Don't Blindly Trust It: How Unreliable Feedback Breaks Tool-Using LLM Agents

2026-06-19

AI Summary

This study investigates the effectiveness of tool-augmented large language model agents under unreliable external feedback, revealing that misleading feedback can cause their performance to drop significantly below that of a no-feedback baseline—a phenomenon termed “performance reversal.” By rigorously controlling agent architecture, prompting, action space, and decoding strategy while varying only the feedback type (faithful, misleading, or absent), the authors systematically evaluate feedback reliability on HotpotQA. Experiments show that Qwen2.5-7B achieves 44.8 F1 with clean retrieval, 22.3 F1 without feedback, but plummets to 4.7 F1 under misleading retrieval; this reversal persists even with strong retrieval systems and locally plausible distractors. The work underscores the necessity of including no-feedback matched controls for accurate tool-value assessment and proposes trajectory failure prediction and evidence rejection mechanisms to mitigate the impact of misleading feedback.

Abstract

Tool-augmented agents are typically evaluated by their gains under reliable external feedback. Yet these gains leave open a key counterfactual: when feedback is unreliable, would the agent be better off receiving no task evidence? We study this question with a controlled matched-loop comparison that fixes the agent loop, prompt, action space, and decoding, while varying only the returned observation: faithful, misleading, or absent. Across question answering and fact verification, persistent misleading feedback produces a value inversion: agents that benefit from clean tools can perform worse than the matched no-feedback fallback. On HotpotQA, Qwen2.5-7B reaches 44.8 F1 with clean retrieval and 22.3 F1 with no feedback, but drops to 4.7 F1 under shuffled retrieval. The inversion persists under stronger clean retrieval and locally plausible distractors, but weakens when later clean evidence can repair the trajectory. Early trajectory signals predict many failures, yet simple repairs remain fallback-limited: rejecting bad evidence helps only when the exposed fallback is reliable. These results show that clean-tool gains can overstate tool value, and that matched no-feedback fallback controls are necessary for evaluating tool-augmented agents.

Problem

• unreliable feedback
• tool-augmented agents
• value inversion
• no-feedback fallback
• LLM agents

Innovation

• tool-augmented agents
• unreliable feedback
• matched-loop comparison
• value inversion
• fallback control