PACE: Anytime-Valid Acceptance Tests for Self-Evolving Agents

Researchers introduce PACE, a statistical testing framework that prevents self-evolving AI agents from committing false improvements to their own prompts and workflows. Unlike naive greedy acceptance rules that accumulate errors through repeated testing, PACE uses sequential hypothesis testing to distinguish genuine improvements from noise, reducing harmful modifications by 30-42% while maintaining accuracy at lower computational cost.

Self-evolving agents represent a frontier in AI development, where systems autonomously iterate on their own instructions to improve performance. The critical vulnerability identified here lies not in proposal generation—the focus of most research—but in acceptance criteria. Greedy acceptance (keeping changes when scores improve) creates a statistical trap: when evaluated repeatedly against noisy development sets, the agent unknowingly exploits random fluctuations, committing false positives that accumulate into drift and degradation. This mirrors p-hacking in scientific research but applied automatically at scale.

PACE addresses this through anytime-valid testing, a framework ensuring that each decision maintains a user-specified false-positive rate regardless of how many evaluations occur or when stopping happens. By comparing candidates against incumbents on identical test instances and using betting-based e-processes, the method controls error rates without requiring predefined sample sizes. Empirical results on Qwen models demonstrate the approach's effectiveness: when genuine improvements exist, PACE commits the real one while rejecting noise; when no improvements are available, greedy agents spuriously modify themselves 13-21 times per run and degrade baseline performance by up to 4.9 points, while PACE maintains stability.

This work has implications for autonomous AI development pipelines increasingly used in production systems. As agents become more self-modifying, acceptance logic becomes a critical safety mechanism preventing unwanted drift. The 18% reduction in evaluation costs also addresses computational efficiency concerns in resource-constrained settings. The research suggests that robustness in self-evolving systems depends equally on acceptance criteria as on proposal quality—a shift in how developers should architect autonomous improvement loops.

• →PACE prevents self-evolving agents from committing false improvements by applying statistical hypothesis testing to acceptance decisions.
• →Greedy acceptance rules cause agents to accumulate 30-42% false and harmful edits when improvements are scarce, degrading baseline performance.
• →The framework maintains user-controlled false-positive rates through anytime-valid testing, allowing early stopping without statistical penalty.
• →PACE achieves comparable accuracy to greedy methods while reducing evaluation costs by 18% and preventing spurious self-modifications.
• →Acceptance criteria are as critical as proposal generation for reliable autonomous agent self-evolution.