acceptodds
Under review as a conference paper at ICLR 2027

Context Poisoning as Extreme-Value Attention Interference in Long-Context Language Models

Abstract

Large language models can process increasingly long prompts, yet their ability to locate and use decisive evidence may degrade as irrelevant or confusable context is added. We formulate this phenomenon, which we call context poisoning, as extreme-value interference in attention: the decisive-evidence score is upper-bounded, while the maximum score among effective distractors grows with their number. Under a softmax retrieval abstraction, we derive a finite-sample upper bound showing that maintaining a fixed accuracy target above base rate requires the evidence margin to scale as Ω(√log N), where N denotes the effective distractor count rather than necessarily the raw context length. The analysis connects long-context degradation to score aliasing, positional aliasing, and softmax dilution. Controlled experiments show that retrieval accuracy decreases as total context grows in the presence of embedded hard negatives, that the same-format condition produces the largest observed accuracy drop among the tested distractor constructions at fixed context length, and that retrieval gating can improve evidence use while its net benefit depends on preserving evidence recall. These results motivate evidence bottlenecks, alias-resistant representations, retrieve-then-reason architectures, verifier-mediated memory, and contrastive anti-poison training.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large language models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_piFeOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper formalizes performance degradation in long-context LLMs as an extreme-value interference problem in attention. By deriving a theoretical upper bound for context poisoning, the paper claims to prove that maintaining retrieval accuracy requires the decisive-evidence margin to scale as Ω(log⁡N)\Omega(\sqrt{\log N}). Based on this, the authors propose several architectural suggestions and evaluate the effects of distractor count and confusability on model recall and reasoning through controlled experiments.

Critical strengths
  1. This paper introduces extreme-value theory to mathematically model attention competition. This perspective shifts the interference source from average noise to"maximum extreme-value noise.
  2. Comparing unrelated filler, same-domain, same-format, and near-alias distractors confirms that score aliasing induced by structural overlap is more damaging than raw context length.
Critical weaknesses
  1. Despite the novel angle, the paper suffers from serious internal contradictions and logical gaps. The weaknesses focus primarily on Eq.(7), whose bound rests on multiple strong assumptions (e.g., Gaussian-distributed distractor scores, and a quantifiable, constant relationship between evidence attention and answer correctness.):

1.1 The Gaussian i.i.d. assumption, the paper's own strong premise, is falsified by the paper's own experiments, causing the proof to collapse. It shows Shapiro-Wilk rejects normality, Ljung-Box rejects independence.

1.2 ρ may be a linear or nonlinear function rather than a fixed value. The cited Du et al. (2025) shows that even with perfect retrieval, long context alone harms performance, implying ρ decreases with N. If so, treating ρ as a constant ignores the core difficulty the paper itself cites.

1.3 An internal contradiction exists between ρ and Φ\Phi. The bound assumes attention high score = correct, which is unproven. But the introduction of ρ (ranging from 0 to 1) shows high score ≠\neq correct, so the causal chain may or may not hold. Yet Φ(⋅)\Phi(\cdot) assumes it holds deterministically. The two assumptions directly contradict.

1.4 The theorem assumes one decisive evidence and one attention head. Real multi-hop tasks require aggregating scattered evidence, and multi-head attention allows cross-head compensation. The 2log⁡N\sqrt{2\log N} constraint lacks corollaries in these settings.

  1. The paper only offers three high-level suggestions, none of which are implemented or empirically validated. The paper's substantive engineering contribution is minimal.
  2. Gemma 3 (12B) has a 96.1% API error rate (72 valid out of 1,840 calls). Claude has 35.4%. Failures concentrate on the hardest, longest queries, so the surviving data systematically under-represent the hardest cases. Core conclusions rest on highly truncated data.
  3. It uses conventional BM25 retrieval gating. At 512K, the gain is only +0.070 EM on 100 questions, with p = 0.092, not statistically significant.
Questions for authors

None

1 comment

We thank the reviewer for identifying assumptions whose formal role and empirical interpretation should be separated more clearly. We respectfully disagree that the appendix diagnostics cause the proof to collapse.

Gaussianity and independence. Theorem 1 is a conditional mathematical result and remains valid under its stated assumptions. The appendix does not observe the raw logits ZiZ_i in Eq. (5): it analyzes the logarithm of sentence-aggregated post-softmax attention after standardization. Because this statistic includes softmax normalization, sentence aggregation, position effects, and pooling across heads and layers, its non-Gaussianity is not a direct falsification of Eq. (5). We will correct the appendix sentence saying that this analysis tests the assumption "directly."

We also note that the Ljung-Box test does not reject independence in every setting; for example, the reported Gemma values are p=0.958p=0.958 and p=0.320p=0.320 at N=20N=20 and N=50N=50. The measured statistic nevertheless departs substantially from the Gaussian i.i.d. abstraction and should be treated only as a diagnostic.

The first step of the existing proof can be stated without Gaussianity or independence. Let

\qquad b=\gamma+\tau\log((1-\rho)/\rho).$$ Under $S_\star\le\gamma$ and the same one-sided faithfulness condition, $$A_N\le a_0+(1-a_0)\Pr(M_N\le b).$$ The i.i.d. Gaussian expression $\Phi(b/\sigma)^N$ is a closed-form specialization that yields the stated extreme-value scale. We will present the result in this order and describe the Gaussian case as an interpretable specialization, not a calibrated model of all transformer heads. **Meaning of $\rho$.** There is no assumption that high attention guarantees correctness. The condition is one-sided: $$\Pr(\hat y=y^\star\mid \alpha_{\star,N}<\rho)\le a_0.$$ For $\alpha_{\star,N}\ge\rho$, the proof uses only the trivial upper bound of one on correctness. Thus downstream reasoning may fail even when evidence attention is high, and there is no contradiction. Fixed $\rho$ is a uniform modeling assumption. If $\rho$ is replaced by $\rho_N$, the bound applies pointwise with the corresponding threshold, but the fixed-$\rho$ asymptotic need not follow; we will state this explicitly. Perfect retrieval alone does not imply a particular functional form for $\rho_N$, because evidence presence, attention allocation, and downstream use are distinct. **Scope and empirical role.** The theorem intentionally analyzes one evidence-selection bottleneck with one decisive item. It does not claim to model cross-head compensation or general multi-hop aggregation; these are limitations of scope, not contradictions in the stated result. The multi-hop gate experiment tests only the qualitative candidate-reduction/recall tradeoff. The endpoint failures limit the Claude and Gemma ablations, but they are not premises of Theorem 1. GPT-4.1 completed the fixed-length study, and the separate 700-pair gate experiment had zero endpoint errors. We will nevertheless report those results conservatively: the strongest gate effect is not statistically conclusive. The proposed losses and architectures will be presented as theory-derived implications and future directions, not as validated engineering contributions.
Reviewer_H91vOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper asks why long-context models get worse when irrelevant or look-alike text is added to the prompt. It calls this "context poisoning". The central claim is that the evidence must beat the strongest distractor, not the average one, and the strongest distractor gets stronger as more are added.

Theorem 1 formalizes this for one softmax attention step. The evidence score is bounded, the NN distractor scores are i.i.d. Gaussian, and the model is no better than base rate whenever the evidence gets less than a fixed share ρ\rho of attention. Under these assumptions, accuracy falls toward base rate unless the evidence margin grows at least like 2log⁡N\sqrt{2\log N}. Proposition 1 shows that a gate keeping KK candidates replaces NN with KK, at the cost of the gate's miss rate. The paper then proposes evidence bottlenecks, alias-resistant positions, an "anti-poison" loss and a verifier. Only a BM25 gate is implemented.

The evidence:

Decay with NN (Sec. 8): GPT-4o finds a fact among embedded decoys in passages of 100–1,000 sentences, with 10 trials at each length. Accuracy is 100% at 100 sentences and 80% at every longer length. Distractor type (Sec. 9): 300 records (about 20K tokens) on Claude Sonnet 4, GPT-4.1 and Gemma 3 12B. Same-template distractors cause the largest drop. Appendix tests (App. D–E): attention statistics on Gemma-3-4B and Llama-3-8B, a sweep over the number of look-alike decoys, and a small lost-in-the-middle rerun. Retrieval gate (App. F): BM25 top-KK on RULER qa_2 with GPT-4.1, using 100 questions at lengths from 4K to 512K tokens.

Critical strengths

The framing is clear and useful. The core idea is easy to follow. The evidence must beat the strongest distractor, so the number of look-alike distractors matters more than raw length. Fig. 1 shows this well, and the paper is candid about its limits. The evaluation advice is practical too. For example, it reports accuracy, evidence recall and accuracy given successful retrieval together, which separates "can't find the evidence" from "can't use it". This framing is the paper's main value.

The gate study (App. F) is the strongest experiment. It uses the same 100 questions at every length and picks KK on a separate split. It also scores the questions where the gate kept all the supporting evidence. On those questions, raw accuracy falls from 0.69 at 16K to 0.60 at 512K, while gated accuracy stays near 0.70. The study also shows the gate's tradeoff directly: with KK fixed at 16, the gate keeps all the needed evidence (recall) less often as context grows, falling from 0.94 to 0.79, and this offsets the gain until KK is raised. This is the paper's best evidence, but it sits in the appendix, and the net gain never reaches significance.

Critical weaknesses

The theory adds little and doesn't predict the paper's own results.

The main message is known. Veličković et al. (ICML 2025) already prove that softmax attention must spread out as the number of items grows. RULER, which the paper cites, shows that look-alike distractors hurt more than filler. The bound is loose. Under the paper's own assumptions, the evidence needs a margin that grows like log⁡N\log N, not log⁡N\sqrt{\log N}. That's because the total weight of all distractors grows faster than the strongest one. It doesn't fit the paper's data. The model treats every filler token as a threat, but the paper's data show filler barely matters. There are no errors with up to 400 filler sentences (App. D), and GPT-4.1 shows no length effect through 32K tokens (App. F). The paper's fix is an "effective distractor count", but it's never measured, so the theory can't say when failure starts. The experiments are too small or too confounded to show the mechanism.

GPT-4o decay (Sec. 8): accuracy is 10/10 at 100 sentences and 8/10 at each longer length. That difference isn't significant (Fisher's p≈0.47p \approx 0.47), and accuracy is flat from 300 to 1,000 sentences. Distractor type (Sec. 9): the one significant result rests on 4 Claude errors in 36 calls (exact paired p≥0.125p \ge 0.125). 35% of Claude calls and 96% of Gemma calls failed. The questions are about famous fiction and there's no closed-book baseline, so models may answer from memory. Decoy sweep (App. D): the decoys appear identical to the evidence except for the number, and accuracy matches random guessing, 1/(k+1)1/(k+1). This measures ambiguity, not attention interference. Mitigations: only a standard BM25 gate is tested, and its best net gain is +0.07 exact match at 512K (p=0.09p = 0.09). In short, the intuition is clear, but the theory restates known limits of softmax and the experiments don't show that this mechanism drives real failures. So I recommend rejection.

Questions for authors

Clarifications

In the decoy-count sweep (App. D), what tells the evidence apart from a high-similarity decoy? For the GPT-4o study, could you show a full example and say whether the number of decoys grows with NN? What is closed-book accuracy on the fiction questions, with no context or with distractors only? On an open model, does the score gap between the evidence and the strongest distractor shrink as NN or decoy similarity grows? Does it predict which questions fail? Your attention pipeline already supports this analysis, and it would connect Theorem 1 to behavior.

Suggestions

Test at least one proposed mitigation, for example the anti-poison loss on an open model. Move the gate study into the main text, in place of the 10-trial GPT-4o study.

2 comments

We thank the reviewer for the detailed assessment and for highlighting the usefulness of the framing, the retrieval-versus-use decomposition, and the paired gate study.

Theoretical contribution and prior work. We agree that the general size-dispersion limitation of softmax is not itself new, and we will add a direct comparison with Veličković et al. We will also avoid claiming that look-alike distractors being more harmful than filler is a new empirical observation. The distinct contribution we claim is narrower: Theorem 1 maps an explicit upper-tail distractor event, through a one-sided evidence-use condition, to a finite-sample task-accuracy upper bound, while separating raw context size, competitive distractor exposure, evidence retention, and evidence use.

We also agree that cumulative dilution yields a stronger log⁡N\log N requirement at fixed temperature. This does not make Theorem 1 incorrect; it identifies a complementary sufficient failure route based on one extreme competitor. We will state clearly that the log⁡N\sqrt{\log N} condition is necessary but non-sharp for the full fixed-temperature softmax denominator.

Relation to the empirical results. In Theorem 1, NN is the number of effective competitive distractors, not automatically the raw number of filler sentences. Therefore, no degradation under generic filler is not a contradiction of the conditional theorem. However, the present experiments do not estimate NeffN_{\mathrm{eff}}, so they cannot predict the onset of failure from raw length. We will make this limitation explicit.

We also agree that Sec. 8 currently overstates the link to the theorem: Eq. (7) is an upper bound, not a parametric equality for A(N)A(N). We will remove the nonlinear fit of a0+(1−a0)qNa_0+(1-a_0)q^N and its R2R^2, and retain the 10-trial GPT-4o result only as a descriptive pilot.

[INSERT THE VERIFIED HARD-NEGATIVE-SCHEDULE SENTENCE HERE.]

We will also include a complete prompt. The current attention pipeline uses post-softmax sentence aggregates, so it does not answer the requested question about the pre-softmax evidence-to-maximum score gap or whether that gap predicts failure.

Statistical and construct-validity limitations. The Claude "significant" label is based on a paired-bootstrap interval. The reported counts imply four discordant matched pairs in one direction, for which the exact two-sided McNemar test is p=0.125p=0.125; we will report this and remove the categorical significance claim. GPT-4.1 completed all calls and shows a smaller same-format point estimate in the same direction, but its interval includes zero. Claude and Gemma will be labeled endpoint-limited, with Gemma treated as descriptive only.

We did not run a no-context fiction baseline. The task also required a cited evidence-record ID, but this does not by itself establish that the answer content came from context; we will therefore narrow the claim to within-prompt behavior and clarify the record-ID scheme. We also agree that the identical-phrasing code-decoy sweep is ambiguous and will not use it as mechanism evidence.

The gate study remains the cleanest empirical diagnostic: it uses 700 paired comparisons, selects KK on a disjoint split, has no endpoint failures, and separates all-support recall from conditional answer accuracy. We will move it to the main text, while reporting the strongest net result (+0.070 EM, exact McNemar p=0.092p=0.092) as suggestive rather than conclusive. Because it is multi-hop, we will describe it as a qualitative test of the candidate-reduction/recall tradeoff, not as validation of the single-item theorem.

Finally, the anti-poison objective, verifier, and positional mechanisms are theory-derived design implications, not evaluated engineering contributions. We will revise the wording accordingly.

I thank the authors for their detailed response, I decide to maintain my score.

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

This paper investigates long-context degradation through an extreme-value model of attention interference. Under a single-bottleneck abstraction with independent Gaussian distractor logits and a threshold-based evidence-use assumption, it derives an accuracy upper bound and an Ω(log⁡N)\Omega(\sqrt{\log N}) necessary evidence-margin condition. It also discusses retrieval gating and architectural mitigation strategies. Experiments examine distractor count and confusability, attention distributions, positional effects, and retrieval gating on HotpotQA-based contexts extending to 512K tokens.

Critical strengths
  1. An interpretable formalization of distractor interference. The paper distinguishes raw context length from effective distractor exposure and separates evidence availability from evidence use. Theorem 1 provides a transparent finite-sample bound within its stated assumptions, helping explain why increasing context capacity need not improve usable memory.

  2. Complementary controlled experiments. The fixed-length confusability study and fixed-count decoy sweep help separate distractor composition from context size. The appendices add open-weight attention diagnostics and real-question evaluations. Separately reporting evidence recall and conditional answer accuracy is useful, and the paper transparently acknowledges endpoint failures, distributional mismatches, and statistical uncertainty.

Critical weaknesses
  1. Proposition 1 is incorrect as stated. The proposition permits a gate retaining “at most KK distractors” but substitutes KK into the decreasing term Φ(⋅)K\Phi(\cdot)^K. For an actual retained count m≤Km\le K, Φ(⋅)m≥Φ(⋅)K\Phi(\cdot)^m\ge\Phi(\cdot)^K, so this substitution does not preserve the upper bound. Furthermore, even with exactly KK distractors, a higher accuracy ceiling does not establish improved gated accuracy. The claimed benefit-versus-recall-penalty interpretation needs stronger justification.

  2. The central margin condition is valid but non-sharp. Under the same fixed-temperature Gaussian model, the full distractor log-sum-exp grows as τlog⁡N+OP(1)\tau\log N+O_P(1). Maintaining a nonvanishing probability of sufficient evidence attention therefore requires a stronger, log-NN-scale evidence score. Section 5 acknowledges cumulative softmax dilution but does not analyze this competing asymptotic or establish when the maximum-based characterization is most informative.

Questions for authors
  1. How should Proposition 1 be corrected for a gate retaining at most KK distractors? Can the authors establish an actual performance comparison rather than only a less restrictive accuracy upper bound?

  2. How does the log⁡N\sqrt{\log N} condition relate to the stronger log-NN requirement from cumulative softmax dilution? In which context-length or temperature regime is the maximum-based analysis most informative?

  3. Could the authors measure raw attention logits, evidence margins, maximum distractor scores, and the full softmax denominator on trials containing failures? What interventions would distinguish the proposed mechanism from alternative explanations?

  4. Appendix D describes high-similarity decoys with identical phrasing that differ only in the number. What observable cue uniquely identifies the intended answer? Could complete prompts clarify whether these conditions preserve an unambiguous task?

  5. Could the retrieval-decay and confusability studies be repeated with larger samples and reliable endpoints? Does the same-format effect persist on fully matched instances without substantial missing responses?

2 comments

We thank the reviewer for the careful analysis and for recognizing the value of the finite-sample formalization and the separation between evidence availability and evidence use. The issues raised require two precise revisions, but they do not invalidate Theorem 1.

Proposition 1. We agree that "at most KK" does not justify substituting KK into a term that decreases with the retained distractor count. This is a local error in Proposition 1 and does not affect Theorem 1, Corollary 1, or their proofs.

Let HH denote retention of the decisive evidence, let MM be the realized number of retained distractors, let p=Pr⁡(Hc)p=\Pr(H^c), and define

\qquad B_m=a_0+(1-a_0)q^m.$$ Assuming Theorem 1's conditions hold conditional on $H$ and $M=m$, the valid statement is $$A_{\mathrm{gate}}\le p+(1-p)\\,\mathbb{E}[B_M\mid H].$$ With only $M\le K$, no $B_K$ upper bound follows, because $q^M\ge q^K$. If instead $M=m$ deterministically on a retrieval hit and $p\le\eta$, then $$A_{\mathrm{gate}}\le \eta+(1-\eta)B_m.$$ We will replace Proposition 1 accordingly and distinguish the total top-$K$ candidate budget from the actual number of retained distractors. We also agree that changing an upper bound does not prove improved realized accuracy; no theorem-level performance comparison follows from the current assumptions. The paired gate experiment provides an empirical comparison, but its strongest net effect (+0.070 EM, exact McNemar $p=0.092$) is suggestive rather than conclusive. **Maximum interference versus cumulative dilution.** The $\sqrt{\log N}$ condition is valid for the single-extreme-distractor event used in Theorem 1, but it is not the sharp fixed-temperature condition for the full denominator. Under the same i.i.d. Gaussian model and fixed $\tau$, $$\frac{1}{N}\sum_{i=1}^N \exp(Z_i/\tau) \longrightarrow \exp\left(\frac{\sigma^2}{2\tau^2}\right)$$ almost surely. Consequently, maintaining a nonvanishing probability of $\alpha_{\star,N}\ge\rho$ requires, up to lower-order terms, $$\gamma_N\ge \tau\log N+\frac{\sigma^2}{2\tau} +\tau\log\left(\frac{\rho}{1-\rho}\right)+o(1).$$ Thus cumulative dilution gives the stronger $\log N$ scale at fixed temperature. This complements rather than invalidates Theorem 1: the theorem remains a finite-sample certificate that one extreme distractor can be sufficient to enter the low-attention regime. We will remove any suggestion that the maximum route is asymptotically sharp or dominant at fixed $\tau$. **Diagnostics and experiments.** Appendix D analyzes transformed, sentence-aggregated post-softmax attention, not the raw logits in Eq. (5). We will correct the statement that it "directly" tests the Gaussian-logit assumption. A direct mechanism test would record the evidence logit, maximum distractor logit, and full log-sum-exp while independently varying count and confusability; the submitted paper does not contain that test. We also agree that the identical-phrasing code condition lacks a unique observable cue. We will relabel it as an ambiguity stress test and remove it from the evidence for attention interference. The small and endpoint-limited experiments will be described as exploratory diagnostics, not validation of the theorem.

Thank you for the thoughtful response. I appreciate the clarification of Proposition 1 and the distinction between the finite-sample bound in Theorem 1 and cumulative softmax dilution. These clarifications address my theoretical concerns, but the response does not provide direct evidence that single-distractor interference explains the observed failures, or show that the same-format effect persists on matched examples with usable responses in both conditions. Since these empirical concerns remain unresolved, I will keep my original rating 2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.