acceptodds
Under review as a conference paper at ICLR 2027

Attention Dilution: Two Statistical Failure Modes of Visual Token Inflation

Abstract

Multimodal language models encode each image as hundreds to thousands of visual tokens. As this count grows, instruction following and fine-grained retrieval degrade, while prune and merge heuristics succeed or fail in ways that appear model dependent. A diffuse attention map need not imply a wrong answer, and no method explains when token count alone causes which failure. We develop a statistical theory of these phenomena. Softmax attention on a two population visual context fails in two ways. Evidence mass on the tokens that carry the answer decays algebraically past a budget set by the margin, and a typical query retains systematically less mass than the average one. Top 1 retrieval of a visual needle collapses at a much larger, exponentially separated scale. When the mass budget lies below the retrieval budget, a diluted yet correct window opens in which attention is already diffuse while the needle remains the argmax. The same split classifies queries by which budget binds and maps when score pruning or covering merges can work for a single frozen attention operation. Monte Carlo studies, a trained attention layer, a CLIP proxy, frozen Qwen3-VL decoders, and a LLaVA-1.5 encoder check instantiate the two laws. Generation accuracy on a referring scene index probe crosses one half in the predicted window once the query is read at the first generated token and the margin is locked from logits alone.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Multimodal large language models, Visual tokens, Attention dilution

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_S2G2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper models one attention query as a few evidence tokens with a margin over many Gaussian noise tokens. It shows that attention mass on the evidence decays as 1/N, while top-1 retrieval fails only at an exponentially larger token count, leaving a window where attention is diffuse but the answer token is still the argmax. The predictions are checked with Monte Carlo, a CLIP proxy, and frozen Qwen3-VL and LLaVA-1.5 models.

Critical strengths
  1. The separation between attention mass and argmax retrieval gives a simple explanation for why a diffuse attention map can coexist with a correct answer. It also connects several known observations about token pruning, such as random dropping being competitive.

  2. The margin is estimated from logits on a held-out split and fixed before evaluation, so the comparison with real models is not a curve fit to the outcome. The Monte Carlo results match the stated laws closely.

Critical weaknesses
  1. The model is oversimplified. It describes one frozen attention operation with i.i.d. Gaussian logits, while real multimodal models have many layers and heads, and the paper does not show how this single-softmax statistic determines the model output.

  2. The experimental setup does not correspond to the theory. The theory concerns one needle token, but the probe is image-level retrieval among several images, and Appendix F states both that the needle image is placed before the distractors and that it is shuffled, with no analysis of needle position.

Questions for authors
  1. Where is the needle image placed in the Qwen3-VL prompts? Please report results with the needle at different positions.A single-image probe with a small target region, such as OCR or referring tasks, would match the single-needle setting of the theory more directly.

  2. The attention is read from the last layer and averaged over heads. Do the two laws hold for individual layers and heads, and how do they combine across layers?

I would consider raising my score if the authors clarify the needle placement protocol, report results across needle positions, and provide per-layer or per-head evidence that the two laws hold in real models.

Reviewer_BhPVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the impact of visual token inflation on multimodal language models from a statistical perspective. It identifies two distinct failure modes of softmax attention: attention mass dilution and retrieval collapse, and derives theoretical bounds characterizing their different scaling behaviors. The analysis further explains why diffuse attention can coexist with correct retrieval and examines the implications for visual token pruning and merging strategies. The authors validate their theoretical findings through Monte Carlo simulations, CLIP-based experiments, and evaluations on frozen Qwen3-VL and LLaVA models.

Critical strengths

(1) The paper distinguishes two different failure modes of visual token inflation: attention mass dilution and retrieval collapse. Analyzing them through different statistical properties of softmax attention provides a useful perspective on why diffuse attention does not necessarily lead to incorrect retrieval.

(2) The theoretical analysis goes beyond qualitative observations by deriving explicit token budgets for both failure modes. The discussion of typical versus average attention mass, extreme-value behavior, and spatial correlations adds depth to the analysis and helps explain how different factors affect the two failure modes.

(3) The paper includes experiments ranging from controlled Monte Carlo simulations and CLIP features to frozen Qwen3-VL and LLaVA models.

Critical weaknesses

(1) Theorem 2 is stated under Assumption 1, which allows different variances for signal and noise logits (σS≠σV\sigma_S \neq \sigma_V). However, the proof in Appendix B.4 explicitly assumes σS=σV\sigma_S = \sigma_V, and Appendix C.3 provides a separate expression for the heteroscedastic case. It is therefore unclear whether Theorem 2 applies only to the equal-variance setting or also covers the more general case. In addition, the asymptotic analysis considers Nv→∞N_v \to \infty with fixed margin and variance, under which both the true retrieval probability and its approximation approach zero. Convergence in absolute error alone may therefore be insufficient to characterize the accuracy of the predicted retrieval threshold.

(2) Theorem 3 shows that exact merging using log-sum-exp logits and softmax-weighted values preserves the original attention output. However, its connection to the attention mass restoration claim in Corollary 3 is unclear. Since exact merging also preserves the total softmax mass of each merged cluster, merging noise tokens in this way should leave the evidence retention rate (η=A/(A+B)\eta=A/(A+B)) unchanged rather than increase it toward a target threshold. Could the authors clarify how attention mass restoration is achieved under this formulation?

(3) The theoretical analysis considers a single frozen attention operation, but the paper extends its conclusions to more complex MLLM behaviors, including instruction following, captioning, and visual retrieval. The connection between these settings is not entirely clear. For example, the theoretical definition of evidence retention does not explicitly account for other text tokens, whereas the Qwen3-VL experiments measure retention relative to visual attention only. Furthermore, real MLLMs involve multiple attention layers, residual connections, and downstream transformations, all of which may affect how attention dilution translates into generation performance. Further discussion of these differences would help clarify the scope of the theoretical conclusions.

(4) In the Qwen3-VL experiments, visual token inflation is mainly achieved by adding more distractor images. However, this also increases the number of candidate images, semantic complexity, and retrieval difficulty, making it difficult to determine how much of the performance degradation is caused by token count alone. Although the synthetic correlation experiments and LLaVA collage study provide some supporting evidence, more controlled experiments would be useful. For example, the authors could vary the visual token count while keeping semantic content fixed, or vary distractor complexity under a fixed token budget. Moreover, the real-model generation experiments mainly focus on the Referring Scene Index Probe, with limited direct evaluation of other tasks frequently discussed in the paper, such as captioning, OCR, and instruction following. While the proposed framework offers a plausible explanation for these tasks, the current evidence seems insufficient to establish whether the two failure modes generalize across such diverse settings.

Questions for authors

Please refer to the Critical Weaknesses section.

Reviewer_m9w3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates attention degradation caused by increasing visual token counts in multimodal language models and develops a statistical framework based on two populations of evidence and noise tokens. The framework distinguishes two failure modes: algebraic dilution of evidence attention mass and retrieval collapse at a token budget exponential in the squared signal-to-noise ratio. This distinction identifies conditions for a “diluted yet correct” regime, where attention is already diffuse while the target remains the highest-scoring token. The paper further analyzes the asymmetric effects of spatial redundancy on these two modes and characterizes when score pruning or covering-based merging can work within a single frozen attention operation. Experiments with frozen Qwen3-VL and LLaVA-1.5 models evaluate the predicted thresholds and their relationship to generation accuracy.

Critical strengths
  • The paper distinguishes evidence mass dilution from top-1 retrieval collapse within a unified statistical framework. This distinction provides a perspective on why diffuse attention can coexist with successful retrieval and clarifies the different quantities involved in attention degradation.

  • The proposed measurement protocol estimates the signal-to-noise ratio from logits on a separate estimation split and fixes it before evaluation. Separating parameter estimation from the measured retrieval and generation outcomes supports a more direct assessment of the theory’s predictive value.

  • The evaluation includes Monte Carlo simulations, a CLIP proxy, and multiple frozen multimodal models. These settings allow the theoretical predictions to be examined at different levels.

Critical weaknesses
  • There are substantive mathematical errors in the retention-restoration claim and the heavy-tail extension. For a fixed attention query with unchanged evidence logits, the exact merge in Theorem 3 preserves total noise mass and therefore leaves evidence retention unchanged, contradicting Corollary 3 and Appendix D.1. Additionally, p. 5, lines 237–240 incorrectly assert that finite variance guarantees a finite exponential moment. A standardized Student-t distribution with five degrees of freedom is a counterexample. Even when the exponential moment is finite, changing the noise distribution can change its value and hence the mass budget.

  • The experiments support the proposed laws in controlled settings, but do not sufficiently establish their predictive relevance to downstream generation or compression-strategy rankings. Although the paper explicitly distinguishes retrieval hit probability from generation accuracy, the quantitative connection between them remains insufficiently validated. The reported compression results also do not establish how the theoretical regimes predict the relative generation performance of different strategies. More concretely, Table 8 reports generation half-crossings of 363/950/1223, all exceeding the corresponding predicted retrieval budgets of 27/110/277. Since these budgets define the upper bounds of the windows in Corollary 1, the reported crossings are inconsistent with the abstract’s claim that they occur within the predicted window. The authors should reconcile this discrepancy and clarify the scope of the predictions concerning generation.

  • The main novelty lies in jointly characterizing evidence mass dilution and top-1 retrieval collapse within a two-population model, and in identifying conditions for the diluted-yet-correct regime. The individual analyses largely build on established concentration and extreme-value results, while the cluster-size correction has a direct precedent in ToMe’s proportional attention. The contribution therefore appears primarily conceptual and integrative, with more limited novelty in the underlying technical ingredients.

  • The connection between the proposed statistical model and specific task categories, such as instruction following, captioning, and OCR, requires further justification, particularly regarding the assumption that signal and noise parameters remain fixed as the visual token count increases.

Questions for authors
  • How does Table 8 support the abstract's claim that generation half-crossings lie within the predicted window?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.