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

The paper presents a study on the impact of unreliable feedback on tool-using LLM agents, specifically evaluating how misleading information can lead to a value inversion where agents perform worse than without feedback. Using the Qwen2.5-7B model on the HotpotQA benchmark, the research shows significant performance discrepancies: 44.8 F1 with reliable retrieval drops to 4.7 F1 under misleading conditions. These findings emphasize the need for robust fallback mechanisms and careful evaluation of tool-augmented agents, as relying solely on external feedback may misrepresent their effectiveness.