arXiv:2608.22510v1 Announce Type: new Abstract: Agent benchmarks often evaluate only final answers even when agents run on stateful runtimes. We argue this under-specifies what is being evaluated: th...
Agent evaluations increasingly benchmark LLMs, but rankings can be swayed by evaluation conditions such as scaffolds or tasks, making reliability claim‑dependent. A Bayesian variance‑decomposition framework applied to 22 benchmarks shows that reliability varies with the measurement goal: fixed model‑scaffold systems rank reliably, while underlying‑model rankings are less stable. Scaffold choice can alter conclusions, and adding more tasks only modestly improves reliability when scaffold coverage is limited; however, pooling diverse benchmarks can substantially raise cross‑task ranking reliability and reduce cost.
BekchiAI introduces a benchmark and platform for evaluating large language model agents. The benchmark comprises 13 tool‑using ReAct agents across seven task categories, totaling 2,057 deterministic test tasks with verifier‑checkable gold answers. The platform offers web‑based observability, token and latency telemetry, and remote run termination for live agents.
arXiv:2608.21601v1 Announce Type: new Abstract: Benchmarks for scientific artificial intelligence are mostly written to be scored: multiple-choice questions, curated agent tasks with reference soluti...
The paper argues that agentic systems waste time and memory by guessing how long tool calls will take, rather than using explicit progress signals from the tools themselves. It demonstrates that tools can report their remaining work or imminent completion, and that incorporating this feedback into serving systems dramatically improves cache decisions and reduces token latency. The authors show that this approach outperforms existing predictors and works robustly across different environments.
The paper introduces ACES (Agentic Continuous Evaluation of Skills), a framework that evaluates reusable skills and capability packages by running paired live trials with and without a target skill, normalizing results into the Agent Trajectory Interchange Format (ATIF), and grading six runtime metrics to compute Skill Lift. ACES demonstrates that scan-only gates miss important aspects of skill performance, while the evaluation protocol reveals significant improvements in skill execution, behavior check, and skill efficiency across 145 real skills and 947 scored cases. The open‑source NVIDIA SkillEvaluator implementation enables reproducible, repository‑native assessment of agentic skills in production environments.
arXiv:2605. 22949v3 Announce Type: replace Abstract: Foundation-model pools are increasingly used as black-box responders in coordinated systems where a coordinator must decide which response to trust.
The paper introduces SPLIT, a method that optimally allocates test‑time compute among sampling, model selection, and verification for language models and agents. Using an exact probability formula that accounts for task difficulty, shared blind spots, and verifier correlations, SPLIT is fitted via empirical Bayes on a small labeled pilot and then used to plan budgets for unseen tasks. Across seven diverse benchmarks, 26 models, and six verifier types, SPLIT achieves near‑optimal performance, outperforming rule‑based baselines by up to 2.5 percentage points and providing insights into when additional LLM judges reduce precision.
GAUGE is a new offline protocol that evaluates whether the common practice of using an LLM-as-a-judge to rank task‑oriented agents actually aligns with a verifiable reward. Across 25 agents from six providers on two benchmarks, GAUGE finds that user satisfaction scores are largely uncorrelated with task success, and that the judge’s ranking loses precision when agents are closely matched in performance. The study highlights a gap between ranking validity and construct validity in current evaluation practices.
The paper investigates how language‑model judges can make version‑dependent errors when evaluating upgraded agents. Using 35 public coding‑agent submissions, two customer‑service agents, and over a thousand expert‑labeled trajectories, the authors show that fixed judges often reject task‑conditioned error invariance and can incorrectly approve failed patches, especially as agent capability increases. Paired audits of current outputs reduce interval width only marginally, and the study concludes that independent human patch review is still necessary.
The paper introduces a Trust Layer for Agent Evaluation, a post‑hoc framework that appends a belief flag to each benchmark score. It checks four properties: alignment with the benchmark’s grading logic, traceable computation of a passing answer, consistency between the agent’s claim and actual behavior, and stability across repeated runs. Applied to five agent configurations on 108 tasks, the study finds that only 22.6% of recorded passes satisfy all checks, revealing widespread issues such as untraceable computation, false completion claims, and unstable results.
arXiv:2607. 22585v1 Announce Type: new Abstract: Public leaderboards for coding agents typically rank systems by model name and pass rate, while the surrounding harness (the scaffold that issues tools, manages context, and decides when to stop) is often under-specified.