Interpretable learned rules improve agent intervention reliability

Learn then Bet: Learning Interpretable Rules for Reliable Agent Intervention

Artificial IntelligenceMachine Learning

Summary

When autonomous agents run without much human help, simple rules often decide when to stop or assist them, but these rules can make mistakes. The authors introduce a method called LeBet that learns easy-to-understand rules from data to decide when to intervene, doing so only when there is strong evidence. Their approach comes with mathematical guarantees that limit false alarms and improve detection based on how the system behaved in training. Tests with a large language model showed better precision and saved communication costs compared to simple rules.

What this means in practice

  • For autonomous system operators: Implement precise intervention rules that reduce unnecessary stops and improve success rates of long-running agents using learned interpretable monitors.
  • For machine learning engineers: Use the LeBet framework to train compact, rule-based monitors that provide reliable intervention decisions with statistical guarantees from non-iid trajectory data.

Authors

You-Lin Chen, Xinyi Lisa Chen, Pedro Herrero-Vidal, Xiaojing Yu, Bingqing Song, Bin Xu

Abstract

Autonomous agents now run for hundreds of steps with little or no human supervision. In practice, practitioners use runtime monitors such as turn limits and keyword filters to decide when a run should be stopped, upgraded, or assisted. However, these heuristics may stop runs that would otherwise succeed or allow failing runs to continue. To address this limitation, we introduce LeBet (Learn then Bet), a statistical learning framework that derives its learning objective from sufficient conditions for valid and powerful intervention. In particular, LeBet learns an interpretable OR-of-AND monitor and accumulates evidence through a fixed betting statistic, intervening when the evidence reaches a prescribed threshold. Under finite within-class exchangeability and leave-one-out stability, we derive finite-sample guarantees for false-positive control (validity) and detection power conditional on the training data, using the same trajectories for learning and calibration without assuming independent and identically distributed trajectories. To address the intractable combinatorial optimization problem in our learning formulation, we use stochastic local search that leverages the monitor's Boolean structure to propose rule edits aimed at improving empirical separation. On SWE-bench Lite with Ministral 3 8B, the learned monitor saves $34.46\%$ of input-plus-output tokens and increases output precision from $18.33\%$ to $21.59\%$. The empirical false-positive rate is $13.64\%$. Our results show how to formalize the learning problem over interpretable rule sets with statistical conditions and establish training-conditional guarantees for reliable intervention based on partial histories.