acceptodds
Under review as a conference paper at ICLR 2027

Before Attention: Functional Visual Token Roles at the Vision–Language Interface

Abstract

Large vision–language models connect visual encoders to pretrained language models through a vision–language interface, yet visual-token function is typically characterized only after decoder processing begins. We ask whether functionally predictive differences among visual tokens are already accessible at this interface. Across projector- and merger-based LVLMs, we observe a recurring post-interface organization: activation mass becomes less concentrated within individual tokens, while a restricted set of hidden coordinates is repeatedly reused across tokens. To test whether this structure is functionally meaningful, we develop PRISM, a lightweight role-recovery procedure that uses paired interface-relative concentration responses to construct interface-response anchors, then recovers image-level contributive and residual roles from dominant coordinate identities and relative ranks. Controlled interventions reveal a consistent functional asymmetry: removing contributive tokens is substantially more disruptive than removing residual tokens, even after matching removal count and subsequent decoder attention. The recovered distinction also extends beyond the initial structural contrast, indicating that PRISM captures structure not available from scalar concentration change alone. Because the partition is computed before query-conditioned decoder processing, the same image-level assignment can be reused across questions. These results show that functional differences among visual tokens are already reflected in representations entering the language model, establishing the vision–language interface as an earlier stage for multimodal analysis.

Primary area:
interpretability and explainable AI
Keywords:
Large Vision-Language Models, Vision-Language Interface, Multimodal Representation Analysis, Functional Token Roles, Model Interpretability

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_YayUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether visual-token representations already predict downstream functional sensitivity before language-decoder processing. Across four 7B vision-language models, it observes decreased fixed-top-coordinate concentration within tokens but strong reuse of dominant coordinate identities across tokens (Section 2, Figure 3). Default binary PRISM uses image-wise tails of the pre/post-interface concentration difference as calibration anchors, then greedily learns a compact disjunction of coordinate-rank predicates; downstream labels and decoder attention are not used to learn this rule (Section 3, Equations 2-8). Matched tokens are called contributive and the unmatched complement residual, based on subsequent removal experiments. Validation includes role removal across four backbones, count-matched interventions within attention quartiles on two backbones, scalar and dense-probe controls, and image-mask reuse across questions (Tables 1-4). I favor weak acceptance for the useful empirical finding, while treating the attention-independence and generalization claims more cautiously.

Critical strengths
  1. The central distinction has substantial intervention evidence rather than only representational visualization. Complete contributive removal is more damaging than residual removal on all three primary metrics across all four backbones (Table 1A); the retention curves provide additional context rather than equating residual with useless (Appendix C, Table 8). The count-matched, attention-stratified comparison further shows large gaps despite lower mean attention for the selected contributive tokens (Appendix G, Table 12). PRISM also substantially exceeds direct concentration-difference ranking on POPE and MME at the original top-decile budget, with three-seed variation reported (Table 2; Appendix F.2, Table 10). These controls make the finding consequential: a simple pre-decoder representation contains predictive structure that is not captured adequately by the tested scalar ordering.

  2. The image-level prior has a concrete, reusable application. Appendix I fixes both the role map and retained mask per image across 5,000 questions on 883 images. The gains over the attention-based comparator persist across the reported question categories (Table 14), supporting useful reuse rather than merely asserting query independence from the method's definition. Restricting the candidate pool while keeping the selection budget and ranking criterion fixed improves aggregate retained performance on all four backbones (Table 4B; Appendix K.1, Table 16). The complementary diversity-selector experiment is also informative, although its improvements are smaller and VILA's two role-biased variants tie (Appendix K.2, Table 17).

Critical weaknesses
  1. The attention control supports a narrower conclusion than exact attention matching. Appendix G matches counts within four attention-mass quartiles using decoder layers 10-30. The matched contributive tokens have 47.2% and 55.9% lower mean attention than residual tokens (Table 1B). This deliberately reversed ordering convincingly challenges a simple explanation based on larger average attention; it tests aggregate attention rather than establishing independence from all attention structure. The matched intervention is reported for LLaVA-1.5 and VILA (Table 12). Unlike the three-seed standard deviations for fixed-budget selection in Table 10, Table 12 presents point estimates without repeated-sampling uncertainty or image-level confidence intervals. Thus, Section 5.2's conclusion should be scoped to the tested attention statistic and backbones. This is fixable through tighter matching, uncertainty estimates, and more qualified wording.

  2. Calibration independence and cross-interface reproducibility need narrower clarification. Appendix H specifies MME as the default calibration source for each backbone, and Appendix A.1 gives 1,187 images for primary LLaVA-1.5 calibration. Appendix H also reports positive removal gaps after TextVQA and POPE recalibration on two backbones (Table 13), providing genuine cross-source evidence. However, an explicit image-level calibration/evaluation split or overlap accounting is not given; Appendix F.3's anchor-region analysis is explicitly on calibration images. The remaining issue is separating calibration-image behavior from unseen-image behavior, not absence of cross-dataset controls. Also, Section 2 indexes paired pre/post representations by the same token index, but the merger-based model's construction of corresponding pre-interface representations is not specified in Appendices A/B. These are fixable reproducibility clarifications, potentially requiring image-disjoint results.

Questions for authors
  1. [Clarification] Appendix A.1 specifies uniform retention sampling within roles. For Appendix G/Table 12, how are the count-matched subsets selected within each quartile, and what are their per-image counts and attention distributions? Do the gaps persist under tighter matching, with image-level uncertainty estimates?

  2. [Clarification] For Appendix A.1 and Appendix H, which images overlap calibration and evaluation for each backbone and benchmark? Can you provide image-disjoint intervention results and clarify whether the three seeds in Table 10 vary only token sampling or also calibration?

  3. [Clarification] For Equations 1-3 and the Qwen2.5-VL experiments, how is each merged post-interface token paired with its pre-interface representation? Given Appendix H's default MME source, do the other backbones use the same calibration image set as LLaVA-1.5?

Reviewer_aF8xOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
3
Read review
Paper summary

This paper asks whether functionally predictive differences among visual tokens are already accessible at the vision–language interface, before decoder processing begins. The authors observe that post-interface visual tokens exhibit lower within-token concentration while strongly reusing a small set of dominant coordinates across tokens. Building on this, they propose PRISM, which constructs anchor sets from the paired pre/post concentration contrast Δi\Delta_i and greedily learns a disjunctive rule over coordinate–rank predicates to partition tokens into contributive and residual roles. Controlled interventions on four LVLMs show that removing contributive tokens is consistently more damaging than removing residual tokens, including under count- and attention-matched settings, and the partition can be reused across questions and as a prior for fixed-budget token selection.

Critical strengths
  • Asking whether token function is already determined before query-conditioned decoder computation is a meaningful for both multimodal interpretability and token reduction.
  • The writing is cautious about its own claims. The authors consistently avoid treating Δi\Delta_i as an importance score and frame the contributive–residual distinction as relative rather than useful versus useless.
  • The method is lightweight and label-free, is calibrated once per backbone, and can be combined with existing token selectors rather than replacing them.
Critical weaknesses

I read the paper with interest and reviewed it carefully and in detail. Below, I present my thoughts on its weaknesses.

  • [W1] The role of the spectral analysis in Sec. 2.1 is unclear. Singular spectra of trained linear layers typically decay, so without a reference (e.g., the spectrum of a randomly initialized or an ordinary LLM linear layer), it is hard to tell whether the observed anisotropy is notable. The two-stage pattern also holds for the three LLaVA-style models but is reversed in Qwen2.5-VL, which the paper does not discuss. In addition, the analysis concerns singular directions rather than individual coordinates, so it does not directly motivate Sec. 2.2, and it is not used in PRISM. Since it is nonetheless listed as a finding in the conclusion, I suggest either adding a simple reference spectrum or presenting it as supplementary motivation.

  • [W2] The concentration decrease in Figure 3A may partly reflect the change in dimensionality. Since the lower bound of C(x)C(\mathbf{x}) is Kc/mK_c/m, values computed in 1024 and 4096 dimensions are not directly comparable. Normalized by this lower bound, the pre- and post-interface values are nearly identical for LLaVA-1.5 (≈5.5×\approx 5.5\times vs. ≈5.7×\approx 5.7\times). Since this decrease is one of the two parts of the Key observation, a brief check with a dimension-normalized measure would strengthen it.

  • [W3] Δi\Delta_i combines terms on different scales, and a pre-only baseline is missing. Δi=C(hi)−C(vi)\Delta_i = C(\mathbf{h}_i) - C(\mathbf{v}_i) subtracts concentrations computed in 1024 and 4096 dimensions, whose scales differ (the uniform lower bound Kc/mK_c/m differs by a factor of four). Within-image ranking removes a common offset but not this scale difference, so the ranking of Δi\Delta_i may be driven mainly by C(vi)C(\mathbf{v}_i), that is, by the encoder output rather than the interface. Table 19 compares Δi\Delta_i with post-interface statistics but not with C(vi)C(\mathbf{v}_i) alone. Adding this baseline would clarify whether the benefit of the paired contrast comes from the interface or from the encoder.

  • [W4] The learned rule is interpretable by design, yet no interpretation is provided. The top-1 coordinate is shared by up to 96% of tokens (Figure 3B), possibly a massive-activation dimension [1], so the discriminative power of PRISM must come from lower-ranked coordinates. The learned rule involves only 26 coordinates (Appendix J), yet the paper does not report which coordinates and rank conditions were selected or where in the image they fire. Listing the predicates and visualizing a few of them would help explain what the contributive role corresponds to.

  • [W5] The middle-region fraction does not by itself demonstrate generalization. Appendix F.3 presents the fact that 58.34% (LLaVA-1.5) and 56.28% (VILA) of PRISM-selected tokens come from the unassigned middle 60% of the Δi\Delta_i distribution as evidence of generalization. Since this region contains 60% of tokens by construction, random selection would give a similar fraction. These numbers do show that PRISM selects a set different from the upper Δi\Delta_i tail, but enrichment relative to the base rate (e.g., 23.67% from A+\mathcal{A}^+ vs. a base rate of 10%) would be a more informative measure, and a softer claim seems appropriate.

  • [W6] The comparison in the cross-question reuse experiment is limited. Table 4A shows that a single PRISM mask reused across questions outperforms FastV [2]. However, any query-agnostic selector can be reused in the same way, and FastV is a weak baseline at this budget. In the paper’s own Tables 16 and 17, FastV retains 88.3 on LLaVA-1.5 while the query-agnostic DivPrune [3] retains 93.3. Since DivPrune is already implemented in the paper, including it under the same reuse protocol would show whether the gain reflects the pre-decoder prior itself. It would also help to clarify how the fixed mask is selected from the contributive pool, since using FastV ranking within the pool would make the mask question-dependent.

  • [W7] The base-LLM overlap analysis needs a chance baseline. Table 3B reports 0/26 overlap between the contributive-rule coordinates and the top-26 high-activation axes of Vicuna. Zero overlap argues against strong inheritance, but it is also the expected outcome for unrelated sets (expected overlap 262/4096≈0.1726^2/4096 \approx 0.17), so reporting this baseline would help readers calibrate the result. By the same baseline, the 4 to 5 out of 26 overlap after decoder layer 1 (Table 15) is well above chance and may deserve more discussion. Finally, since the comparison uses only text-token axes, repeating it with the high-activation axes of the visual tokens would more directly test whether the rule coordinates reflect outlier dimensions from the encoder or projector.

[1] Mingjie Sun et al., Massive Activations in Large Language Models, COLM, 2024
[2] Liang Chen et al., An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models, ECCV, 2024
[3] Saeed Ranjbar Alvar et al., DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models, CVPR, 2025

Questions for authors
  • [Q1] Could the authors specify the interface architecture of each backbone (number of layers, activation, input and output dimensions) and the exact checkpoints used? In particular, VILA releases differ in projector design, and the rank=1024 annotation in Figure 2 presupposes a 2-layer MLP with 1024-dimensional input.

  • [Q2] The Qwen2.5-VL merger combines each 2×2 patch group into a single token, so pre- and post-interface tokens have no one-to-one correspondence. How is vi\mathbf{v}_i defined in this case (e.g., the concatenated 5120-dimensional vector or the mean of the four patches)?

1 comment

Thank you for your detailed responses and additional experiments.

I appreciate the substantial effort you have made to address my concerns. After reviewing your responses, I believe that my major concerns have been adequately addressed, and I will therefore raise my rating to 3.

Reviewer_TWDrOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether functional differences among visual tokens can already be identified at the vision–language interface, before query-conditioned decoder processing. The authors propose PRISM, which constructs pretext anchors from paired pre/post-interface concentration changes and then learns compact coordinate-rank rules to partition visual tokens into contributive and residual roles. Experiments across four LVLMs show that removing contributive tokens causes substantially larger performance degradation than removing residual tokens, and the recovered roles can also serve as reusable priors for fixed-budget visual-token selection.

Critical strengths
  • The paper goes beyond correlational analysis and validates the recovered roles through direct interventions across multiple LVLM architectures.
  • PRISM provides a lightweight and reusable pre-decoder token prior. The role assignment is computed before decoder traversal and does not require query-conditioned attention, gradients, or additional language-model forward passes. The same image-level role map can be reused across different questions, and using contributive tokens as the candidate pool improves fixed-budget token selection over standard FastV across multiple backbones.
Critical weaknesses
  1. The novelty over closely related pre-decoder visual-token role analyses is not sufficiently established. The paper itself identifies [1] as closely related work that already categorizes projected visual tokens into different functional groups. PRISM uses a distinct coordinate-rank formulation, but the current experiments do not directly compare its contributive/residual partition with existing token-role assignments. Therefore, it remains unclear whether PRISM discovers a genuinely different functional organization or largely provides another proxy for previously identified informative and redundant visual-token groups.
  2. The paired concentration contrast used to construct PRISM's initial anchors requires stronger justification. PRISM defines Δi=C(hi)−C(vi)\Delta_i=C(h_i)-C(v_i) using the same fixed top-KcK_c concentration statistic, although the pre- and post-interface representations viv_i and hih_i can lie in spaces with different dimensionalities. Since fixed-KK concentration can depend on representation dimensionality, it is unclear to what extent the observed decrease and resulting anchor ordering reflect an interface-specific transformation rather than properties of the two representation spaces. A dimension-normalized or matched-dimensionality control would strengthen the central premise.

References:

[1] Yingqi Fan, Junlong Tong, Anhao Zhao, and Xiaoyu Shen. What do visual tokens really encode? uncovering sparsity and redundancy in multimodal large language models. CVPR 2026.

Questions for authors
  1. Can the authors report results using a dimension-normalized concentration statistic, such as a fixed top proportion rather than a fixed KcK_c, to verify that the recovered functional asymmetry remains unchanged?
  2. Can PRISM be calibrated on an image set completely disjoint from all downstream evaluation benchmarks and still obtain comparable intervention and token-selection results?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.