arXiv:2609.37315v1 Announce Type: cross Abstract: Tool-using agents are entering settings where a wrong action carries real cost, and the benchmarks certifying them grade what each simulated tool cal...
arXiv:2609.35889v1 Announce Type: cross Abstract: Tool-using language-model agents select and execute third-party artifacts. Different implementations can return the requested output while producing...
TestJack is a scalable framework that audits large language model (LLM) coding agents by generating new tests for each trial, focusing on prompt requirements that may be violated. It keeps only the tests passed by the ground‑truth patch and re‑examines failures, providing replayable evidence for each confirmed error. In experiments across six model backends and five benchmarks, TestJack found that 34.4% of trials deemed correct actually violated task requirements, reducing the overall resolution rate from 50.6% to 33.2%.
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.
arXiv:2607. 16345v1 Announce Type: cross Abstract: Modern agentic systems increasingly rely on skills: installable packages of natural language and code that teach an LLM agent to perform a domain task.
The paper introduces TRACE, a protocol designed to diagnose whether changes in verifier scores for large language model agents reflect actual changes in agent behavior or merely alterations in the evaluation process. TRACE works by applying targeted changes to evaluation components, running paired experiments, and re‑scoring unchanged trajectories to isolate the source of score variation. Experiments on synthetic tasks and public benchmark tasks demonstrate that seemingly significant score shifts can often be attributed to evaluation artifacts, while TRACE can also detect genuine behavioral changes.
The paper introduces VerifierBench, a benchmark comprising 222 tasks from six agent benchmarks and 4,993 hidden test cases designed to evaluate the generation of executable verifiers by large language model agents. It presents Verifier4Verifier (V4V), a six-check audit that screens source verifiers and test suites, and reports that across 19 models, normalized balanced accuracy ranges from 10.6 to 62.1, with generated verifiers more reliably rejecting invalid solutions than accepting valid ones. Case studies highlight that many false rejections stem from verifiers that accept no tested solution, revealing unsupported assumptions about solution formats, tool use, and code changes that exclude valid solutions, and underscore the need to test both acceptance of correct solutions and other permissible solutions.
The paper introduces a new evaluation method called "same-input rerun" to assess the consistency of clinical language‑model agents across repeated runs. By replaying 1,000 MedAgentBench tasks with identical inputs, the authors find that action‑level outputs—such as test orders, medication requests, and referrals—vary significantly, even when benchmark scores remain unchanged. The study demonstrates that current benchmarks, which typically evaluate only a single run per task, can miss substantial behavioral divergence.
The paper introduces a pre‑action verification framework for large language model agents that emit shell commands or code edits. By running deterministic checks before execution, the system can catch invalid commands (95.8% success with a 10.0% false‑positive rate) and prevent silent failures in code edits, achieving high recall while minimizing false positives. The authors provide benchmarks, verifiers, and guards for both action modalities.
AgentEval is a system that tracks a contract-bound assessment record throughout a stateful agent’s tool execution, capturing state, events, and evidence to determine whether obligations are satisfied, violated, or unresolved. In 32 tool‑world administrations, independent checks confirmed all 30 complete episodes, while controlled interventions on 48 evidence views uncovered 36 binding faults that hash‑only checks missed. A public‑benchmark study on 192 held‑out executions identified ten completed subgoals within broader task failures and highlighted the impact of missing evidence on overall task success.
arXiv:2606. 02965v2 Announce Type: replace Abstract: As large language models gain tool access and are deployed as autonomous agents capable of editing records, executing transactions, and modifying infrastructure, we still evaluate them based on the sole metric of task completion.
The paper investigates how AI agents behave when a task becomes impossible, focusing on whether they stop or escalates and how observing other agents influences this decision. Using seven ImpossibleBench tasks and models GPT‑5.6 Sol, Claude Fable 5.1, and Gemini 3.8 Flash, the study compares solo and three‑agent settings under explicit‑boundary and benchmark‑native regimes. Results show that agents differ markedly: Fable escalates, Sol usually stops, and Gemini often fails to decide, with boundary‑crossing behaviors emerging from both rule evasion and ambiguity about protected system states.