acceptodds
Under review as a conference paper at ICLR 2027

The Intrusion Pathway: How Non-Evidential Context Degrades Language Model Performance

Abstract

Language models lose accuracy when a prompt contains non-evidential context: irrelevant statements from unrelated domains, or random off-distribution tokens, with no competing answer and no semantic overlap with the query. Such noise pervades retrieval-augmented, tool-using, and long-context deployments, yet the failure has lacked a mechanistic account. We trace the corruption to the Intrusion Pathway, which starts with MLPs enriching both query-subject and irrelevant-token representations, parroting heads carry prefix-token content to the prediction position, and relational readers largely avoid unrelated sources. Nevertheless, preserved subject enrichment and selective reading do not prevent intrusion in the broader computation. We then localize part of this interference to a sparse set of attention heads, which we call intrusion heads. Their prefix-derived writes promote distractor-related content, while their reduced support for the correct answer arises primarily from weaker non-prefix contributions rather than direct suppression by the prefix. The same heads transfer across tasks and distractor types and are largely weight-disjoint from the heads that integrate relevant contextual evidence, enabling a class-conditional intervention. Replacing the corrupted last-token residual with its clean counterpart, an oracle probe that requires a matched clean forward pass, recovers 97.6% on average of the distractor-induced drop across four downstream tasks. A deployable mask level knockout of about ten heads, roughly 1% of the model’s attention heads, recovers 57% at bounded cost to helpful context.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, contextual distraction, irrelevant context, language model robustness, causal intervention, retrieval-augmented generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_P1HEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates the problem of distractors appearing in the context of language models. The authors argue that a certain number of attention heads attend more to this distracting context while reducing their contribution to the correct answers (intrusion heads), and propose a remedy that masks these heads’ attention to the context. In simple math and question answering tasks, across 4 language models, the authors report that the remedy recovers about 57% of the accuracy drop from distractors.

Critical strengths
  • How to avoid irrelevant context tokens adding noise is an interesting question.
  • Some figures are nice (eg Figure 4)
Critical weaknesses
  • The paper is not well written. Besides how unclear some parts are individually, it overall lacks a coherent and well thought structure and flow. The language is not polished at all, and it is frequently hard to parse. In addition, the paper is overly complex to understand given the simplicity of the main idea. Overall the paper is far from being the best version of itself from a clarity and flow point of view.
  • The paper lacks clarity and rigor in the definitions, in the equations (see question about Equation 3, for an example), in the experimental setup of each experiment (almost never provided in the main paper). Overall, it is very hard from a careful read to understand what is going on, besides the overall concept of “finding a way to silence the contribution of irrelevant tokens”.
  • Significant parts of the paper are not backed by results nor properly explained experimental setups within the main paper. As an example, section 3 seems to show a lot of findings without showing a single numerical result (besides the last paragraph out of 5), nor clarifying what the experimental setup is and clearly stating what the take away is. It is not very informative and also hard to verify without spending a considerable amount of time looking at the appendix. If the authors believe these paragraphs should be part of the main paper, then they should be self-contained and should not be necessary to go to the appendix to understand basic experimental setup explanations or the results themselves. Otherwise, the entire section should be moved to the appendix.
  • Related to the previous point, the appendix is used way too much. Most paragraphs have references to at least one appendix paragraph. Sometimes more. The main paper refers to the appendix twenty-eight times. Overall, the main paper is far from being self-contained within the page limit.
  • The paper only plays with very simple tasks, and it is unclear how much any of these findings transfer to real world settings, and how to operationalize them without these controlled settings.
  • Some results reported in the main paper are quite odd: for example, Figure 7 shows a clean (no distractors) performance that changes across the number of distractors.
Questions for authors
  • Why does Equation 3 need to contain the nctx quantity? Can you better explain what Equation 3 is measuring? Why is the intrusion score defined this way?
  • In Figure 7, why does clean change with the x axis? If x axis means number of distractors, and clean has no distractors, it should be a horizontal line and not decrease with more distractors.
  • Which models are used in Figure 7 and Table 2? How are the results put together?
4 comments
AuthorsOpenReview ↗

Response to Reviewer P1HE

Thank you for the detailed review. Your review helped us make the main text more self-contained and helped sharpen the presentation.

Q1. "Why does Equation 3 need to contain the nctx quantity? Can you better explain what Equation 3 is measuring? Why is the intrusion score defined this way?"

The score asks two things of each head. Does it read the irrelevant text, and does it lose support for the correct answer when that text is added? The nctx run is the same question with no context, so it is the reference for the second part. The difference inside the max is the support for the answer that the head loses once the irrelevant text is added, and the max keeps only losses. Without this reference, a head that never supports the answer could not be told apart from a head that lost its support. The weight a_h is the head's attention to the irrelevant text. A head therefore scores high only if it both reads that text and loses support for the answer. The text after Eq. 3 now states this for each term (Section 4). We have added this explanation in the revised main paper as well. Kindly refer to General Response Point-4.

Q2. Clean accuracy in Figure 7 changes with the number of distractors.

This follows from how GSM-DC (Yang et al., 2025) is built, which differs from the other benchmarks we use. Every GSM-DC problem is a dependency graph of reasoning steps, and irrelevant context is added as k extra nodes in that graph, not as text appended to a fixed question. Therefore, in our evaluation each value of k has its own prompts and its own clean reference, which is that k's problems without the injected nodes, so the whole prompt changes from one k to the next. The clean accuracy at a given k is the accuracy on that k's problems without the irrelevant nodes and not one fixed problem set, and hence its accuracy moves with k. For this reason each k has to be read on its own, and every comparison in Figure 7 is made within the same k. Appendix D.1 noted this in the previous version as well, and the caption of Figure 7 now states it in the revised main text.

Q3. Models and aggregation in Figure 7 and Table 2.

Figure 7 shows the mean over the four 7B-scale models (Llama-3.1-8B, Mistral-7B-v0.1, Qwen2.5-7B and Pythia-6.9B) at each k, with binomial standard errors. Figure D.1 gives each model separately. In Table 2, GSM-DC, HotpotQA and PubMedQA use the same four models, and GSM-IC uses Llama and Mistral with chain of thought. Each entry is the share of the lost accuracy that the intervention recovers, and the last row is the mean of the four tasks. Both captions now state the models.

Simple tasks. We use the simple tasks only to find the mechanism: tracing needs one predicted token and prompts that differ only in the added text. The fix is tested on harder tasks (Section 5): multi-hop QA (HotpotQA), biomedical QA (PubMedQA) and multi-step arithmetic (GSM-DC, GSM-IC).

Use outside controlled settings. Controlled prompts are needed once per model, to find the heads. Applying the mask needs three things: access to the model's attention, the head list, and the position of the context. A system that assembles its own prompt already has that position. No clean run, no retraining [and no label of which text is irrelevant] is needed. These are still benchmarks with injected distractors; tool outputs and very long contexts are untested, and the Limitations section now says so.

Experimental setup and results in the main text.

Thank you for raising this point, which made the main text stronger. Section 2 gives the setup shared by all experiments, that is the prompts, tasks, models, metrics and notation. Sections 3 and 4 follow the irrelevant text through the model one step at a time, so they give the setup of each analysis next to its result instead of in a separate methods section. The introduction now ends with a short passage that says so. Following your comment, we made the setup of each analysis explicit in the text and added the key result as a number for most of the analyses in these two sections.

Use of the appendix.

The paper traces a pathway, and each step of it rests on several analyses and their controls. In the main text we gave priority to the chain of reasoning, so that a reader can follow each step and the checks that rule out other explanations for it. The complete results behind every step, that is the full tables, per-model values and controls, do not all fit in the main text, so we placed them in the appendix and pointed to them where they are used. The main text now also carries the setup and the key numbers, as described above, and the pointers lead to the full detail behind them.

Thank you again for the close reading. We would be glad to clarify any point that remains unclear.

AuthorsOpenReview ↗

Follow-up to Reviewer P1HE

Thank you again for your detailed review. We acknowledge your concerns about the writing and the heavy use of the appendix. Each statement in the paper rests on its own result, and all of them cannot fit within the page limit. In the revised version, each analysis now gives its setup and key result in the main text, and the appendix is used only for the fuller detail. We have also rewritten large parts of the paper with more explanation and less jargon, and our response answers your three questions on Equation 3, Figure 7 and the models used.

Since the revised paper differs considerably from the version you reviewed, we would be grateful if you could take a look at it and at our response. If they address your concerns, we hope you would consider reflecting this in your assessment.

I thank the authors for their rebuttal. However, I am still not satisfied, and my opinion remains mostly unchanged.

I re-read the paper, and the abstract and introduction are now clearer, but the other presentation problems remain. Sections 3 and 4 still read like dense summaries of too many results without details, contextualization, clearer word choice, or appropriate pacing, with too many important appendix references. My view has not changed: either these results are important and should be properly self-contained within the main text, or they are not important and should be left out altogether. Currently, reading the paper and understanding all steps is still too complicated because of these issues.

Besides the presentation issues, I have a concern about the experimental side. I was looking at the code and it looks like there might be an issue with HotpotQA and PubMedQA prompts. The gold passages are always added at the end of the context. The default maximum context is 3000 characters in the provided code for the downstream tasks results (I assume to avoid the risk of going past the context limit of the models). In practice, this cuts the gold paragraphs from “distracted” contexts in many cases depending on k, and omits them altogether for high values of k. I am fairly certain the cut happened in the paper experiments because my reproduction closely matches the accuracies reported in the paper for “distracted”. Removing the limit restores most of the “clean” accuracy for the “distracted” prompts. Could the authors confirm whether the paper experiments used this limit, and what changes without it?

Thank you for re-reading the paper and for the detailed follow-up. We are glad the abstract and introduction now read more clearly. We take your point that Sections 3 and 4 are still dense and lean on the appendix, and we will keep restructuring them so that each result is either self-contained in the main text or moved out of it. We are also checking the prompt-construction point you raise for HotpotQA and PubMedQA and will act accordingly.

Reviewer_YUHrOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks the question of why irrelevant context lowers the logit of the correct answer. It studies five open models on factual recall problems and find that heads that attend to the prefix and lose direct support for the answer, and calls them intrusion heads. Masking about ten of them recovers 57% of accuracy loss on four benchmarks. Patching the last-position residual from another clean run found to be recovering 97.6%.

Critical strengths

The main strength from me is that the question is genuinely interesting. Useful question, careful analysis. Interventions are compared against matched random, layer, and score controls with bootstrap intervals. The paper checks whether selected heads are separable from heads that use helpful context.

Critical weaknesses

The mechanistic story doesn’t necessarily connect to the downstream result. On factual recall, where the heads were selected, the K10 mask moves Llama’s gold logit by about 0.01 to 0.02 against a prefix-induced drop of 1.5 to 1.9. Yet the same heads recover 57% of accuracy loss downstream. The paper never explains this gap.

Downstream evidence is thin. Sample sizes promised at line 396 are missing, HotpotQA and PubMedQA have no confidence intervals, and the mask’s cost on clean inputs isn’t reported downstream. The 97.6% oracle needs a clean forward pass, so it says nothing about deployment.

Questions for authors
  • How do you explain a mask that barely moves the factual-recall gold logit recovering about half of the accuracy loss downstream?
  • What are the sample sizes and confidence intervals for Tables D.5 and D.6? What does the mask do to accuracy at k=0?
  • Cross-task overlap is almost 10/10 in Table D.1, but bootstrap resamples on the same task give a Jaccard of only 0.49 to 0.67 Table C.1. Why are the sets so stable across tasks?
  • I come across the paper arxiv 2605.10828 and wonder if the conclusion share anything in common
5 comments
AuthorsOpenReview ↗

Response to Reviewer YUHr (1/2): the link between the mechanism and the downstream result

Thank you for the careful review. Your first question is the central one for the paper, so we answer it at length here, together with the oracle point. Sample sizes, intervals, k = 0, head-set stability and the related paper follow in a second comment.

Q1. "How do you explain a mask that barely moves the factual-recall gold logit recovering about half of the accuracy loss downstream?"

We cannot yet explain it in full, and we prefer to say so plainly. Below is what the evidence supports, what we tried, and what stays open.

What the evidence supports and some new experiments The same heads are involved in both settings. Recomputing the intrusion score on each downstream task selects 9 or 10 of the same ten heads in every model (Table D.1). The gain is also specific to them. Masking ten random and layer-matched heads instead recovers 1% and 2.5% of the lost accuracy on HotpotQA respectively, against 48.5% for the intrusion heads. It scales with the intervention as well. With a graded mask, recovery rises with the mask strength, from 15% to 49%. These are new experiments and both their results will be added in the updated Appendix D.

What we tried and what stays open. We could not turn these observations into a quantitative account of the downstream gain. A controlled toy task of 20 GSM IC chain of thought quries with each token labeled reproduced the effect of the mask, but its numbers do not carry over to the downstream tasks. The control and the graded mask show that the gain depends on these heads. They do not show how a lower distractor logit becomes a higher accuracy. Showing that requires tracing and patching the generated reasoning token by token, with each token labeled as coming from the question, the evidence or the irrelevant text, so that the shift in the output distribution and in confidence can be followed along the chain. That is a study in its own right and The Limitations and future work now state this as open.

What the paper claims. The paper does not claim that the factual-recall logit effect accounts for the size of the downstream recovery. Section 4 states that on the one-token task the mask's effect on the answer is small next to the damage. The claim is that heads found through the mechanism on a simple task are effective and reusable intervention targets on longer tasks, and the downstream results test that claim directly.

And finally, we thank the reviewer for providing us with an important direction for future work to further strengthen the field's understanding of irrelevant context.

The 97.6% oracle.

We agree. The patch needs a clean run of the same question, so it says nothing about deployment. The abstract, the contributions and the Limitations state this explicitly, and Section 5 uses the patch only as a reference for how much of the loss is reversible at the answer position. Every deployment claim rests on the mask.

AuthorsOpenReview ↗

Response to Reviewer YUHr (2/2): downstream evidence, head-set stability and related work

Q2a. Sample sizes.

You are right that they were promised and missing. Appendix D now gives them. HotpotQA questions are drawn from the distractor dev set (7,405 questions), PubMedQA from the expert-labeled PQA-L set (1,000 questions), and GSM-IC from the 4,000-example subset used by Shi et al. (2023).

Q2b. Confidence intervals for Tables D.5 and D.6.

Both tables now give a 95% interval for every accuracy and for the mask's gain over the distracted run.

Q2c. The mask at k = 0.

At k = 0 the prompt holds only the question and its supporting passages, so this measures what the mask costs when all of the context is useful. On a prompt with no context at all the mask changes nothing, because there is no attention edge for it to remove. We also now call the intervention a mask throughout. Table 2 and Figure 7 called it a knockout, which wrongly suggested that the heads are removed.

Q3. "Cross-task overlap is almost 10/10 in Table D.1, but bootstrap resamples on the same task give a Jaccard of only 0.49 to 0.67 Table C.1. Why are the sets so stable across tasks?"

The two numbers describe different sets. Table D.1 compares the ten-head sets used downstream. The Jaccard in Table C.1 is for twenty-head sets under resampling, where the lower-ranked heads have close scores and trade places. The same table shows that 10 to 14 of the twenty are selected in at least 80% of resamples, and the ranking as a whole is stable, with split-half Spearman correlations of 0.97 to 0.99 (Section 5). The ten heads used downstream sit at the top of the ranking and are mostly the same under resampling as well. All ten are selected in at least 80% of resamples ( we will update the new drafts appendix with this new experiment in Table C.1 )

Q4. arXiv 2605.10828.

Thank you for the pointer. It is a very relevant paper, and we now cite and discuss it in the introduction. Gao et al. (2026) study the same problem from the side of the context. They vary the share of hard, topically related distractors in a long context and find that the first small share causes most of the accuracy drop, because these distractors draw a disproportionate share of attention. We share the conclusion that the damage runs through attention to the distracting text. The papers differ in two ways. Their distractors are related to the question, while ours, in the mechanistic analysis, are unrelated to it and contain no competing answer. They describe how attention is divided over the whole context, while we locate the specific heads whose support for the answer falls and intervene on them.

Their paper also bears on your first question. Their account is at the level of the whole context and ours at the level of a single predicted token. The gap you point to lies between the two, in generated reasoning, where irrelevant content can enter at every step and build up. Studying how it accumulates along a reasoning chain is the next step for us, and their results will help us take it.

Thank you once again for the careful and expert review. The mechanistic analysis you credited is kept in full, and Sections 3 and 4 now give each setup and key result in the main text so that it is easier to follow. We would be glad to clarify any point that remains unclear.

AuthorsOpenReview ↗

Follow-up to Reviewer YUHr

Thank you again for your careful and expert review. We found it extremely helpful, and your first question has opened a future-work direction for us. Connecting the mechanistic analysis to the downstream result- that is, how a change in logits becomes a gain in accuracy- is a study in its own right, and we will pursue it in our future work. Our response also covers your other questions, with new specificity and graded-mask controls, the missing sample sizes and confidence intervals, the head-set stability, and a discussion of Gao et al. (2026). The new results reported in our response will be added to the paper, and we will upload the updated version once you acknowledge.

We would be grateful to hear whether our response addresses your concerns. If it does, we hope you would consider reflecting this in your assessment.

Reviewer_9SqKOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper investigates why language model performance degrades when irrelevant context is included in the prompt. The authors find that the model's knowledge about the question stays generally intact, and the problem, which they call the intrusion pathway, is caused by a small set of attention heads in later layers that focus on the unrelated text. The authors find that these heads are generally disjoint from those that use helpful context. Finally, they find that blocking these heads from attending to the context recovers about 57% of the lost accuracy on four benchmarks.

Critical strengths

The paper answers an interesting open question that is of practical relevance, namely why irrelevant context decreases support for the correct answer. While previous work showed that including irrelevant context in the prompt lowers the accuracy of language models and that the models have attention heads that promote tokens from the irrelevant context, this question was largely unaddressed. The paper includes a mechanistic explanation for this effect and offers a simple solution, making it a useful contribution to the mechanistic interpretability literature.

Critical weaknesses

I believe the main weakness of the paper is presentation. I found it jargon-heavy and generally quite hard to follow. The abstract and introduction contain a lot of terms (e.g., parroting, intrusion heads, intrusion pathway, relational readers, etc.) that are not properly introduced before being used. I believe most of these are not standard. Even if they are, adding an intuitive, high-level explanation would make the paper much more accessible to a broader audience. Some sentences, particularly in the abstract, are very convoluted and took me several reads to parse. Also, I suggest emphasising the contributions more in the introduction and positioning the paper better in relation to prior work. I think this could be a good paper, but it seems like it would take quite a bit of rewriting.

Questions for authors
  1. If I understand correctly, your method does not need to know which context is useful and which is not. Is this correct?
  2. How would the method be used in practice when the boundary between the context and the question is unclear?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 9SqK

We sincerely thank you for the careful reading and the encouraging assessment. Your comments helped us make the paper readable for a general audience and place it more clearly against prior work.

Answers to your questions :-

Q1. "Your method does not need to know which context is useful and which is not. Is this correct?"

Yes, that is correct. The mask does not need to know which parts of the context are useful. It blocks the selected heads from reading all context tokens, helpful or not, and the selectivity comes from which heads are masked. Intrusion heads are chosen to exclude the heads that read useful evidence, so masking them costs little when the context is helpful. When the context states the answer (golden context), masking the intrusion heads lowers the probability of the correct answer by only 0.7 to 2.4% of its value across the five models (Table C.7).

Q2. "How would the method be used in practice when the boundary between the context and the question is unclear?"

The mask needs to know where the context is. In most deployments this is already known, because the system itself assembles the prompt. Retrieval pipelines insert passages as delimited blocks, tool outputs arrive as separate tool messages, and chat templates mark every role, so the mask can use these markers directly. The context also does not have to come before the question. When we moved the irrelevant text after the question, or between the subject and the relation, the intrusion heads' attention followed it and the damage stayed within the confidence interval of the original setting (Appendix A.6).When no marker is available at all, the same ten heads can instead be switched off entirely by zeroing their output. This needs no boundary and recovers 39% of the lost accuracy on HotpotQA and 52% on PubMedQA, against 48.5% and 63.3% for the mask ( We will add this new experiment to Appendix D ) . Switching a head off also removes what it contributes from the question (Section 4), so the mask remains our default whenever the context span is known.


Presentation and Positioning :-

We accepted all of your comments on presentation, and they changed both how the paper reads to a broader audience and how it stands against earlier work. The revised PDF is uploaded, and our general response shows these changes before and after.

Simpler language. Your comments showed us where the language had to be simpler. And how we can better introduce the Jargons and overall reduce their use in the main paper. Kindly refer to General Response Point-1.

Readers outside the subfield. You also pushed us to explain the work for a general audience, because irrelevant context is a problem for the whole field and not only for interpretability researchers. We tried doing that as well, and made changes in that account. Kindly refer to General Response Point-2.

Contributions. The introduction now lists three numbered contributions: the mechanistic account, the finding that intrusion heads are separate from the heads that read useful evidence, and the test-time fix.

Positioning against prior work. A new related-work paragraph in the introduction (lines 67 to 76) places the paper against earlier work. Behavioral studies show that irrelevant context lowers accuracy (Shi et al., 2023; Wu et al., 2024; Yang et al., 2025). The two closest mechanistic accounts explain neighboring cases, namely topically related distractors that compete for attention on retrieval heads (Gao et al., 2026) and heads that promote words already present in the context (Niu et al., 2025). Neither explains why support for the correct answer itself falls when the text is unrelated, which is the question we answer. Existing remedies usually retrain the model, through noise-robust fine-tuning (Yoran et al., 2024) or by learning to critique retrieved passages (Asai et al., 2024). Our fix works at test time instead. It changes no weights and needs no retraining or extra forward pass. The heads are chosen once per model on a simple factual-recall task and reused unchanged on all four downstream tasks, so they do not need to be found again for each task.

We believe the abstract and introduction can now be read without a background in mechanistic interpretability, and we would welcome a note on any passage that is still hard to follow. We hope the revision is the paper you thought this could be.

AuthorsOpenReview ↗

Follow-up to Reviewer 9SqK

Thank you again for your thoughtful and insightful review. We have answered your questions in our response: the mask does not need to know which context is useful, and when the context boundary is unclear, the same heads can be switched off entirely with most of the gain kept. Your comments on presentation were extremely helpful, and we have rewritten the abstract and introduction in plainer language with clearer contributions and positioning. We believe the paper is now a better version of itself because of your feedback. The new result reported in our response will be added to the paper, and we will upload the updated version once you acknowledge.

We would be grateful if you could look at our response and the revised PDF and let us know whether they address your concerns. If they do, we hope you would consider reflecting this in your assessment.

Thank you very much for carefully addressing my comments and for the new draft. I didn't have time to read the whole revision carefully, but the abstract and introduction seem vastly improved since the submission. If the paper is accepted, I encourage the authors to think carefully and keep iterating on the main text to improve its readability. I was already in favor of the acceptance of this paper and I will maintain my score.

We thank the reviewer for engaging with our revisions and for acknowledging the improvements in the writing. We will continue to revise the paper and improve the structure and presentation of the main text. We also thank the reviewer for recognising the value of our methodology and contributions, and for maintaining their positive score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: summary of revisions

We thank all reviewers for their careful and constructive reviews. Reviewers found the question interesting and practically relevant, the analysis careful, with matched random, layer and score controls, and the fix simple. The shared requests were for clearer writing, a more self-contained main text, sharper definitions and fuller downstream reporting. The revised PDF addresses each of these, and the main changes are summarised below with their locations in the revised PDF.

1. Inclusive language for a broad audience

  • The abstract and introduction are rewritten in plain language. New terms are now introduced with an intuitive explanation before they are used, several are replaced by everyday wording, and long sentences are split so that each carries one idea (lines 13 to 107).
  • Section 2 now explains copying with and without a cue, using induction as the familiar reference point, so parroting is defined by example before Section 3 uses it (lines 143 to 161).

2. Clear contributions and positioning

  • The introduction now lists three numbered contributions: the mechanistic account, the finding that intrusion heads are separate from the heads that read useful evidence, and the test-time fix (lines 91 to 104).
  • A new related-work passage places the paper against behavioural studies (Shi et al., 2023; Wu et al., 2024; Yang et al., 2025), retraining-based remedies (Yoran et al., 2024; Asai et al., 2024) and the two closest mechanistic accounts (Niu et al., 2025; Gao et al., 2026) (lines 67 to 76).

3. A self-contained main text

  • Section 2 now gathers the shared setup in one place: prompts, tasks, models, metrics and the recovery measure used downstream (lines 110 to 142).
  • Each analysis in Sections 3 and 4 now gives its setup next to its result and reports the key number in the main text. For example, lines 217 and 327 now report that copying heads shift at least three times more than control heads and that intrusion heads move 9 to 25% of their attention onto the irrelevant sentence.
  • The introduction now ends with a short roadmap of how the sections build on each other (lines 106 to 107).

4. Sharper definitions and captions

  • The text after Eq. 3 explains each term of the intrusion score. The attention weight measures whether a head reads the irrelevant text, the no-context run is the reference for how much answer support the head loses, and the max keeps only losses (lines 296 to 299).
  • The intrusion-head set is now defined concretely in the main text, including where the heads sit in the network and how stable the selection is under resampling (lines 300 to 303).
  • The abstract and contributions now clearly separate the oracle patch, which needs a clean run and serves only as a reference, from the deployable mask, which needs neither a clean run nor retraining (lines 27 to 31 and 101 to 104).
  • The captions of Table 2 and Figure 7 now state the models and how results are combined. The Figure 7 caption also explains that in GSM-DC (Yang et al., 2025) each number of distractors comes with its own problem set, so every comparison is made within the same k (lines 480 to 483).

5. Fuller downstream reporting

  • Appendix D now gives the sample size for every downstream task (lines 1512 to 1516).
  • Tables D.5 and D.6 now report a 95% interval for every accuracy.
  • The Limitations and future work section now sets out the scope of the evaluation and the natural next step: following irrelevant content through generated reasoning, token by token (lines 510 to 532).

6. New experiments, to be added to Appendices C and D

  • Specificity. Masking ten random or layer-matched heads recovers 1% and 2.5% of the lost accuracy on HotpotQA, against 48.5% for the intrusion heads.
  • Graded effect. With a graded mask, recovery rises steadily with mask strength, from 15% to 49%.
  • No context boundary needed. Switching the same ten heads off entirely needs no context boundary and still recovers 39% on HotpotQA and 52% on PubMedQA.
  • Stable head set. All ten heads used downstream are selected in at least 80% of resamples.

We believe these changes make the paper clearer, easier to verify and better grounded, and we would welcome any further suggestions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.