Reducing Redundant Actions in ReAct Agents via Semantic Filtering

ReAct-style language agents often exhibit redun- dant action execution, repeatedly issuing seman- tically similar tool queries that yield no new in- formation. This leads to inefficient reasoning and unnecessary computation, yet its impact on agent performance remains underexplored. We propose a simple semantic filtering mechanism that re- moves redundant actions during inference based on embedding similarity. When a duplicate action is detected, execution is skipped and a feedback signal is injected into the model context to guide subsequent decisions. Experiments on HotpotQA with Qwen2.5 and LLaMA variants show that our method improves F1 by up to 10% in a sam- pled subset of the HotpotQA evaluation set while reducing LLM calls by 5–6%, achieving up to 17% efficiency gains. We further observe a strong model-specific reasoning behavior: filtering ben- efits Qwen models but degrades performance on LLaMA models, suggesting that redundancy in- teracts differently with model-specific reasoning behavior.