When the Distractor Shares the Word: Sense Collisions as a Controlled Probe of Contextual Distraction

Language models are easily thrown off by irrelevant text in their prompts. They are also, on clean single sentences, very good at telling which meaning of an ambiguous word is intended. This proposal asks what happens when these two facts meet. Suppose a passage placed next to the question uses the same word in a different meaning: a paragraph about lamb as meat beside a question about lambs as animals, or a retrieved document about a river bank beside a question about bank loans. Such text is the most natural kind of distractor a search system can return, because it matches the query's surface form. It is also an unusually clean experimental tool. Ambiguous words come with human-measured degrees of relatedness between their meanings, so we can vary how similar the distractor's meaning is while keeping the shared words fixed, and the reverse.

Our central claim is that a model's vulnerability to a same-word, different-meaning distractor can be predicted from how well the model separates the two meanings in clean context. Meanings that the model already blurs internally should be the ones a distractor most easily hijacks. If this holds, distraction is not only a generic "attention dilution" problem. It has a lexical component that we can measure ahead of time, item by item, from clean-context representations. We test the claim on word-sense benchmarks and on question answering with retrieved passages. We then ask whether general noise-robust training methods transfer to these sense collisions or need sense-specific data.

What Is Known and What Is Missing

The distraction literature motivates the question but confounds its key variable. Large Language Models Can Be Easily Distracted by Irrelevant Context built GSM-IC by inserting irrelevant sentences into math problems. It found that "in-topic" distractors with overlapping role names are the most damaging, with macro accuracy often falling under 30%. In that design, sharing surface words and sharing meaning always move together, so the study cannot say which one does the damage. Breaking Focus: Contextual Distraction Curse in Large Language Models used tree search to generate model-targeted distractors, causing roughly 45% average degradation (GPT-4o-mini fell from 85.7% to 22.0% on CommonsenseQA). It argues that the failure concerns ability rather than missing knowledge. That supports our framing, but its distractors are optimized rather than controlled, so they do not show which properties make text distracting. On the retrieval side, Making Retrieval-Augmented Language Models Robust to Irrelevant Context documents a "copying trap," where models copy entities from noisy passages. It proposes NLI filtering (keeping only passages a natural-language-inference model judges to support the answer) and fine-tuning on mixed gold and noisy passages (SA-RetRobust). The authors note that their noise was irrelevant rather than conflicting. Sense collisions sit between those two categories, and this is exactly where we probe.

The word-sense literature supplies the measuring instrument, but always in clean context. Patterns of Polysemy and Homonymy in Contextualised Language Models compared BERT, ELMo and Word2Vec against human similarity ratings for 28 polysemous nouns. It reports human cross-sense similarity of 0.17 for homonyms (unrelated meanings) and 0.73 for polysemes (related meanings), and a BERT Large correlation of $r = 0.69$. It also shows that models capture some regular alternations (chicken as animal versus meat) but miss parallel ones (lamb, pheasant). That item-level inconsistency is what our central prediction rests on. WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representations poses sense identity as a binary judgment over sentence pairs, with BERT Large at 63.8% against an 80.0% human ceiling. Do Large Language Models Understand Word Senses? shows GPT-4o reaching 82.5% F1 on the Maru et al. benchmark, on par with the supervised ConSeC (83.0%). The same paper shows GPT-4o falling to about 45.6% on hardEN and favoring candidate definitions listed earlier. All of these evaluations use isolated sentences. None places a competing-sense passage in the context.

The obvious study at this intersection would wrap WiC sentences in random paragraphs and report an accuracy drop. We do not propose that. Random padding mostly tests long-context degradation, which is already documented, for example the inverted-U curve in Beyond RAG vs. Long-Context: Learning Distraction-Aware Retrieval for Efficient Knowledge Grounding. Our design changes two things. First, it manipulates distractor meaning and distractor wording independently. Second, it tests a falsifiable, item-level link between clean-context representations and contextual fragility.

Core Claim

Fix a model and an ambiguous target word that appears in a query sentence with a gold meaning. A distractor is a passage that contains either the same word or a paraphrase of it, and uses either the gold meaning or a competing one. Let the separation score $s_i$ for item $i$ be the model's clean-context distance between the target word's representations in its gold-meaning and competing-meaning uses. For open-weight models we compute this from hidden states. For closed models we use a behavioral proxy: clean WiC-style accuracy on that sense pair. Let the hijack rate $h_i$ be the probability that a prediction that was correct in clean context switches to the distractor's meaning once the distractor is added.

We predict three things. First, $h_i$ is much larger for same-word competing-meaning distractors than for paraphrased competing-meaning distractors, which isolates the effect of shared surface form. Second, among same-word distractors, $h_i$ is higher for related meanings (polysemy) than for unrelated ones (homonymy). Third, and most important, $h_i$ is negatively correlated with $s_i$ even after controlling for human-rated relatedness. In other words, the model's own blind spots, not just the linguistic difficulty of the word, predict which items get hijacked.

Research Program

We first build a controlled distractor suite. Target items come from WiC, the Haber and Poesio metonymy dataset, the Maru et al. WSD benchmark, hardEN and 42D. For each item we create four distractor passages of matched length and position. These cover {same word, paraphrase} crossed with {gold meaning, competing meaning}. We add an off-topic control passage and a random-padding control. Competing-meaning passages are drawn from sense-annotated sentences for the same lemma, with the meaning verified by human annotators. Paraphrase passages replace the target word with a synonym of the stated meaning. We vary distractor length (one sentence up to several paragraphs) and position (before, after, or between the WiC sentence pair). For the multiple-choice benchmarks we also vary candidate-list order. This tests whether a competing-meaning distractor amplifies the first-position bias reported by Meconi et al.

Second, we build a retrieval setting. We select questions from NQ and TriviaQA whose answers hinge on an ambiguous term, filtered by the answer model's clean-context correctness. We pair each question's gold passage with retrieved passages that use the alternate meaning of that term. These come from a dense retriever's top results, filtered by annotators for meaning. This adapts the noisy-retrieval protocols of Yoran et al. and of Improving Zero-shot Reader by Reducing Distractions from Irrelevant Documents in Open-Domain Question Answering. The difference is that the noise is defined by meaning collision rather than by random or topical selection.

Third, we analyze the mechanism in open-weight models (Llama-3.1-8B-Instruct and Llama-3.3-70B-Instruct), with BERT Large as a reference encoder that links back to Haber and Poesio. We track how the target word's hidden state moves toward the centroid of the competing meaning as distractor length grows. We also measure how much attention the target word's position sends to the distractor's copy of the word. This tests a specific version of the dilution story: the hijack runs through attention to the repeated word form, and is not simply a general spreading of attention. As an intervention, we mask attention from the target position to the distractor's copy of the word. If masking removes most of the hijack effect, the surface-form channel is causal.

Fourth, we test remedies. The question is whether general noise-robust methods handle sense collisions, or whether sense collisions need their own training signal.

Baselines

The naive reference point is the same model with no distractor. Its clean accuracy defines the ceiling against which hijack rates are computed. The second reference is the same model with the distractor added and nothing done about it. Any remedy must close a meaningful share of the gap between these two.

The first prompting baseline is the instruction "ignore irrelevant information and use the meaning intended in the target sentence." This follows the instructed prompting of Shi et al., which raised code-davinci-002 from 72.4% to 77.8% micro-accuracy on GSM-IC. It costs nothing to run, so any trained method must beat it to matter. We run it with the same models and decoding settings. We also try three instruction paraphrases and report the best, so that prompt wording does not decide the comparison.

For question answering, the strongest inference-time competitor is Distraction-aware Answer Selection (DAS) from Cho et al. It combines document-selection instructions with an answer-score adjustment based on the likelihood of the query given each document. It reports an average 64% Exact Match gain over naive zero-shot readers, rising to 231% with 100 DPR documents on SQuAD. This method is the most serious threat because query likelihood may already penalize wrong-meaning passages. We run it on the same passage sets with the same reader model.

The strongest training-based competitor is SA-RetRobust from Yoran et al. This is fine-tuning on roughly 1,000 examples that mix gold and distracting passages. It reports gains of over 30 points relative to no retrieval on 2WikiMQA and stable accuracy under random distractors. Alongside it we run Ext2Gen from Aligning Extraction and Generation for Robust Retrieval-Augmented Generation. Ext2Gen trains the model to extract relevant sentences and then answer, aligned with preference optimization. It reports a rise for Llama-3.1-8B-Instruct from 0.473 to 0.557 average QA score. We train both on their original generic noise, using Llama-3.1-8B-Instruct and the same step budget.

The key ablation trains SA-RetRobust on an equal number of examples in which the distractors are competing-meaning passages, with no generic noise. The lemmas in this training set are disjoint from the test lemmas. We then compare generic-noise training against sense-collision training on held-out lemmas. Our hypothesis is that sense-specific training is needed. It counts as supported if sense-collision training lowers held-out hijack rates by a clear margin, which we set in advance at five points absolute, while generic-noise training does not. It counts as refuted if generic training matches it.

We also run NLI filtering as specified by Yoran et al., since it is the obvious filtering baseline.

Finally, for the multiple-choice WSD sets, the clean reference numbers are GPT-4o at 82.5% F1 on Maru et al., 76.8% on 42D, and about 45.6% on hardEN, all from Meconi et al. We found no reported result on Lacuna for any instruction-tuned LLM on WiC, nor for any model on WiC with distractors. We therefore establish those references ourselves.

Why not just…?

Why not just tell the model to ignore irrelevant text? Breaking Focus reports that such instructions fail against distractors that are coherent and plausible, and same-word distractors are coherent by construction. Still, the evidence comes from optimized distractors, not lexical ones, so we do not assume the outcome. The instructed-prompting baseline runs in every condition. If it closes most of the clean-to-distracted gap, we will report that sense collisions are easy to fix, and the diagnostic findings will stand without the remedy section.

Why not just filter passages with an NLI model or a cross-encoder reranker? Yoran et al. found NLI filtering too strict, because it discards partially helpful passages. A reranker, however, might cleanly reject wrong-meaning passages, since the query context usually signals the intended meaning. This is a real threat in the retrieval setting, and it does not apply to the WSD setting, where the distractor sits inside the model's own context rather than in a retrieved list. We report reranker and NLI rejection rates for competing-meaning passages separately from accuracy. If filtering removes them reliably, the retrieval part of the problem reduces to retrieval quality, and the contribution narrows to the in-context mechanism.

Why not just apply existing noise-robust fine-tuning? This is the most likely way our remedy claim fails, and the ablation above is designed to settle it. Yoran et al. explicitly left conflicting context open. Competing-meaning passages share the target word with the query, which may defeat a model trained to skip passages that look unrelated. If SA-RetRobust or Ext2Gen trained on generic noise matches sense-collision training on held-out lemmas, we will conclude that no special treatment is needed.

Why not just sample many answers and vote? Shi et al. report 99.7% recall with 20 self-consistency samples. Recall is not accuracy, though, and a systematic hijack pulls most samples toward the same wrong meaning. We test whether majority voting recovers hijacked items at 20 times the inference cost. If it does, this indicates the hijack is noise in decoding rather than a shift in the model's internal representation. That would argue against our mechanism, and we would report it as such.

Decisive Checks

The proposal stands or falls on three results. The first is whether same-word competing-meaning distractors produce substantially higher hijack rates than paraphrased ones of equal length. The second is whether the clean-context separation score $s_i$ predicts $h_i$ with a significant negative partial correlation once human relatedness is controlled. The Haber and Poesio items are the cleanest test here, because their lamb-versus-chicken inconsistencies give known low-separation cases. The third is whether masking attention to the distractor's copy of the word restores clean-context accuracy. If the first fails, lexical overlap is not the driver, and Shi et al.'s in-topic effect is about meaning rather than wording. If the second fails, distraction vulnerability cannot be predicted from representations, and the dilution account stands without a lexical refinement. Either outcome is informative. The remedy comparisons then show whether the field needs sense-aware robustness data or can rely on generic noise training.