with running systems through black-box interfaces and must achieve verifiable security effects. The best-performing model, Claude Opus 4.8, succeeded in 48.8% of tasks with high guidance but only 12.5% under minimal black-box information, showing that reliable exploitation from external observations remains challenging.

·3 min
Evaluation & benchmarks

LLM agents fail to forget user information on demand, benchmark shows

Xie et al. introduce Bemula, the first benchmark dedicated to memory unlearning in LLM agents. The benchmark pairs each dialogue with a control version lacking forget requests to isolate genuine forgetting from failure to encode. Evaluating six mainstream memory systems under three backbone LLMs, the authors find that almost no forget request is successfully handled end-to-end, with most systems failing completely.

·3 min
Evaluation & benchmarks

CopRoute reduces inference calls by exploiting complementary failures among language models

The authors propose CopRoute, a dependence-aware conformal routing family that builds model sets sequentially by ordering models on how many new queries each answers beyond those already selected, then calibrates a cutoff to meet a target all-miss risk. On RouterBench and RouterEval, CopRoute achieves the fewest certified calls on nine of ten datasets at target risk 0.05 and eight of ten at 0.10, with up to 79 percent fewer calls (median 24 percent) compared to RACER.

·3 min
Evaluation & benchmarks

LLM creativity is best measured by performance on rare, long-tail outputs

The authors propose that what creativity benchmarks share is the evaluation of task performance in the long tail of a reference input-output distribution, where inputs or outputs are improbable. Analyzing 15 benchmarks and 26 open-weight models, they show creativity benchmarks sit further into the tail than standard benchmarks, and rankings derived from tail subsets align more closely with human creativity judgments in LM-Arena. They introduce q@Tail(α), a lower-cost diagnostic that measures mean quality on the most novel fraction of a benchmark.

·3 min
Evaluation & benchmarks

LLM agents discover hidden rules in social dilemmas but fail to sustain cooperation

The authors introduce OpenSocialJax, a benchmark extending Clean Up and Commons Harvest with procedurally generated layouts and hidden rules from tens of millions of configurations. Evaluating eight large language models in homogeneous and mixed teams, they find that while models often discover hidden rules and improve after initial trials, rule discovery does not reliably lead to sustained cooperation: cleaning becomes concentrated among fewer agents and shared resources are depleted.

·3 min
Evaluation & benchmarks

Coding agents rarely build the agent a user actually asked for.

The authors introduce AgentSWE, a benchmark for agent software engineering in which coding agents are commissioned to build, edit, or optimize a complete agent from a natural-language requirement. Across 25 tasks, they evaluate five models in the Codex harness plus other pairings, grading frozen submissions against requirement-derived rubrics on held-out cases. The strongest builders average 19.1 on creation and 31.6 on editing, and none closes more than 5 percent of the remaining gap in optimization.

·2 min
Evaluation & benchmarks

New benchmark tests AI judges on long computer-use tasks by swapping instructions

The authors introduce AgentHorizon, a benchmark of 1,373 instruction-trajectory pairs built from 166 hours of human-recorded computer-use sessions across three operating systems. By swapping instructions between closely related tasks, they create negative examples that test whether an AI judge can spot subtle failures in trajectories that look complete. Evaluating eleven judge models, they find that the best, GPT-5.5, achieves 80.9% balanced accuracy on the hardest split, and that providing tools helps some models but hurts others.

·3 min
Evaluation & benchmarks

LLMs match human annotators on some MT quality tasks, but both remain unreliable

The authors evaluated four LLMs as annotators for two machine translation quality schemes, MQM and ESA, comparing their agreement with human annotators across a 70-language-pair test set and WMT23/WMT25 data. LLM-human agreement matched or exceeded human-human agreement for some tasks, but both varied widely across annotation schemes, language pairs and domains, and remained unreliable for most settings.

·2 min
Evaluation & benchmarks

Benchmark and agent for verifying images against real-world facts

Chen et al. introduce WorldFact-Bench, a dataset of 1,274 real–fake image pairs spanning ten domains, each containing a factual conflict while preserving visual plausibility. They also propose PERSIST-Agent, an iterative verification system that maintains a persistent state linking candidate facts, observations, evidence, and verification statuses. On 8B backbone models, the agent raises pair accuracy by up to 10.4 percentage points over direct judgment, but overall performance remains modest, highlighting the gap between visual plausibility and factual correctness.

·3 min
Evaluation & benchmarks

Benchmark of 82 unresolved problems in theoretical sciences tests AI beyond known results

The authors introduce OpenProblemBench, a benchmark of 82 open problems drawn from the mathematics and theoretical physics literature. Each problem supplies research context, assumptions, and prior progress. Four evaluator models independently assess submissions for correctness and progress without reference solutions. Across seven evaluated configurations, GPT-6-Astra attains a mean judged solve rate of 14.0%, compared to 2.4–6.7% for other models.

·3 min
Evaluation & benchmarks

Clinical AI benchmarks poorly reflect how clinicians actually use language models

Vishwanath and colleagues compared eight months of real-world use of an institutional language model assistant with 58 public clinical AI benchmarks. They found a substantial mismatch: clinicians mostly requested help with documentation, administration and knowledge retrieval, while benchmarks emphasized diagnosis and assumed questions could be answered from the information provided.

·3 min
Evaluation & benchmarks

LLM-assisted peer review should prioritize error detection over imitation of human reviews

The authors introduce a verification-centric perspective for LLM-assisted peer review, focusing on error detection rather than mimicking human reviews. They present a scalable benchmark built by synthetically inserting logical contradictions into conference papers, and propose a Multi-Layered Review (MLR) framework that conducts multiple passes over the manuscript before generating feedback. Across evaluations, MLR shows strong alignment with human review scores, achieves high error detection performance, and offers complementary reviewer focus, though it remains vulnerable to adversarial manipulation.

·2 min
Evaluation & benchmarks

HELM Safety's HarmBench fails to measure a single harmful refusal construct.

The authors audited HELM Safety's HarmBench dataset for construct validity of 'harmful refusal'. They found that HarmBench does not measure a single attribute: multidimensional item response theory required several latent dimensions, and differential item functioning flagged items where developers affect scores beyond refusal ability. This undermines the assumption that safety benchmark scores reflect unidimensional constructs.

·3 min
Evaluation & benchmarks

Input-blind controls mimic oracle gains for layer programs in multiple-choice evaluation

Using 32 layer-skipping and repetition programs on two language models (Qwen3-4B-Base and Llama-3.1-8B) across 4,413 multiple-choice items, the authors compare oracle gains with gains from input-blind controls at the same sites. They find that with shared option order, these controls produced 10.2-11.8 and 15.6-19.4 percentage points of headroom—exceeding the real programs' 9.0 and 10.1 in all random-direction draws per model. This indicates that substantial headroom can persist without establishing a benefit specific to the selected layer computation.

·1 min
Evaluation & benchmarks

Coding-agent benchmarks should mirror the task flows of their target users

The authors collected 4,782 agent sessions of software engineers in JetBrains IDEs and analyzed the long sessions with at least three user messages, which account for 33% of the sample. They find that real task flows mix and shift among types such as explanation, planning, review, and refactor, unlike issue-derived benchmarks, and propose SWE-TaskFlow to calibrate benchmark interaction flows. A pilot on 700 SWE-Bench Pro tasks shows that changing the interaction protocol alone can roughly double agent cost without stable gains in resolve rate.

·3 min
Evaluation & benchmarks

Benchmark shows image editing models fail to parse scribble-only intent

The authors introduce ScribbleEdit, a benchmark for image editing in which the user's only input is a scribble drawn on the image. Across 23,772 synthetic training pairs and 250 manually labeled evaluation pairs, existing vision-language editing models fail to correctly identify the intention of the scribble. The authors show that understanding the scribble, rather than generation quality, is the main bottleneck.

·2 min
Evaluation & benchmarks

Text classifier predictions shift with serving context, even with fixed model and input

Kasa et al. systematically measure how serving-context factors—batch size, batch composition, padded length, precision, quantization, device, and inference backend—affect the outputs of transformer text classifiers. They train 180 models across discriminative, pseudo-generative, and fully generative formulations, and evaluate each under 3,960 serving contexts, holding the checkpoint and text fixed. The authors find that label stability does not imply score stability: changing padded length alone flips about 10% of labels, while bf16 precision moves up to 56.7 percentage points of probability mass without altering labels in some settings.

·3 min