arXiv:2607. 12338v1 Announce Type: new Abstract: Agent benchmarks often compare two agents after all tasks have run, but costly evaluations make partial runs tempting.
Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.
arXiv:2608. 05519v1 Announce Type: new Abstract: Agent benchmarks usually measure task completion and treat resource use as an auxiliary statistic.
The paper investigates how test‑time computation can enhance language models and at what cost, introducing the SELF‑POT benchmark to evaluate this across competition mathematics, competitive programming, and agentic workflows. SELF‑POT separates candidate coverage from final accuracy, tracks correctness transitions under revision, and measures protocol completion alongside task success. Using a unified budget rule, the study compares direct inference, parallel sampling, and self‑revision across five low‑cost reasoning models, revealing that selection rules and failure handling significantly influence gains and cost savings.
The paper introduces $ au^ au$-Bench, a benchmark that turns the construction of AI agents into a measurable task. In this environment a developer agent receives real business records, client requirements, a production API, an existing codebase, and constraints on cost and models, and must deliver a complete customer‑service agent. The benchmark evaluates performance by deploying the agent against simulated users, revealing that current state‑of‑the‑art models achieve only 23.9% success while an expert‑written reference scores 82.2%.
arXiv:2609.01603v1 Announce Type: cross Abstract: Evaluating software engineering agents on realistic benchmarks is costly, since each task may require multi-step code exploration, modification, and...
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 TEAM-Design, a rule that assigns two replay probabilities to each task—one for a human-only replay and one for an agent-only replay—based on how difficult it is to predict the missing baseline outcome and the cost of replay. It addresses the challenge of deciding whether to keep a human-AI workflow or replace it with a single actor when only one outcome can be observed after deployment. The authors prove that TEAM-Design solves the budgeted design problem and controls error rates, and demonstrate its effectiveness on clinical and coding benchmarks, noting it excels when one comparison is clearly harder than the other.