Title: Beyond Task Completion: The Imperative of Assessing Abstention Capabilities in Autonomous Agents

Abstract:

Current evaluation frameworks for autonomous agents primarily assess task completion, a metric that systematically overlooks a critical question: whether the agent ought to have acted at all. Agents optimized via human feedback exhibit a structural propensity to proceed despite insufficient inputs, lacking evidence, or possessing no authorization for safe execution. We label this tendency "compliance bias," noting that both the reward signals and benchmark scoring systems inherently favor action, treating continuation as the default correct behavior regardless of the absence of preconditions for safety.

This paper presents three primary contributions. First, we demonstrate that compliance bias stems from reward hacking within human-feedback pipelines and is reinforced by major agent benchmarks. These benchmarks either penalize agents for halting or lack the architectural capacity to differentiate between a principled pause and a silent failure. Second, we introduce a three-gap taxonomy to categorize scenarios warranting abstention: specification gaps (missing required information), verification gaps (unconfirmable world states), and authority gaps (lack of explicit authorization). This framework establishes a principled foundation for developing abstention-aware benchmarks. Finally, we propose evaluation protocols comprising three metrics: Safety Rate, Usability Rate, and Informed Refusal Rate. Preliminary results across 144 enterprise agent scenarios and five model families reveal that a runtime-enforced abstention mechanism can block up to 89.2% of hazardous actions while maintaining 87.5% usability in authorized contexts. These findings indicate that the safety-usability tradeoff is not inherent but tunable, with its characteristics varying significantly across different model families. We present this as foundational work, offering the proposed taxonomy and composite metrics to initiate broader discourse.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC