Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Semantic Distractor Robustness in LLMs: Exploring the Underlying Mechanisms

AbstractLarge language models drop in performance when their context contains semantic distractors: context topically related to the question but off the reasoning chain. Prior mechanistic work tracks a single injected distractor token, missing failures in which the model predicts other context tokens or out-of-context but topically related ones. We instead track the model’s actual top non-gold prediction, providing a localization target for every observed top-1 failure in the analyzed subset. Failures converge on a small set of mid-to-late-layer answer-selection heads (AS-Heads): patching their clean outputs into distracting runs recovers 60-86\\%$ of failed predictions across four base models from two families (random heads recover $\leq 8\\%$), with the mechanism also replicating in two instruction-tuned models. Clean-prompt ablations indicate that these heads also contribute to normal answer selection, so their role is not specific to distractors. Redirecting their attention onto an entity span causes the model to transcribe that span at $22$-$83\\%$ (vs. $\leq 7.4\\%$ for random heads), establishing a causally consequential intervention point. The localized heads transfer across MuSiQue, HotpotQA, 2WikiMultiHopQA, and non-Wikipedia RAG settings. At inference, their attention provides a training-free context-filtering signal that lifts F1 by up to $+35$ points on disrupted samples and is competitive with retrieval-head methods.

More from EMNLP 2026

All EMNLP 2026 papers →