acceptodds
Under review as a conference paper at ICLR 2027

From What Is Seen to What Is Meant: How VLMs Infer Implication from Images

Abstract

Vision-language models can describe what an image depicts yet misinterpret what it means. This paper asks where that inference fails and examines four components: visual-evidence grounding, source-relation selection, target-domain mapping, and answer binding. We first conduct a four-stage audit of 975 incorrect II-Bench responses and assign 76.1% to source-relation selection or target-domain mapping, compared with 16.9% to explicit visual-evidence errors. Because response labels alone do not establish the computations behind these errors, we conduct controlled tests. Removing image content reduces accuracy, while three visual-attention summaries do not reliably distinguish correct from incorrect interpretations. When five true relations from the same image compete, the question-relevant relation ranks first in only 46.3% and 34.9% of full-context cases across the two models. Transferring relation-conditioned states changes both relation and answer preferences. With the correct relation fixed, implication-specific states steer preferences for held-out implications, and supplying an aligned implication improves final answer selection. These results show that the key bottleneck in image-meaning inference is not whether a model sees an image, but whether it can form and use the implication supported by the right relation among visible elements.

Primary area:
interpretability and explainable AI
Keywords:
Visual language model, Implication understanding, Mechanism Analysis

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

Vision-language models can describe what an image depicts yet misinterpret what it means. This paper asks: once a VLM has accessed the relevant visual evidence, what happens between what is seen and what is meant? To study this transition, this paper examines four components: visual-evidence grounding, source-relation selection, target-domain mapping, and answer binding. Experimental results show that the key bottleneck in image-meaning inference is whether it can form and use the implication supported by the right relation among visible elements.

Critical strengths

Figures and Tables are aesthetically well-presented.

Critical weaknesses
  1. Every causal claim in the paper rests on the same 35 examples, and the selection of those examples is never described.

  2. Section 4.1 needs to report the absolute mean-fill accuracies and add a black-image and a no-image condition so that the image-necessity bound is explicit.

  3. In Figure 4(a,b), the condition that adds the answer options appears to help more than the condition that adds the image. This is an interesting and slightly awkward finding: if answer options constrain relation choice more effectively than the image does, then the full-context condition may be measuring option-driven reasoning rather than image-driven relation selection, which is close to the opposite of the paper's framing. The paper does not comment on this ordering anywhere. There needs to be a top-1 rate for each context condition numerically in the main text, and say why the options are so informative.

Questions for authors

Q1: What are the selection criteria/filtering procedure for the 35 examples?

Q2: What are the absolute mean-fill accuracies in Section 4.1, and how does the performance compare under the black-image versus the image-free conditions?

Please refer to the "Critical Weaknesses" section for further details.

5 comments

Dear Reviewer ffpB:

We thank the reviewer for the thoughtful and rigorous feedback.


Q1

Every causal claim in the paper rests on the same 35 examples, and the selection of those examples is never described. What were the selection criteria or filtering procedure for these 35 examples?

A1

Thank you for pointing this out. The 35 examples strictly follow the official development split provided by II-Bench. We adopted this predefined split directly. This ensures an objective and standard test setting. At the same time, we combined fine-grained interventions on this development set with large-scale validation on the full test set to ensure that our findings are reliable. We will modify Section 3.1 of the manuscript to make these details clearer.

The reasons why we used this development set for detailed intervention experiments are as follows. First, we separated exploration from validation in our research design. We wanted to carry out detailed analyses and draw conclusions on the development set. Then we tested and verified those conclusions on the complete test set. This helped avoid data leakage on the test set. Also, the controlled tests in Sections 4 through 7 required expensive human annotations. We could not easily apply these across the entire benchmark. These annotations include

  • Region-level annotation of question-relevant visual regions (Key ROIs) projected onto 8×88 \times 8 visual-token grids
  • Hand-crafted creation of 4 visually true same-image competitor relations for each sample (Table S16)
  • Creation of 3 target implication candidates for each sample, with each candidate written in 3 different ways (Table S17)

To make the findings of the paper more reliable, we also did the following work

  • Controlled evaluations across six A to F answer-label rotations and five A to E relation-label rotations across the development-set experiments, producing thousands of forward passes and internal state interventions
  • A four-stage error audit across all 975 wrong answers from the full test examples in II-Bench (Figure 2)
  • Image-content interventions across the 1,371-example II-Bench evaluation set and 765 test examples in the external CII-Bench (Table 1, Table S4)
  • Structured-LoRA adaptation tests across three random seeds on II-Bench and external CII-Bench transfer (Table 2, Tables S12 to S15)
  • Wide evaluations across two model families (Qwen3-VL and InternVL3.5) and three model sizes (2B, 4B, and 8B) to test general performance

We intend to do our best to verify these findings on a larger set and include them in the final manuscript.

Q2

Section 4.1 needs to report the absolute mean-fill accuracies and add a black-image and a no-image condition so that the image-necessity bound is explicit. What were the absolute mean-fill accuracies, and how does performance compare under the black-image versus no-image conditions?

A2

Following your suggestion, we evaluated the absolute accuracies across 5 input conditions on the exact same evaluation splits used in Table 1, Table 2, and Table S4 (the shared 1,371-example evaluation set on II-Bench and the 765-example test split on the external CII-Bench, under the fixed-order evaluation protocol of Table 2 and Tables S13 to S15, while the 62.4% and 67.9% reported in Section 1 are free-form generation accuracies across II-Bench examples before excluding the 28 unresolved audit records). Overall, the black-image and no-image conditions perform similarly and remain well above random chance due to language priors in multiple-choice questions, while adding the original image brings clear and consistent accuracy gains over all visual-ablation baselines.

  1. Original image, full multimodal input (upper baseline, matching the identity-order baselines in Table 2 and Tables S13 to S15)
  2. Mean-fill, replacing pixels with average color values (removes shapes but keeps overall color)
  3. Black image (zero-fill), setting all image pixels to zero (keeps image token placeholders without visual content)
  4. No image (text-only), removing the image completely (pure text baseline or blind guessing)
  5. Random chance, 6-choice baseline (16.67%)

The complete absolute accuracy results and paired drops are summarized below.

BenchmarkModelOriginalMean-FillΔAcc\Delta\text{Acc} (Mean-Fill minus Orig.)Black-ImageNo-Image (Text-Only)Random Chance
II-Bench (N=1,371N=1,371)Qwen3-VL-2B63.31%46.24%−17.07 pp46.68%47.34%16.67%
II-Bench (N=1,371N=1,371)InternVL3.5-2B68.20%54.92%−13.27 pp55.00%53.32%16.67%
II-Bench (N=1,371N=1,371)Qwen3-VL-4B75.86%56.60%−19.26 pp57.70%58.72%16.67%
II-Bench (N=1,371N=1,371)InternVL3.5-4B75.20%63.31%−11.89 pp64.77%61.12%16.67%
CII-Bench (N=765N=765)Qwen3-VL-2B38.69%26.27%−12.42 pp27.06%28.50%16.67%
CII-Bench (N=765N=765)InternVL3.5-2B46.54%34.77%−11.76 pp35.16%33.33%16.67%

Analysis and Findings

  1. Textual priors (language shortcuts)
    Without images or with black images, models reach 46.68% to 64.77% accuracy on II-Bench and 27.06% to 35.16% on CII-Bench. Both ranges are much higher than random chance (16.67%). This confirms that multiple-choice questions provide common world knowledge and answer clues. Models can rule out bad options using language priors alone.
  2. Visual necessity
    Across all tested models, adding the original image brings a clear increase of 11.76 to 19.26 percentage points over mean-fill, 10.43 to 18.16 percentage points over black-image, and 10.19 to 17.14 percentage points over the text-only baseline. The mean-fill differences exactly match the paired drops reported in Table 1 and Table S4 (all adjusted qq values below 0.001). Existing vision-language models can clearly recognize basic objects and use this visual information to remove incorrect options.

We will add these complete absolute accuracies to Section 4.1 and Appendix D.1 to enrich the paper.

Q3

In Figure 4(a,b), the condition that adds the answer options appears to help more than the condition that adds the image. This is an interesting and slightly awkward finding. If answer options constrain relation choice more effectively than the image does, then the full-context condition may be measuring option-driven reasoning rather than image-driven relation selection, which is close to the opposite of the paper's framing. The paper does not comment on this ordering anywhere. There needs to be a top-1 rate for each context condition numerically in the main text, and say why the options are so informative.

A3

Thank you for this helpful question. In short, answer options provide a top-down semantic filter to rule out unrelated local relations, while adding the image triggers strong visual competition among multiple true relations in the picture, and this visual interference directly lowers accuracy in the full context, showing that the process is not purely option-driven. We first report the exact Top-1 selection rates for the gold relation across the five context conditions.

  • Qwen3-VL-2B, relations only R (42.1%) to question plus relations Q+R (44.6%) to question plus image plus relations Q+I+R (34.3%, sharp drop) to question plus options plus relations Q+O+R (55.4%, sharp rise) to full context Q+I+O+R (46.3%)
  • InternVL3.5-2B, relations only R (42.0%) to question plus relations Q+R (45.1%) to question plus image plus relations Q+I+R (24.9%, sharp drop) to question plus options plus relations Q+O+R (50.3%, sharp rise) to full context Q+I+O+R (34.9%)

The main reasons for this ordering are as follows.

  1. Why does adding the image Q+I+R decrease accuracy compared to Q+R?
    In our design, the 4 competitor relations are visually true same-image competitors that describe real local details in the picture (Table S16). Therefore, the model cannot simply rely on object recognition to drop bad choices.
  • Under the text-only condition (Q+R), these small local details look less related to the main question. Language patterns therefore make it easier to pick the broader gold relation.
  • When we add the image (Q+I+R), the vision encoder creates strong visual signals for these visible objects. The model faces competition between multiple visually true details in the picture. It gets distracted by secondary elements. This causes the Top-1 selection rate to fall by 10.3 to 20.2 percentage points.
  1. Why do the answer options O provide so much information? (Q+O+R is the highest)
    Image implication inference connects a concrete source-domain relation R to a target implication expressed in the answer options O.
  • The options describe high-level target implications, such as social or psychological ideas.
  • Giving options O works as an effective top-down filter. By comparing candidate relations with target options, the model easily removes minor local details that do not support the intended meaning.
  1. Why this does not mean the process is purely option-driven
    If relation selection depended only on options and ignored the image, the full context (Q+I+O+R) would perform as well as or better than Q+O+R.
    However, we see the opposite result. Adding the image back into the full context causes accuracy to fall by 9.1 percentage points in Qwen (from 55.4% to 46.3%) and by 15.4 percentage points in InternVL (from 50.3% to 34.9%).
    This provides clear evidence. Visual interference from multiple visually true same-image relations is very strong. It directly weakens the top-down guidance provided by the text options. We will add these exact numbers and explanations to Section 5.1 of the revised manuscript.
Reviewer_bMrcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates why vision-language models can recognize image content yet fail to infer its intended meaning. The authors organize the task into four components: visual grounding, source-relation selection, target-domain mapping, and answer binding.

Critical strengths

The paper asks an interesting question, and I found the distinction between selecting a concrete relation and interpreting its meaning useful. Several experimental choices are thoughtful, particularly the use of true same-image competitors, answer-label rotations, and held-out paraphrases with random and cross-sample intervention controls.

Critical weaknesses
  1. The intervention direction is constructed from supplied implication texts for the same sample, and evaluation uses another paraphrase of that implication. The held-out paraphrase control is valuable. Still, the result establishes that an implication-conditioned direction can favor semantically related text. It does not by itself establish how the model normally derives that implication from the image. Testing whether the effect depends on the image and the selected source relation would help distinguish the proposed mapping mechanism from more general semantic steering.

  2. The claim that relation selection and target mapping account for some errors rests on annotations of generated explanations. A response can mention the relevant objects while still misunderstanding an attribute or the direction of an interaction. Conversely, the explanation may not faithfully expose the computation that produced the answer. This matters because Table S3 requires an explicit incorrect visual claim to assign a grounding error. That is a defensible conservative annotation rule, but it leaves some visual failures unobservable. The resulting distribution should therefore be presented as a distribution of identifiable response errors, with appropriate caution about what it says about internal bottlenecks.

  3. I would also like to see actual annotation agreement and adjudication statistics. Appendix C describes two GPT-5.5 annotators followed by human review, but does not report the agreement rates or a stage-level confusion matrix.

Questions for authors

See Weaknesses.

6 comments

Dear Reviewer bMrc:

We thank the reviewer for the positive assessment and the constructive feedback.


Q1

The intervention directions are constructed from implication text for the same example, and evaluated on a held-out paraphrase. The held-out paraphrase control is valuable. Still, the results show that an implication-based direction may bias the continuation toward semantically related text. It does not by itself show how the model normally derives that implication from the image. Testing whether this effect depends on the image and the chosen source relation would help separate the proposed mapping mechanism from more general semantic steering.

A1 (page 1)

Thank you for this insightful question. Combining existing controls in the paper with new experimental results, we address this concern from three aspects. Both natural inference tests and controlled interventions confirm that target-domain mapping directly depends on real visual input and the specific source relation in the image rather than general text steering.

First, under natural inference without injecting any external direction vectors, we added a paired comparison between original images and equal-sized black images to test whether visual input directly drives natural implication representations.

  • We keep the question text, the six answer options, and the prompt identical, without providing any relation or implication hints, and replace the input with an equal-sized black image as a control.
  • We fix the text reference representations by using three human-written correct implication paraphrases as positive references and the incorrect options as competing references.
  • We take the hidden states of the implication segment across layers 19 to 27, compute the margin as the average cosine similarity to positive references minus that to competing references, and perform a paired test between original and black images.

After Benjamini-Hochberg correction, the paired gains in both model families reach statistical significance with q<0.05q < 0.05, and the 95% confidence intervals do not cross zero.

ModelComparisonMargin Difference [95% CI]BH qq
QwenOriginal minus Black+0.01401 [+0.00609, +0.02135]0.02015
InternVLOriginal minus Black+0.01287 [+0.00391, +0.02297]0.02985

Under identical text prompts, options, and reference anchors, language priors on the question text alone cannot achieve the same level of implication alignment. Only when the real image is present does the model shift its late hidden states toward the correct target implication during natural reasoning. This confirms that natural mapping directly depends on visual input.

A1(page 2)

Second, to test whether natural implication states also depend on the specific visual relation in the image, we added a controlled image-editing experiment across Original A, Content Edit B, and Appearance Control S.

  • We compare three image versions for each case. In case dev-23, Original A shows the Eiffel Tower under "The resume" and an ordinary utility pylon under "The candidate", which means a resume looks grand while actual ability falls far short. Content Edit B changes "The candidate" into the Eiffel Tower too, so actual performance matches the resume and the contrast disappears. Appearance Control S changes the color tone while keeping the Eiffel Tower versus utility pylon contrast, so the mismatch meaning remains.
  • We feed all three images with the exact same question ("What does this comic mean?") and the same incomplete answer prefix ("From the scene in the image, this comic implies..."), without giving any true or competing implications in the input. We then take the hidden states of the implication segment in layers 19 to 27.
  • We also build two fixed reference sets with 3 paraphrases for the original meaning and 3 paraphrases for the edited meaning. We feed these texts into the model under the original image context, take the average state of their content words as reference vectors, and share them across A, B, and S. We then measure which reference the test state is closer to by computing D as closeness to the original meaning minus closeness to the edited meaning.
ModelComparison on Relative Tendency DBH qq
QwenOriginal A and Appearance S versus Content Edit B0.0109
InternVLOriginal A and Appearance S versus Content Edit B0.0438

Original A and Appearance Control S keep a higher D, while Content Edit B significantly lowers D and shifts the hidden state toward the edited meaning in both models (Qwen BH q=0.0109q = 0.0109, InternVL BH q=0.0438q = 0.0438). This shows that during natural inference, late implication states depend directly on the visual relation in the image.

Third, existing controls and interventions across Sections 5 to 7 establish that the mapping mechanism depends on the image and the chosen source relation.

  • Section 6.2 (Figure 6c) rules out unconditioned general semantic drift. Direction vectors are constructed and evaluated on top of a shared image, question, and selected gold relation. Contrast S shows that the gold implication direction is significantly stronger than directions built for the other two competitor implications, whereas norm-matched random vectors (R) and cross-sample vectors (X) produce no effect (all q<0.001q < 0.001).
  • Section 5.3 (Figure 5b) shows that downstream answer choice depends on the selected relation state. Holding the image and question fixed, replacing a competitor relation state with the gold relation state in the residual stream significantly raises the gold-answer margin across layers 21 to 26 (q<0.05q < 0.05).
  • Section 7 (Figure 6d) shows that only an implication aligned with the image and source relation raises discrete top-1 answer accuracy from 61.0% and 65.2% to 77.0% on InternVL3.5-2B and Qwen3-VL-2B, whereas same-image competitor implications fail to provide comparable gains.

In the revision, we will add both natural visual-dependence experiments to the main text and clarify the relationship between natural mapping and controlled directional interventions in the discussion.

Q2

The claim that relation selection and target mapping explain parts of the errors is based on annotations of generated explanations. A response may mention relevant objects but still misinterpret an attribute or interaction direction. Conversely, explanations may not faithfully reflect the computations that produced the answer. This is critical because Table S3 requires an explicit visual error to categorize a response as a grounding error. This is a reasonable conservative rule, but it makes some visual errors unobservable. Therefore, the distribution should be presented as an observable response-level error distribution, and the internal bottlenecks it reveals should be treated with caution.

A2

Thank you for pointing this out. We acknowledge that the annotation rule in Table S3 is conservative, classifying an error as visual grounding only when the response contains an explicit visual error. This rule can leave unstated fine-grained attribute or interaction-direction errors in later stages. In addition, generated natural language explanations may not fully reflect internal computations on every sample.

In the revised Section 1, Section 3.2, Figure 2, Appendix C, and Table S3, we will discuss this boundary and define the audit of the 975 incorrect responses strictly as an observable response-level error distribution.

However, we want to emphasize that the four-stage text audit is a large-scale descriptive audit of observed response patterns, providing clear research motivation rather than serving as proof on its own. We conducted a comprehensive set of activation measurements and controlled interventions independent of free-form explanations to establish the internal bottlenecks.

Internal evidence for source-relation selection (Section 5)

  • Explicit candidate selection (Section 5.1, Figure 4a and 4b). We do not ask models to generate explanations. Instead, we provide five visually true relations from the same image in text for models to choose from. Even under full context, the accuracy is only 46.3% on Qwen3-VL-2B and 34.9% on InternVL3.5-2B. This shows that relation selection remains a bottleneck after removing unstated visual description differences.
  • Internal score verification under blind review (Appendix A.3, Appendix C.1, Section 5.2, Figure 4c and 4d). The audit labels were finalized before anyone checked internal model measurements. Under this blind review, responses labeled as relation errors also have the lowest gold-relation margin inside the models (q=0.0002q = 0.0002 in Qwen, q=0.0563q = 0.0563 in InternVL).
  • Residual state replacement (Section 5.3, Figure 5). Holding the context fixed, we replace the relation-conditioned state in the residual stream with the gold relation state. Replacing this state from layer 15 onward increases the gold-relation margin, and replacing it across layers 21 to 26 increases the gold-answer margin (q<0.05q < 0.05).

Internal evidence for target-domain mapping (Sections 6 and 8)

  • Semantic state shifts and direction injection (Section 6, Figure 6a to 6c). We start different branches from the same image, question, and gold source relation. Different target implications produce extra residual state movement across layers 19 to 27 (q<0.006q < 0.006), and these states separate by meaning rather than surface wording (q<0.001q < 0.001). Injecting the corresponding direction vector increases the model probability on held-out paraphrases of that implication (q<0.001q < 0.001).
  • Two-way separation in layer-wise tuning (Section 8, Table 2 and Table S13). Tuning relation selection in late layers (layers 15 to 27) degrades accuracy and S6 scores in both model families. In contrast, tuning target mapping in late layers (layers 21 to 27) improves both accuracy and S6 scores in both families.

Direct verification of visual reliance (Section 4)

  • Full-image ablation and attention diagnostics (Section 4, Table 1 and Table S4). Mean-fill and related image ablations show that removing image content lowers accuracy significantly (all q<0.001q < 0.001), while three internal visual attention metrics cannot reliably separate correct and incorrect responses.
  • Patch-level ROI Occlusion. On 35 development examples, masking unrelated background barely affects performance. In contrast, masking key visual regions reduces task margin from 1.66 to 0.53 on Qwen3-VL-2B and from 0.46 to 0.13 on InternVL3.5-2B. This confirms that models physically rely on key local visual cues.

In summary, the text audit provides large-scale failure clues. The internal bottleneck status of relation selection and target mapping is established by the probability, activation, and controlled intervention evidence above.

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

The paper's main claim is that Vision-language models can describe what an image depicts but misinterpret what it means. The authors organize the task into four stages: visual-evidence grounding, source-relation selection, target-domain mapping, and answer binding. An audit of 975 incorrect II-Bench responses assigns 76.1% of the errors to source-relation selection or target-domain mapping. The paper then presents controlled experiments on a 35-example development set. Replacing images with their channel-wise means substantially reduces accuracy, while three attention-based measurements do not reliably distinguish correct from incorrect responses. When models have to select the necessary relation to solve the question from five true relations, they choose the correct relation in less than half of the cases. The authors also use residual-state interventions to investigate if given relation and given implication states affect later preferences, and show that providing an aligned implication improves final answer selection. Finally, scaling and LoRA experiments suggest that relation selection and target-domain mapping may benefit from different training objectives and layer placements.

Critical strengths

1- The paper studies an underexplored problem. Distinguishing between understanding what an image depicts and what it implies is highly relevant for evaluating the reasoning capabilities of vision-language models.

2- The paper also examines model behavior through relation choosing experiments, natural trajectories, internal-state interventions, answer-binding experiments, and adaptation experiments. This provides several complementary insights of the problem.

3- The same-image relation competition experiment is intuitive and useful. Using visually true but question-irrelevant relations as competitors creates a more meaningful diagnostic.

4- The authors control for answer-position bias through cyclic label rotations. Relation candidates occupy every possible label position, and the final answer experiments average over all six answer orders. Reporting both probability margins and discrete top-1 accuracy also makes the behavioral results easier to interpret.

Critical weaknesses

1- The paper relies too much on a four-stage decomposition, while these stages, especially source-relation selection and target-domain mapping, could overlap in many cases. For complex questions, it may be challenging to determine whether an error reflects failure to identify the relevant relation or failure to map that relation to the correct target domain. This overlap could undermine some of the paper’s stage-specific conclusions.

2- In Section 6, the authors try to demonstrate that target-domain mapping is associated with distinct internal states by showing that different implication continuations diverge more from a shared anchor than relation controls do (Figure 6). However, this pattern might be a general property of transformer language models: different sequences naturally produce different hidden states because of lexical or syntactic differences. While the authors use paraphrase and intervention controls, it is still unclear whether this finding is specific to visual implication reasoning.

3- In Section 4, the authors evaluate visual-evidence grounding mainly through raw attention metrics. However, such metrics are often noisy and do not reliably reflect a model's true feature reliance or internal computations. Although the authors also perform image-removal experiments, these do not show which visual features are actually used. Gradient-based analyses or localized patching ablations would provide stronger support for these claims.

4- While broad and strong causal claims are made in the paper, many of the experiments rely on only 35 examples. These small sample sets may limit the generalizability, and make the reported confidence intervals less reliable. It would strengthen the paper to show that the findings hold on larger datasets.

5- The paper relies heavily on GPT-5.5 evaluations in several parts, including failure-stage annotations and relation–implication descriptions. The failure-stage audit uses two independent GPT-5.5 reviewers followed by selective human adjudication, while the authors review the generated relation–implication descriptions. However, the paper does not report an aggregate agreement score, the adjudication rate, or an independently measured LLM annotation error rate. This reliance on LLM-generated annotations could introduce a circular bias into the evaluation.

Questions for authors

1- How do the authors distinguish an error in source-relation selection from an error in target-domain mapping when the relation itself is abstract?

2- In Section 4, the authors use raw attention metrics to dismiss visual grounding failures. However, standard attention weights are unreliable proxies for true feature reliance [1, 2]. Even if a model looks at the right spot, internal feature routing can still fail downstream, something cross-attention stats completely miss. Why weren't standard attribution methods like Integrated Gradients or localized patching ablations used?

3- What was the inter-annotator agreement score (e.g., Cohen's or Fleiss' Kappa) between the two independent GPT-5.5 reviewers before human adjudication?

[1] S. Jain and B. C. Wallace, "Attention is not explanation", NAACL-HLT, 2019.

[2] S. Serrano and N. A. Smith, "Is attention interpretable?", ACL, 2019.

11 comments

Dear Reviewer FaHY:

We thank the reviewer for the thorough and careful feedback.


Q1

Weakness 1 The paper relies heavily on a four-stage breakdown, and these stages, especially source relation selection and target domain mapping, may overlap in many cases. For complex questions, it is hard to tell whether an error comes from failing to identify the relevant relation or failing to map that relation to the right target domain. This overlap may weaken the validity of some stage-level conclusions in the paper.
Question 1 How do the authors separate source relation selection errors from target domain mapping errors when the relation itself is abstract?

A1(page 1)

Thank you for this question. We separate these two stages through clear cognitive boundaries and an ordered annotation rule, and internal attention routing, cross-stage activation features, and layer-wise fine-tuning results consistently confirm that they correspond to distinct internal processes rather than one overlapping stage.

When building the four-stage breakdown, we drew on classic theories from cognitive science and linguistics. Conceptual Metaphor Theory (Lakoff and Johnson, 1980 [1]) and Structure-Mapping Theory for analogical reasoning (Gentner, 1983 [2]) point out that human metaphor understanding usually involves extracting relational structure in a concrete source domain and mapping it systematically to an abstract target domain. Inspired by these cognitive theories, we combined visual perception with final decision making and proposed a four-stage framework for vision-language models. In practice, we strictly follow this four-stage split and apply clear operational boundaries and causal rules.

For abstract relations, we restrict the source relation strictly to the world inside the image. It describes concrete physical actions, spatial layouts, or local interactions between entities in the picture, such as cutting the binding between two people or serving cactus as food. We restrict target mapping strictly to abstract implications beyond the picture world. It must leave specific characters behind and project onto social norms, common psychology, or general principles, such as individual independence within a partnership or absurd service outcomes. Even when interactions in the image carry emotional tone, such as a hostile posture, we still classify them as source relations if they describe local interactions between specific characters. We classify an element under the target domain only when it is summarized as a general principle.

During annotation, we follow a strict ordered annotation rule that records the earliest blocking deviation. We assign an error to the target mapping stage only when the model response already states the correct image source relation and then fails in the later reasoning step. If the model misreads the entity actions or interaction structure in the image from the start, we assign the error directly to the source relation stage. For a very small set of responses whose wording is too vague to judge reliably, we place them in the Unresolved category (3.1%) rather than forcing a label.

A1(page 2)

More importantly, although correct and wrong samples show no clear difference in overall attention patterns, different stages within the same response show systematic differences in attention routing and internal activations. On Qwen3-VL-2B and Qwen3-VL-4B, we compared internal features between the implication segment and the relation segment (implication segment minus relation segment). After Benjamini-Hochberg correction for multiple testing, all metrics reach significance (with a significance threshold of q<0.01q < 0.01, all metrics in the table actually satisfy q<0.001q < 0.001, and none of the 95% confidence intervals cross zero). The exact results are shown below.

Internal Feature Difference MetricQwen 2B Diff [95% CI]Qwen 2B BH qqQwen 4B Diff [95% CI]Qwen 4B BH qqCross-Stage Pattern
Mean image attention (percentage points)−1.92 [−2.20, −1.68]0.000198−1.37 [−1.60, −1.13]0.000198Implication segment attends less to the image
Mean prior-text attention (percentage points)+2.87 [+2.33, +3.41]0.000198+2.24 [+1.67, +2.76]0.000198Implication segment attends more to prior text
Directional similarity of attention output and residual+0.01253 [+0.00987, +0.01523]0.000057+0.00975 [+0.00632, +0.01329]0.000057Attention update aligns more with current state
Effective rank of local state window+0.237 [+0.160, +0.321]0.000057+0.154 [+0.078, +0.244]0.000568Local state changes spread across more directions

From a functional view, the relation stage organizes what happened in the image, while the implication stage summarizes what it means. For example, the relation stage organizes the contrast between specific people (such as people often talking about Depp and Pitt while ignoring Fraser from the same period), and the implication stage draws a general judgment from it (such as collective memory selectively ignoring some people). The metrics above reflect this stage shift in three ways.

  • Attention routing shifts toward generated prior text. After entering the implication segment, the model assigns less attention to the image and more attention to the generated prior text. This shows that the model relies on the relation description already organized earlier to continue summarizing and looks back at the picture less.
  • Attention updates move more along the current state direction. When entering a layer, the attention output and the input residual stream align more closely in direction during the implication segment, meaning updates push further along the current state on average. This may suggest that the model has already formed a clear line of expression from the prior text (such as someone being ignored), and then keeps summarizing along this line rather than turning frequently to search for new entity relations.
  • Local state changes spread across more directions. Within a local window of up to 16 adjacent tokens, the hidden states around the window mean show a higher effective rank during the implication segment, meaning local state changes are not concentrated in just a few directions. This may suggest that when shifting from concrete people to abstract ideas such as public attention, collective memory, and selective neglect, local states vary across more dimensions.

In addition, controlled intervention experiments also support the conclusion that these two stages do not overlap as one single process.

  • Geometric separation in representation space (Figure 6b). When we fix the same source relation anchor, internal states for different target implications still diverge clearly on their own.
  • Asymmetry in layer-wise fine-tuning (Table 2 and Table S13). Fine-tuning on source relations causes a clear performance drop in deep layers, while fine-tuning on target mapping achieves its best gains in deep layers. This opposite layer response indicates that the two processes rely on different internal mechanisms.

We will add these cross-stage internal feature comparisons and boundary details to the revised manuscript.

References
[1] Lakoff, G., and Johnson, M. Metaphors We Live By. University of Chicago Press, 1980.
[2] Gentner, D. Structure-Mapping, A Theoretical Framework for Analogy. Cognitive Science, 7(2), 155 to 170, 1983.

Q2

Weakness 2 In Section 6, the authors try to show that target domain mapping is linked to distinct internal states by showing that different implication continuations deviate more from a shared anchor than relation controls (Figure 6). However, this pattern could be a general property of Transformer language models, where different sequences naturally produce different hidden states due to lexical or syntactic differences. Although the authors use paraphrase and intervention controls, it remains unclear whether this finding is specific to visual implication reasoning.

A2(page 1)

Thank you for this question. Combining our existing text controls with two new visual control experiments, we confirm that these deep hidden states are not surface wording artifacts or pure text effects, but directly depend on the real image and its visual relation to support final multimodal choice. First, in our existing controlled experiments in Section 6, we already ruled out pure surface lexical and syntactic artifacts through three designs.

  • Relation controls in Figure 6a. We introduced relation controls (repeating the selected relation or describing another visually true same-image relation) that also contain unseen words and different sentence structures. Target implications still produce clearly larger shifts than both relation controls. This shows that state changes do not come simply from adding new words, but from moving to an abstract concept level.
  • Cross-paraphrase clustering in Figure 6b. For the same implication, different paraphrases use completely different words and sentence structures, yet they cluster closely together in representation space. The distance between different implications is clearly larger than the distance between different paraphrases of the same implication.
  • Leave-one-paraphrase-out steering in Figure 6c. We evaluated the extracted implication direction vector on unseen test paraphrases. This vector raises the prediction probability specifically for the unseen paraphrase, while random vectors and cross-sample vectors show no such effect. If the model were driven by surface wording, the vector could not generalize in this way.

Second, to verify visual specificity during unguided natural problem solving, we added a paired experiment comparing the original image with a same-size black image.

  • We kept the original question text, the six choices, and all prompt settings identical, gave no hints about the true relation or implication, and replaced only the input image with a same-size black image as the control.
  • We fixed the same text reference set, using 3 human-written paraphrases of the true implication as positive references and the wrong choices from the original question as competing references.
  • We took the hidden states of the implication segment in layers 19 to 27, computed the Margin (mean cosine similarity to positive references minus mean cosine similarity to wrong choices), and ran a paired test between the original image and the black image.

After Benjamini-Hochberg correction, the paired gains reach significance in both model families (q<0.05q < 0.05, and neither 95% confidence interval crosses zero).

ModelPaired ComparisonMargin Diff [95% CI]BH qq
QwenOriginal minus Black+0.01401 [+0.00609, +0.02135]0.02015
InternVLOriginal minus Black+0.01287 [+0.00391, +0.02297]0.02985

Because the question and choices are identical, this difference can come only from the real image. This shows that triggering the right implication depends on visual input, and having the image present makes it easier to form deep states aligned with the true implication.

Comments on the paper

3 comments

Dear Reviewers FaHY, bMrc, and ffpB,

We sincerely thank you for your careful, professional, and helpful feedback. Your questions identified the key evidence needed for studying image implication reasoning and directly guided this revision. We have completed the following changes. New material is shown in blue.

  • We clarified the boundary between source relation selection and target domain mapping.
    This change addresses Reviewer FaHY Q1 and Reviewer bMrc Q2. Appendix C now provides clear definitions, the earliest blocking rule, and the criteria for the Unresolved category. It also discusses the limits of using generated explanations to observe model errors. We further added a cross stage comparison between relation and implication spans. The results show clear differences in attention routing and local state structure.

  • We tested whether implication states depend on real images and specific visual relations.
    This change addresses Reviewer FaHY Q2 and Reviewer bMrc Q1. Appendix G now includes a paired comparison between Original and Black Image conditions. It also includes controlled comparisons among Original, Content Edit, and Appearance Control images. Real images move late response states closer to the correct implication. Changing a key visual relation moves the states toward the edited implication.

  • We added a local occlusion experiment on key visual regions.
    This change addresses Reviewer FaHY Q3. Appendix D.2 compares the original image, key ROI occlusion, and equal area background occlusion. We used 10 random background locations for each example, giving 350 background occlusion runs in total. Key ROI occlusion reduces task accuracy, answer margin, and relation selection accuracy. Background occlusion has little effect on task accuracy.

  • We explained the complete source of the 35 examples and clarified the work supporting the validity of our conclusions.
    This change addresses Reviewer FaHY Q4 and Reviewer ffpB Q1. The controlled experiments use all 35 examples in the official II-Bench development set. We applied no additional selection or filtering. We also clarified the use of label rotations, paired designs, full test set image ablations, different model families and scales, three random seeds, and direct transfer to CII-Bench.

  • We reported annotation agreement and the scale of human review.
    This change addresses Reviewer FaHY Q5 and Reviewer bMrc Q3. The two independent annotation runs reached 79.4 percent raw agreement before human review, with a Cohen’s κ\kappa of 0.72. Human review covered 201 records, or 20.6 percent of the audit. The final audit retained 30 responses as Unresolved. About 68 percent of disagreements occurred between source relation selection and target domain mapping.

  • We added complete input ablation results and explained the different roles of images and answer options.
    This change addresses Reviewer ffpB Q2 and Q3. Appendix D.1 now reports absolute accuracy under Original, Mean Fill, Black Image, No Image, and chance conditions. Section 5.1 reports the exact relation selection accuracy under all five context conditions. The discussion explains that answer options provide top down filtering from the target meaning. Images support several true relations from the same scene and therefore create visual competition.

The questions and suggestions reflect deep research experience. We greatly value your professional and helpful feedback. We also sincerely appreciate the time and effort that you devoted to reviewing our work.

If these changes have addressed your main concerns, we sincerely ask you to reconsider your evaluation and consider raising your score. Your recognition and support mean a great deal to us and to this work. We remain happy to answer any further questions.

Thank you again for your time, patience, and expert guidance.

Best regards,
The Authors

Dear kind Reviewers FaHY, bMrc, and ffpB,

Thank you once again for all the time, care, and hard work you put into reviewing our paper during your busy schedules. We truly hope one of you might kindly break the silence and kick off the discussion with us. If our answers and additional experiments have resolved your concerns, we would be deeply grateful if you could be the brave first voice to step forward and raise your score for our work. Your support means the world to us, and we are always here and happy to answer any further questions.

Warmest wishes,
The Authors

Dear AC and Reviewers,

We would like to express our deepest gratitude to the AC for your tremendous effort, dedication, and thoughtful oversight throughout this demanding review cycle, and we sincerely thank all reviewers for their time and constructive feedback. As the discussion period concludes, we provide a brief summary of our paper's contributions, the strengths noted by the reviewers, and the improvements made during the rebuttal.

Summary of Contributions.

  • Diagnostic framework and large-scale audit. We introduce a four-stage diagnostic framework covering visual-evidence grounding, source-relation selection, target-domain mapping, and answer binding, and audit 975 incorrect II-Bench responses to show that 76.1% of observable failures occur in relation selection or target mapping.
  • Mechanistic bottleneck localization. Through controlled diagnostics and causal interventions, we show that visual access alone does not distinguish correct from incorrect interpretations, whereas competition among true same-image relations is a central bottleneck, with models ranking the gold relation first in only 46.3% and 34.9% of full-context cases, and that relation- and implication-conditioned internal states causally steer downstream preferences.
  • Actionable full-scale training validation. Across 1,371 II-Bench and 765 CII-Bench test examples, we demonstrate that late-layer target-mapping LoRA in layers 21 to 27 improves accuracy and six-order robustness while relation LoRA degrades performance.

Strengths Recognized by the Reviewers. Across the initial reviews, the reviewers highlighted the novelty, relevance, and rigor of our study. In particular, they noted that our work investigates an underexplored and interesting problem—distinguishing between what an image depicts and what it implies—and found the distinction between selecting a concrete relation and interpreting its meaning useful. They also praised our thoughtful and rigorous experimental design, including the intuitive use of visually true same-image competitors, cyclic label rotations to control for position bias, held-out paraphrases with random and cross-sample intervention controls, complementary behavioral and internal-state analyses, and clear, well-presented figures and tables.

Summary of Rebuttal and Discussion. During the rebuttal, we conducted extensive additional experiments and thoroughly revised the manuscript, as highlighted in blue in the updated PDF. Specifically, we added cross-stage internal feature comparisons in Table S4, a 350-run patch-level ROI occlusion study in Tables S6 and S7, complete absolute-accuracy bounds across Original, Mean-Fill, Black-Image, and No-Image conditions in Table S5, natural-inference Black-Image and controlled Image-Editing experiments in Appendix G, and full inter-annotator agreement statistics with 79.4% raw agreement and Cohen's kappa of 0.72 in Appendix C.2.

These additions largely addressed the reviewers' questions, and the reviewers acknowledged or implicitly accepted the effectiveness of our new experiments in strengthening the manuscript. We would also like to respectfully note for the AC's attention that Reviewer FaHY responded only on the final day of the discussion period and overlooked several key insights in our rebuttal, namely that our bottleneck claim is supported not solely by state interventions or free-form audits, but also by natural-trajectory internal margins in Figure 4c and 4d, unguided natural visual-specificity tests in Appendix G, explicit relation competition bypassing free-form generation in Section 5.1, and full-scale layer-wise LoRA validation in Section 8. We have posted a concise follow-up clarifying these points.

We kindly ask the AC and the reviewers to consider the full body of evidence and the substantial improvements in the revised manuscript, and to render a fair and holistic final decision. We deeply appreciate your invaluable service to the community and wish the AC and all reviewers the very best in your ongoing research and endeavors.

Warm 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.