acceptodds
Under review as a conference paper at ICLR 2027

RegionTrace: Measuring How Image Regions Shape Multimodal Generation

Abstract

Reliable multimodal assistance requires understanding whether answers are grounded in relevant visual evidence. We present RegionTrace, a framework for tracking how image regions support the same answer throughout generation. At each stage, it holds the text generated so far fixed and measures how restoring an image region changes the answer score. Keeping the answer and regions fixed reveals where attribute errors find support, how support evolves during explanation, and whether visual revisits restore it. Experiments across three models and four benchmarks reveal that models often borrow attributes from the wrong object: in color-misbinding errors, the color-bearing distractor outweighs the queried region in 68.0–97.9 % of cases. Across 7,206 explanations, a complete-image probe matches the eventual answer in 86–92 % of traces before explanation begins, while regional support weakens after the first sentence in all six model–task settings. Explicitly revisiting the image does not reliably restore this support or improve accuracy. RegionTrace connects spatial attribution with conditional answer-support trajectories throughout multimodal generation.

Primary area:
interpretability and explainable AI
Keywords:
vision-language models, multimodal reasoning, multimodal interpretability, visual grounding, regional attribution

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_h7UCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces RegionTrace, a toolbox for measuring how image regions shape an answer throughout multimodal generation, and demonstrates its use through three empirical studies. Study I contrasts questions about target, elsewhere, and absent colors, using regional perturbations to identify visual evidence supporting incorrect attribute answers. Study II varies regional visibility while holding recorded reasoning prefixes and the eventual answer fixed at selected checkpoints to trace changes in visual support. Study III compares explicit Look-Back spans with equal-length ordinary reasoning continuations to assess changes in regional support and answer accuracy. The results show stronger contributions from distractor regions than from queried regions to color errors, declining incremental visual support during explanations, and no consistent Look-Back gains in regional support or accuracy.

Critical strengths
  1. Clear and systematic study design. The three studies connect spatial attribution of attribute errors with changes in conditional visual support during explanation generation and explicit visual revisits, demonstrating the toolbox’s diagnostic value.

  2. Broad quantitative evaluation. The experiments cover multiple models and four benchmarks, complemented by alternative masking, scoring-scale, and matched-continuation analyses that strengthen the empirical assessment.

Critical weaknesses

The proposed toolbox offers limited methodological novelty over existing attribution and reasoning-analysis methods. EAGLE already performs perturbation-based visual-region attribution for selected generated tokens conditioned on preceding text. Similar reasoning-analysis techniques have also been used in prior work, such as PDCR [1]. PDCR defines a Visual Dependence Score as the log-likelihood ratio of a reasoning step under real versus blank images, conditioned on the same preceding text. RegionTrace’s main distinction is tracking signed, context-dependent regional effects on the same candidate answer across generation checkpoints, with Shapley averaging as a summary. While useful, the methodological advance provided by this integration over existing fixed-prefix perturbation and attribution approaches remains insufficiently established.

[1] PDCR: Perception-Decomposed Confidence Reward for Vision-Language Reasoning, CVPR 2026.

Questions for authors

Could the authors more clearly articulate RegionTrace’s core incremental contribution over existing work and its implications for future research on visual grounding in VLMs?

How well does its fixed-target analysis generalize to practical VLM applications, particularly open-ended generation beyond the evaluated Yes/No and multiple-choice settings?

In Study I, how often does the elsewhere region have a positive signed contribution among cases where it outranks the queried region? Separating positive support from weaker suppression would clarify the interpretation that distractor regions support incorrect answers.

6 comments

W1 & Q1: Core contribution

Thank you for raising this point. RegionTrace connects regional attribution to a central question in visual reasoning: what does reasoning actually change about the visual evidence supporting an answer? A wrong answer can draw support from a real visual cue in the wrong region. As an explanation grows, the generated text can sustain the same answer while regional support weakens. Even an explicit second look does not reliably restore that support.

RegionTrace makes these questions accessible within one framework by following the same answer and image regions across reasoning stages, while retaining whether a region supports or opposes the answer in different visual contexts. Its core contribution is to connect where an answer’s visual support comes from with what happens to that support during reasoning. This gives researchers a way to examine whether explanations and visual revisits actually change the evidence behind a decision, alongside whether they improve accuracy. We compare RegionTrace with EAGLE and PDCR below.

Comparison with EAGLE

EAGLE combines sufficiency and necessity to rank image regions for selected generated tokens. A ranking alone, however, does not tell us whether a region consistently supports the answer. The same region may support an answer when shown alone but oppose it when other regions are visible. RegionTrace preserves these opposing effects and tracks how support for the same answer changes across visual contexts and reasoning stages (Table 20, p. 36).

Across six Study I settings, 63.0–67.2% of false-Yes cases contain a region that supports the wrong answer in one visual context but suppresses it in another. EAGLE’s combined score can hide this reversal: opposing effects can cancel. RegionTrace retains the signed effects for each context (Appendix F.2, Table 22, p. 38).

Using the same regions, target answer, probability scores, and image masking, we compared RegionTrace with a matched EAGLE-style ranking on 366 PACO cases where the complete-image answer was Yes. RegionTrace increased Gold Top-1 from 76.5% to 78.4% and Drop AUC from 0.379 to 0.387 (Appendix F.1, Table 21, p. 37).

MethodGold Top-1 ↑Probability-drop AUC ↑
RegionTrace: exact Shapley78.4%0.387
RegionTrace: uniform-context average78.1%0.386
EAGLE-style sufficiency–necessity76.5%0.379
Full-context LOO76.0%0.364
Empty-context insertion76.2%0.370

We timed each method’s regional pipeline on the same 37 images, with Qwen2.5-VL-7B on one NVIDIA A40 (Appendix F.3, Table 23, p. 38). RegionTrace takes 4.11 seconds per image, including automatic region construction, versus 168.04 seconds for efficient EAGLE: approximately 41× faster, with comparable peak memory.

PipelineMean time/image ↓Peak GPU memory
RegionTrace: DINO + SAM + attribution4.11 s18.07 GiB
EAGLE: efficient V2, batch 8168.04 s18.05 GiB
EAGLE: full search, batch 8752.71 s17.46 GiB
EAGLE: full search, serial921.29 s16.44 GiB

Comparison with PDCR

PDCR replaces the image with a blank one and measures how the probability of a reasoning step changes, while keeping the preceding text fixed. It uses increases in confidence in the correct answer to guide training. RegionTrace addresses a complementary question: which image regions support or oppose the same answer, and how does that support change during reasoning? At each checkpoint, it measures signed regional effects under different visible contexts, keeping the answer target fixed whether correct or incorrect.

In Figure 2 (p. 5), the black mug body supports the incorrect answer by +5.87 log-score units, while the queried white rim opposes it by −0.44. Separating these contributions localizes the misleading support. In Study II, regional support decreases across all six settings from the first sentence to the final checkpoint, even as the text increasingly sustains the same answer (Appendix B.3, Table 9, p. 24). We added the PDCR comparison and citation to Section 6 (p. 10).

Core contribution and implications

RegionTrace traces an answer’s visual support through reasoning and revisits. It could also test whether a GUI agent’s action depends on the intended interface element. See our response to reviewer P8vJ, Q2 for further implications.

Q2: Application to open-ended generation

In fact, we had already explored open-ended generation while developing RegionTrace. We examined which image regions influence a freely generated description and where their effects are strongest. We used 500 AMBER images and InternVL3-8B, which produced descriptions averaging 114 tokens without predefined answer options.

We first record each description, then hide or restore two selected image regions and rescore the recorded tokens. Each token and its preceding text stay unchanged during the comparison. To measure the strength of regional influence, we add the absolute effects of the two regions on each token. The table reports the average for the first and last quarters of each description, in log-probability units.

First quarterLast quarterChange (95% CI)Lower in last quarter
1.43390.3781−1.0558 [−1.1201, −0.9909]470/500 (94.0%)

We computed the confidence interval by resampling the 500 images. This measure was about 74% lower in the last quarter, with a decrease in 94% of descriptions. A plot across ten parts of each description shows a sharp early drop, followed by generally weaker effects. We have added this plot, with 95% confidence bands, to the revised manuscript (Figure 16, p. 40). The plot is also provided in the supplemental material, with the corresponding summary statistics on p. 3 (Table 24).

In these descriptions, the selected regions had a larger average effect on early tokens than on later ones. This illustrates a practical use of RegionTrace in open-ended generation: locating the parts of a response most affected by particular image regions.

Q3: Positive support from distractor regions

Thank you for pointing out this distinction. We checked the sign of the elsewhere region’s contribution among false-Yes cases where it outranks the queried region. Its contribution is positive in 97.9–99.6% of PTR cases and 91.0–94.9% of PACO cases across models, as detailed below.

DatasetModelPositive elsewhere contributionNon-positive elsewhere contribution
PTRQwen2.5-VL-7B265/266 (99.6%)1/266 (0.4%)
PTRInternVL3-8B746/752 (99.2%)6/752 (0.8%)
PTRInternVL3-14B428/437 (97.9%)9/437 (2.1%)
PACOQwen2.5-VL-7B75/79 (94.9%)4/79 (5.1%)
PACOInternVL3-8B201/221 (91.0%)20/221 (9.0%)
PACOInternVL3-14B126/136 (92.6%)10/136 (7.4%)

The table shows that the elsewhere region usually supports the wrong answer, rather than simply suppressing it less than the queried region.

Hiding only the elsewhere region also switches the constrained Yes–No decision to No in 76.0–83.6% of PTR errors and 35.5–53.9% of PACO errors (Section 3.2, Table 2, p. 6). These rates cover all elsewhere-color false-Yes cases, not just those where elsewhere ranks higher. The sign breakdown is included in Appendix A.3, Table 5 (p. 20) of the revised manuscript and in the supplemental material, with the minority of zero or negative contributions explicitly identified.

Thank you for the additional analyses. My main concern remains that the paper does not establish a substantial methodological advance over existing perturbation and fixed-prefix analyses. As P8vJ points out, keeping the reasoning prefix fixed measures how visual evidence affects the answer given the recorded reasoning, leaving open how changes to the image would alter the reasoning itself. I therefore see RegionTrace as a useful diagnostic extension, but my overall rating remains unchanged.

Reviewer_P8vJOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RegionTrace, a framework to understand how image regions affect VLMs' answers to visual reasoning tasks. The paper presents three studies to showcase the findings using RegionTrace on VLMs.

Critical strengths

The paper is well motivated: it develops a framework to trace which part of the input image is important for the VLM to generate its answers.

Critical weaknesses
  • It is unclear to me what significant insight this paper provides, and I'm not fully convinced by the proposed framework. Perturbation-based analysis for classic machine learning models and modern transformer models has been well studied. The paper's focus is narrowing down to one particular setting: mask out the image by regions, and see how the scores change for the same answer. Study I shows results of this setting for one attribute: color in the visual reasoning task. It is unclear how this generalizes to other attributes or image properties. Then, for the chain-of-thought setting, I'm not convinced that fixing the reasoning sequence while the input (region) is changed is the right approach: in the unconstrained case, the model will output different reasoning steps if an image is masked differently. Finally, the whole approach is based on image regions marked by humans or segmentation tools, but it is unclear if the VLMs studied are actually understanding the image in the unit of those regions.
  • This paper reads more like a technical report of experiment logs and findings than a strong ICLR paper with clean writing and careful organization. The paper is overly descriptive in many straightforward settings (e.g., Sec 2.1-2.3) but then skims over important details: e.g., how the candidate-normalized probability of a fixed answer is computed. Oftentimes, the paper will use words or terms without context or explanation. In general, I found this paper sometimes hard to follow due to the quality of the writing. I have observed similar writing styles in heavily-AI assisted/generated writings. The presentation of the paper needs to be improved.
Questions for authors
  1. For the region effect score (e.g., +5.87 in Figure 2), does the raw number provide any useful information? It seems the sign matters most: e.g., positive values support, and negative values suppress the answer.
  2. Can the authors discuss what is the significance of the findings by the proposed RegionTrace framework?
  3. Can the authors discuss how a predefined region aligns with what a VLM actually sees in the image?
5 comments

W1: Fixed-prefix analysis and generalization

We would first like to clarify a possible misunderstanding about what we hold fixed in our analysis. Keeping an answer fixed for scoring does not mean requiring the model to keep producing that answer. RegionTrace is an analysis toolbox: our aim is to trace how image regions change the model’s preference for that answer.

As is well known, an autoregressive model computes a whole vector of logits at each decoding step, from which we can obtain the log-probabilities of different candidate tokens. The token eventually selected is only one outcome of this distribution. We can examine the scores of other candidates as well.

Take a simple question, “Is the cup red?” Suppose the model prefers Yes with the complete image. After masking a region, it may instead prefer No:

Visual conditionP(Yes)P(No)Model’s preferred candidate
Complete image0.800.20Yes
Region rr masked0.300.70No

After masking, the model may prefer No, while the probability of Yes falls from 0.80 to 0.30. Keeping Yes as the scoring target lets us measure this loss of support; it does not require the model to keep producing Yes.

Of course, we could also generate a new response for each modified image and compare it with the original response. Huy et al. (2026) perturb images while keeping the question fixed and quantify semantic disagreement among the generated answers. Such comparisons measure output variability, but the answer and any preceding reasoning may both change. They therefore do not directly isolate changes in support for a fixed answer. We instead score the same target under the same textual context, so that the score difference reflects the effect of the image-region intervention.

At each reasoning checkpoint, we vary image regions while holding the recorded prefix fixed. As the prefix grows, we track regional support for the same answer. This conditional comparison does not assume that modified images would reproduce the recorded reasoning. Studies II and III use it to test whether regions provide additional support after reasoning is written and whether an explicit visual revisit restores that support.This design allows RegionTrace to reveal how the regional visual support for an answer changes during reasoning, even when the answer itself remains unchanged.

Scope of Study I: generalization beyond color

Study I uses color binding to localize visual support for incorrect answers, rather than systematically test cross-attribute generalization. To address the reviewer’s concern, we tested InternVL3-8B on 40 PTR images selected from scene annotations before inference. Each image had two left–right questions with opposite correct answers.

We used three fixed vertical regions: one for each queried object and one for other objects. We computed exact Shapley Region Effects across all eight visibility combinations, using uniform-gray masking and the correct answer’s log-probability margin over the alternative. The contribution advantage is the mean contribution of the two queried-object regions minus that of the third region.

MeasureResult
Answer accuracy, original images78/80 (97.5%)
Answer accuracy, uniform-gray images40/80 (50.0%)
Contribution advantage > 070/80 (87.5%)

The mean advantage was 2.01 log-score units (image-bootstrap 95% CI: 1.59–2.42; Appendix A.5, Table 7, p. 21). This mirrors the pattern observed for correctly answered color questions: task-relevant regions provide stronger support for the correct answer than other regions. Together, these results show that the regional-support pattern identified by RegionTrace extends beyond color to spatial reasoning.

What RegionTrace adds to perturbation-based analysis

We agree that perturbation-based analysis is well established. RegionTrace connects regional attribution with changing support for the same answer across visual contexts and reasoning stages, including before the answer is written. At each checkpoint, we fix the textual prefix and measure how restoring a region supports or suppresses that answer. In 63.0–67.2% of false-Yes cases across six Study I settings, a region supports the same answer in one visual context but suppresses it in another (Appendix F.2, Table 22, p. 38). We discuss this contribution further, including comparisons with EAGLE and PDCR, in our response to Reviewer h7UC (W1/Q1).

Huy et al. Do Vision Language Models Recognize Visual Ambiguity? ECCV 2026.

W2: Clarity, organization, and answer scoring

Thank you for the specific feedback on clarity and organization. We agree that the presentation can better balance routine setup with essential methodological details. We have streamlined Sections 2.1–2.3, explained technical terms at first use, and removed repetitive descriptions.

In particular, the candidate-normalized probability used in Studies II and III should have been explicitly defined in the main text. Let ℓt,d(S)\ell_{t,d}(S) denote the model log score for answer candidate dd, evaluated with image ISI_S, the saved prefix at checkpoint tt, and the answer-scoring cue. For the fixed eventual answer d∗d^\ast, we compute

=\frac{\exp\left(\ell_{t,d^\ast}(S)\right)} {\sum_{d\in\mathcal D}\exp\left(\ell_{t,d}(S)\right)},$$ where $\mathcal D=\{\mathrm{Yes},\mathrm{No}\}$ for MM-GCoT and $\mathcal D=\{A,B,C,D\}$ for VCR. The target $d^\ast$ remains fixed across checkpoints and image interventions, whether correct or incorrect, and the answer itself is excluded from every saved prefix. This normalization is implemented in the `target_score` function (https://github.com/regiontrace-anon/regiontrace/blob/main/code/regiontrace/core.py#L6-L22) in our **publicly available anonymous code repository** and used by `measure_trajectory` (https://github.com/regiontrace-anon/regiontrace/blob/main/code/regiontrace/trajectory.py#L23-L48). This definition is now explicit in **Section 2.4**, with the answer-token specifications and scoring-cue details in **Appendices B.1 and C.1**. We appreciate the reviewer’s feedback on the organization and presentation. The three studies were designed as complementary parts of a single investigation, using a common framework to examine where an answer’s visual support comes from, how it changes during reasoning, and whether visual revisits restore it. We have revised the presentation to make this shared motivation and progression more explicit. Specifically, to investigate visual grounding during generation, RegionTrace measures support for the same answer across image regions, visual contexts, and generation stages. Studies I–III examine spatial misbinding, changes in visual support during explanation, and visual re-examination, using paired questions, image-removal and deletion checks, and length-matched continuation controls. Together, they show that real visual evidence can support a spatially incorrect answer, that its additional contribution can decline as the explanation grows, and that the tested Look-Back variants do not consistently restore this support or improve accuracy. We have revised the presentation to make this connection between the research question, method, experimental evidence, and conclusions more explicit. ### Q1: Interpreting Region Effect magnitudes The sign indicates whether a region supports or suppresses an answer. The magnitude lets us compare how strongly different regions within the same image contribute to that answer. For example, consider +5.87 and +0.19 in **Figure 2 (p. 5)**, under the same scoring setup: | Region effect | Direction | Magnitude | | ------------- | ------------ | ------------------------------------------------------ | | +5.87 | Supports Yes | Larger average increase in the Yes–No log-score margin | | +0.19 | Supports Yes | Smaller average increase in the same margin | Both regions have the same sign, but if they support a wrong answer, we still want to know which one provides stronger support for that error. The magnitude makes this distinction possible. Specifically, Figure 2 uses the Yes–No log-score margin: $$V(S)=\log p_\theta(\mathrm{Yes}\mid q,I_S) -\log p_\theta(\mathrm{No}\mid q,I_S).$$ The reported +5.87 is the body region’s Shapley-weighted average marginal increase in this margin across visual contexts. It quantifies support on this scoring scale, not a probability increase of 5.87 percentage points. This matters for our findings. In Study I, we compare whether the color-bearing distractor provides stronger support than the queried region. In Studies II and III, we track whether regional support weakens or returns during reasoning. **An effect can remain positive while its magnitude decreases substantially**, which the sign alone would miss. With the scoring setup held fixed, the magnitude lets us compare how strongly different regions support or suppress the same answer.

Q2: Significance of the three studies

Before discussing the findings of the three studies, we would like to explain the broader purpose of RegionTrace. As more methods are developed to improve VLM reasoning, we also need tools that help us understand what these methods actually do and provide a clearer view of model behavior. Which visual evidence supports an answer? How does its contribution change during reasoning? Does a visual revisit restore its influence? RegionTrace provides a way to investigate these questions, helping researchers make more informed decisions about how to use and improve reasoning methods.

With this goal in mind, we selected the three studies to demonstrate different uses of this analysis.

In Study I, we investigate where an incorrect answer obtains visual support. The color-binding experiments show that genuine evidence from one region can support an incorrect answer about another: the relevant color is present, but on the wrong object or part. This provides a more specific diagnosis of the error and shows why assessing visual grounding requires identifying the source of support, beyond establishing that the answer depends on the image.

In Study II, we examine how that support changes during explanation generation. As the explanation grows, the recorded text increasingly sustains the eventual answer, including incorrect answers, while the selected regions provide less additional support given that text. This helps us understand how an answer can persist through an explanation and lets us examine its continuing dependence on visual evidence.

In Study III, we evaluate a reasoning intervention. The tested Look-Back variants do not consistently restore regional support or improve accuracy compared with equally long ordinary reasoning. This directly tests whether an explicit visual revisit achieves its intended effect, illustrating how RegionTrace can evaluate a method’s behavior alongside its final accuracy.

These studies illustrate uses of RegionTrace without defining its full scope. The approach could be adapted to other settings where meaningful image regions and a target output can be specified and scored. Potential applications include examining whether a GUI agent’s action depends on the relevant interface element, or whether a medical VLM’s answer is supported by the image region under discussion. More broadly, RegionTrace gives researchers a tool to investigate their own questions about visual evidence and model behavior.

Q3 and the final part of W1: Input regions and model representations

Thank you for raising this question. The predefined regions specify which parts of the input image we change; they do not prescribe how the VLM should divide or understand the image. We hide or restore the pixels within a region before the image enters the visual encoder. The modified image then goes through the model’s usual image-processing pipeline.

This is how our regions relate to what the model receives. The model can use the restored content in any way: it may rely on small details within the region, combine them with information elsewhere, or change its interpretation of the surrounding scene. RegionTrace measures how these changes, taken together, affect the score of the same target answer. The question and preceding text remain fixed during this comparison. We therefore do not need to assume that the model treats each predefined region as a separate object or part.

Appendix G.2 helps explain this distinction. Restoring one region can affect more than the model’s processing of that region alone. It can also change how information from different parts of the image is combined. Under its stated assumptions, Proposition G.2 (Eq. 25, p. 43) shows how these contributions together produce the final score change, without requiring the changes inside the model to stay within the restored region.

We choose meaningful objects or parts as regions so that the intervention and its results are easy to interpret. An effect assigned to a region describes how restoring its image content changes support for the answer; it does not establish that the model internally represents that region as a separate unit. We have clarified this distinction in Section 2.2 (“Defining the regions”) and directed readers to Appendix G.2.

Reviewer_uZ6UOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces RegionTrace, a framework for tracking how image regions support a fixed answer throughout multimodal generation. Through region masking and restoration, it studies attribute misbinding, changes in visual support during explanation, and the effects of explicitly revisiting the image.

Critical strengths
  1. The framework tracks support for the same answer across generation stages, providing a useful connection between spatial attribution and the evolution of visual evidence use.

  2. Experiments across multiple models and benchmarks provide informative findings, particularly that explicitly revisiting an image does not reliably restore visual support or improve accuracy.

Critical weaknesses

The main concern is the limited empirical analysis of internal model behavior under masking. The paper compares regional attention with Region Effect in selected examples, but a systematic comparison of attention before and after masking or restoring regions would strengthen the analysis. In particular, layer- and head-level results could help explain how these interventions redistribute attention and how those changes relate to the measured Region Effects.

Questions for authors

Could the authors provide layer- and head-level attention comparisons before and after masking or restoring regions across a representative set of examples?

4 comments

W1 & Q1: Attention changes across layers and heads

Thank you for this suggestion. Our new analysis shows that masking a region changes attention more in middle-to-late layers, especially layers 19–24, with clear differences across heads. Regions that cause larger attention changes also tend to have larger effects on the same answer's probability. This relationship is strongest before the explanation and weaker near the final answer.

In the revised manuscript, Fig. 5 (p. 8) gives an overview of the attention changes. Appendix G.1 provides the complete before/after comparisons in Fig. 17 (p. 41), the per-head associations in Fig. 18 (p. 42), and the paired statistics in Table 25 (p. 42). These figures and the corresponding statistics are also collected in the supplemental material, retaining their numbering from the revised manuscript.

We examined all 28 layers and 28 heads per layer of InternVL3-8B on all 404 valid MM-GCoT judgement questions in our frozen evaluation set (350 images). At each of the three stages below, we keep the text prefix and original final answer yy (Yes or No) fixed while masking one region at a time. The answer itself is excluded from the prefix.

Let II be the complete image and I−iI_{-i} the image with region ii masked. At checkpoint tt, at,j(ℓ,h)(J)a_{t,j}^{(\ell,h)}(J) denotes the attention weight from the position predicting the first answer token to key position jj, in layer ℓ\ell and head hh, for image JJ. We measure the attention change using total variation distance (TV):

Di,t(ℓ,h)=12∑j∣at,j(ℓ,h)(I)−at,j(ℓ,h)(I−i)∣.D_{i,t}^{(\ell,h)}=\frac{1}{2}\sum_j\left|a_{t,j}^{(\ell,h)}(I)-a_{t,j}^{(\ell,h)}(I_{-i})\right|.

The sum covers all key positions—visual, text, and special tokens—so TV measures how much attention is redistributed after masking.

For answer label dd, let Pt(d∣J)P_t(d\mid J) be its model likelihood with the text input fixed at checkpoint tt. The normalized probability of the original answer yy and its regional effect are

py,t(J)=Pt(y∣J)Pt(Yes∣J)+Pt(No∣J),Ei,t=py,t(I)−py,t(I−i).p_{y,t}(J)=\frac{P_t(y\mid J)}{P_t(\mathrm{Yes}\mid J)+P_t(\mathrm{No}\mid J)},\qquad E_{i,t}=p_{y,t}(I)-p_{y,t}(I_{-i}).

A positive Ei,tE_{i,t} means masking lowers the answer's probability, so the region supports that answer; a negative value means it suppresses it. Because TV measures the magnitude of attention change, we correlate it with ∣Ei,t∣|E_{i,t}|. For the table, TV is computed per head before averaging across all 28×2828\times28 layer–head pairs. The other attention measure is the original attention mass on the region, weighted by its fractional coverage of visual tokens and averaged across the same layers and heads:

Correlation with ∣Ei,t∣\lvert E_{i,t}\rvertBefore explanationAfter first reasoning stepBefore answer
Attention paid to the region before masking0.106 [0.024, 0.187]0.044 [−0.035, 0.124]−0.018 [−0.091, 0.100]
Change in attention after masking (TV)0.651 [0.605, 0.692]0.444 [0.385, 0.503]0.153 [0.103, 0.203]

The Pearson correlations compare regional differences within questions; brackets show 95% confidence intervals from resampling whole images. These associations do not establish which individual heads cause the answer effect.

Dear Reviewer,

Thank you for taking the time to review our paper and share your feedback. We’d be happy to clarify anything in our response if helpful.

Best regards, The Authors

Dear Reviewer,

Thank you for taking the time to review our paper and share your feedback. We’d be happy to clarify anything in our response if helpful.

Best regards, The Authors

Thank you for the thorough layer and head level analysis, which addresses my main concern. I will keep my positive score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of discussion and remaining concerns

Dear Area Chair,

Thank you for overseeing the discussion. We would like to summarize the remaining concerns and highlight the evidence relevant to your assessment. Reviewer uZ6U has confirmed that our additional layer- and head-level analysis addresses their main concern and has maintained their positive recommendation. The remaining disagreement with Reviewers P8vJ and h7UC concerns the fixed-prefix setting and the significance of the contribution.

The fixed-prefix analysis addresses a specific diagnostic question: given the text generated so far, how much additional support does an image region provide for the same answer? Holding the prefix fixed makes this comparison possible without simultaneously changing the textual context. Studying how image interventions alter freely generated reasoning addresses a complementary question; it does not by itself establish whether the conditional measurements support our stated conclusions.

Our response to Reviewer h7UC provides a concrete example. The intact and region-masked images produce identical reasoning and final answers, token for token. Yet, after their shared first reasoning step, the normalized probability of Yes is 99.14% with the intact image and 37.75% with the region masked. Comparing the generated responses alone would miss this difference. Across 404 questions, 58 had at least one single-region mask that preserved the entire generated response. These observations illustrate the information that the conditional analysis can reveal, including when the recorded prefix is also generated under the intervention.

The contribution should also be assessed through the diagnostic findings enabled by the framework. RegionTrace follows signed regional support for the same answer across visual contexts and generation stages. The three studies use this framework to locate support for attribute errors, measure changes in additional regional support as explanations grow, and test whether explicit visual revisits restore that support. Our responses also provide comparisons with EAGLE and PDCR to clarify this contribution relative to existing methods.

We would appreciate your assessment of RegionTrace’s contribution in light of its stated research question and the evidence presented in the manuscript and discussion. In particular, we ask that the fixed-prefix design be evaluated against its intended diagnostic purpose, and that the framework’s contribution be considered together with the findings enabled by the three studies.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.