What Benchmarks Don't Measure: The Case for Evaluating Abstention Competence in Autonomous Agents
The paper discusses the limitations of current benchmarks for evaluating autonomous agents, particularly their failure to assess when agents should abstain from action. It introduces the concept of compliance bias, where agents are incentivized to act even without sufficient information or authorization. The authors propose a new taxonomy and evaluation protocols to better measure abstention competence in agents.
- ▪Current benchmarks for autonomous agents do not evaluate whether agents should proceed with actions.
- ▪The authors identify compliance bias as a tendency for agents to act without adequate inputs or authorization.
- ▪They propose a taxonomy of abstention scenarios and new evaluation protocols to improve safety and usability in autonomous agents.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2606.02965 (cs) [Submitted on 1 Jun 2026] Title:What Benchmarks Don't Measure: The Case for Evaluating Abstention Competence in Autonomous Agents Authors:Victor Ojewale, Suresh Venkatasubramanian View a PDF of the paper titled What Benchmarks Don't Measure: The Case for Evaluating Abstention Competence in Autonomous Agents, by Victor Ojewale and 1 other authors View PDF HTML (experimental) Abstract:Benchmarks for autonomous agents measure whether agents complete tasks, yet this framing is systematically blind to whether an agent should have proceeded at all.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
