Retry, Switch, or Abstain? Learning Strategy-Aware Tool-Use Policies via Controlled Error Injection

2026-08-12

AI Summary

This work addresses the vulnerability of large language models (LLMs) to transient, persistent, or silent failures during real-world tool invocation, where effective recovery mechanisms are often lacking. To enhance robustness without retraining, the authors propose BENCH2ROBUST, a framework that transforms standard benchmarks into stochastic environments via controlled error injection. It integrates structured runtime recovery strategies—retry, switch, and abort—with Bayesian Tool Memory (BTM) and curriculum reinforcement learning. Experimental results demonstrate that BTM alone improves robustness by up to 16.8 percentage points; when combined with curriculum reinforcement learning, the approach achieves success rates of 40.8–45.5% under error-injected conditions on the Retail task while preserving performance in error-free settings.

Abstract

Tool-using LLM agents are commonly trained and evaluated in environments where tool calls succeed reliably, yet deployed tools can fail transiently, persistently, or silently. Robust recovery therefore requires more than repeated retries: an agent may need to retry the same path, switch to an alternative, or recognize that no viable path remains. We present BENCH2ROBUST, a framework that converts failure-free tool-use benchmarks into controlled stochastic environments with scenario-controlled solvability, where episodes explicitly require retrying, switching, or stopping after available paths are exhausted. We use BENCH2ROBUST to study two complementary interventions: structured runtime recovery context through Bayesian Tool Memory (BTM), and curriculum-controlled reinforcement learning. Across 7 models from 4 families and two multi-turn benchmark families, tool failures produce a near-universal robustness gap. On held-out Retail tasks, BTM improves robustness by up to 16.8 percentage points without retraining, while RL learns complementary recovery behavior that remains beneficial without inference-time BTM. Combining the two reaches 40.8-45.5% under injection while preserving failure-free performance. These results suggest that robust tool use benefits from combining environment-specific recovery knowledge with learned recovery behavior.

Problem

• tool-use robustness
• failure recovery
• LLM agents
• stochastic tool failures
• strategy-aware policies

Innovation

• tool-use robustness
• Bayesian Tool Memory
• controlled error injection
• reinforcement learning
• strategy-aware recovery