acceptodds
Withdrawn from ICLR 2027

What Do VLM Language States Remember? Low-Rank Recovery of Instance-Level Visual Evidence

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Does a multimodal language model retain, in its contextualized language states, evidence pointing to the specific visual instance a referring expression denotes: the dog on the left, not merely dog? We probe this with a capacity-limited linear decoder, asking whether language states can linearly recover the corresponding spatially localized visual evidence from frozen vision representations. Across Qwen3-VL, LLaVA-OneVision, and Llama-3.2-Vision on RefCOCO, a low-rank linear mapping recovers the instance-level correspondence, reaching 88.4%, 88.6%, and 71.7% Pointing, with most gains saturating near rank 32; rank sweeps and cross-modal geometry localize this structure to a few shared dimensions, while native-attention, CLIP, and statistical-projection baselines perform substantially worse. CKA/CCA, semantic mismatch, anchored concept swapping, and unseen-category transfer further show that recovery follows cross-modal structure already present in frozen representations rather than arbitrary bbox associations. Spatially matched suppression and retention show that the recovered evidence has a location-specific causal effect on outputs in Qwen and LLaVA, distinguishing decodable from functionally used information. Token-level analyses show instance specificity emerges only as the expression is contextualized (bike sharpening into bike on the far left), and across architectures the location and strength of recoverability depend on how visual representations are injected into the language stream. Thus, the models retain, in their language states, a compact low-rank structure of linearly recoverable instance-level visual evidence, beyond what native attention reveals.

Primary area:
interpretability and explainable AI
Keywords:
Vision-Language Models, Multimodal Representation Learning, Cross-Modal Semantic Correspondence, Low-Rank Linear Probing, Visual Grounding

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

The paper explores if there is more information than just object categories, such as instance-level spatial relationship information in VLMs. The authors use three model families (Qwen3VL, LLaVA-OneVision, LLaMa-3.2-vision) to train low-rank linear mapping that maps the representations from the vision encoder and LLM backbones. They evaluate the performance on several benchmarks to show the instance spatial level information exists. They also do analysis by using representation similarity, semantic mismatching, concept swapping to show the mapping recovers cross-modal structure, which shows the language representation can retrieve matching visual patches.

Critical strengths

1.The question is clearly defined. The paper use 4 levels to explain the results to distinguish “decodable information” from “information actually used by the model”. 2. The are extensive controllable experiments to rule out the trivial explanations, including geometry analysis, semantic mismatching, concept swapping and so on. 3. Mapping-guided reweighting improve the several downstream tasks performance, including v*, ADE20K Count, POPE random.

Critical weaknesses
  1. The authors argues about the language states “remember a specific instance”. However, the localization jointly use the language representations and image patches instead of using the using language representation itself. It’s more like language representation include the instance-level conditioned visual query information.
  2. The baseline comparison is not clear. The authors don’t provide the experimental setup for the baselines like CLIP, groundingDINO. Since the mapping in paper is supervised trained, it will be unfair to compare the CLIP directly.
  3. There is no explanation for the high pointing accuracy but low IoU scores.
  4. The pipeline is not clear enough. If the image and text pair is the input, then the text token will also have the information based on the self-attention from the image tokens. It will be better to compare language representation itself only (no image input) as the query.
  5. It will be better to have an ablation study for the shuffled image patches with text representations.
  6. Missing related work.
    1. Li, Jiaang, et al. "Seeing or Knowing? Visual Context Sensitivity in Multimodal Large Language Models."
    2. Huh, Minyoung, et al. "The platonic representation hypothesis."
    3. Li, Jiaang, et al. "Do vision and language models share concepts? a vector space alignment study."
    4. Esmaeilkhani, Parsa, and Longin Jan Latecki. "Preserving Localized Patch Semantics in VLMs."
Questions for authors

See weaknesses.

Reviewer_7AhHOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether contextualized language states in vision-language models retain spatially specific evidence for the instances denoted by referring expressions. To examine this, the paper introduces a Local Grounding Bridge that projects target-phrase states and visual patch representations into a shared low-rank space while keeping the VLM frozen. Experiments on RefCOCO across Qwen3-VL, LLaVA-OneVision, and Llama-3.2-Vision demonstrate instance-level localization, with most gains saturating near rank 32. Geometric analyses, semantic perturbations, and unseen-category transfer provide evidence of shared cross-modal structure, while token-level analyses trace how instance specificity develops as contextual qualifiers accumulate. Spatially matched suppression and retention experiments in Qwen and LLaVA further show that the identified visual regions affect model outputs.

Critical strengths
  • The paper investigates a clearly defined question: whether contextualized language states support recovery of specific visual instances beyond category-level semantics. Evaluating a low-rank probe across three frozen VLM architectures provides a comparative basis for studying this correspondence.

  • The analysis includes rank sweeps, representation geometry, semantic mismatch, concept swapping, and unseen-category transfer. These complementary controls help assess the role of semantic pairing, supervision, and generalization in the recovered correspondence.

  • The suppression and retention experiments use spatially shuffled score maps that preserve the intervention coefficients. This provides a useful control for assessing whether the locations identified by the Bridge have behavioral relevance.

Critical weaknesses
  • The main grounding comparison is incomplete. Table 1 compares the Bridge against CLIP (2021), OWL‑ViT (2022), and Grounding DINO (2024), but does not include more recent methods from 2025 and 2026.

  • Figure 1 Panel (b) appears to combine unrelated examples, showing a tennis-player image alongside the query "the second bus from the right". More importantly, the figure does not explicitly show how visual information enters the language model before the target-span representations are extracted. The inference procedure also needs clarification: the Bridge uses language states from T19 in Qwen and T24 in LLaVA, while the downstream interventions are applied at earlier layers, L11 and L9, respectively. It is unclear how the evidence from these later layers is obtained before intervening at earlier layers, for example through an initial forward pass or cached activations. Clarifying this execution order in Figure 1 and the method description would make the framework easier to understand and reproduce.

  • Qwen T19–V6 achieves 63.03% held-out AP in the layer analysis (Appendix A.3), but only 61.10% at rank 64 in Table 7. The reason for this discrepancy is not explained. The model-selection protocol is also unclear: Appendix A.1 describes configuration selection using internal validation data, while Appendix A.3 cites external RefCOCO grounding performance when selecting the final Llama configuration. The authors should reconcile the reported AP values and clarify whether the external RefCOCO evaluation used for configuration selection was independent of the final test set.

Questions for authors

See weaknesses.

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

This paper investigates whether VLM language states preserve information about specific spatial information from vision embeddings. The authors use a low-rank linear Bridge to recover spatial information from frozen vision and language representations. They evaluate Qwen3-VL, LLaVA-OneVision, and Llama-3.2-Vision, primarily on RefCOCO. This paper finds that instance-level visual information can be recovered through a low-rank mapping, with performance largely saturating around rank 32. The authors also find that instance specificity emerges as referring expressions become contextualized. Layers where spatial information can be recovered depend on how visual representations are injected into the language model, which is a very nice architecture-specific finding! The paper distinguishes between successful answering, representational presence, decodability, and causal use. The authors use suppression, retention, and reweighting interventions to investigate whether the recovered visual evidence influences model behavior.

Critical strengths
  • The authors show that representations progressively shift from category-level to instance-level localization. It would be interesting to investigate how this process evolves across language model layers.

  • Much of the cross-modal correspondence can be recovered with a low-dimensional “bridge”. The authors also include several useful controls, including semantic mismatch, concept swapping, and unseen-category transfer. I found the finding that incorrect labels cannot easily override geometry anchored by other categories compelling.

  • The suppression and retention experiments compare layers identified by the bridge against spatially permuted random controls. This provides evidence that the recovered locations are causally relevant.

Critical weaknesses
  • The paper misses several relevant works on VLM mechanistic interpretability, including [1], [2], and [3]. These investigate visual information routing, multimodal integration, and causal mechanisms in VLMs. The authors should better position their contribution relative to this literature.

  • Some of the methodological choices are not fully explanied. For example, the authors select different Bridge layers for each architecture, such as T19-V6 for Qwen. In section 4.5 the authors do describe a layer-screening procedure, but I am not sure how sensitive the results are to these choices. Figure 2 does not specify which model is used, or whether the similarity metrics are computed from the selected Bridge layers before or after projection.

  • The fact that a supervised bridge can recover visual information does not necessarily imply that the VLM itself represents this information in a low dimensional subspace. A targeted ablation of dimensions could strengthen this claim.

  • The bridge outperforms base ViTs in pointing accuracy. However, these systems differ substantially in architecture, training, and supervision. Is the improvement due to the VLM representations, multimodal training, or a supervised transformation?

  • The main evaluation uses RefCOCO. Although additional datasets are used for downstream experiments, the generalizability of the low-rank recovery finding is not established across different tasks.

References

[1] Gandelsman et al., 2024, Interpreting the Second-Order Effects of Visual Representations in Multimodal Models

[2] Ortu et al., 2026, When Seeing Overrides Knowing

[3] Golovanevsky et al., 2026, What Do VLMs NOTICE? A Mechanistic Interpretability Pipeline for Gaussian-Noise-free Text-Image Corruption and Evaluation

Questions for authors

See weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.