Clarification Timing Does Not Mitigate Anchoring Bias in Tool-Using LLM Agents

Abstract

Tool-using LLM agents exhibit anchoring bias on ambiguous queries, disproportionately selecting the first-listed candidate regardless of correctness. We test whether forcing clarification \emph{before} retrieval exposure---an Intent Reconstruction Gate (IRG)---can mitigate this bias. Through controlled experiments on InteractComp-Anchor (210 instances) with Qwen2.5-7B-Instruct, we compare three conditions: search-first baseline, IRG clarify-first, and a candidate-order control. Our results show that IRG produces negligible improvement (+0.95 percentage points, $p = 0.42$) over the search-first baseline, while simply reordering candidates to place the target first achieves 13.33 percentage point gains ($p < 0.001$). The dominant factor in anchoring bias is candidate presentation order, not clarification timing. Practitioners should focus on candidate presentation strategies rather than clarification timing to mitigate anchoring bias in tool-using agents.