acceptodds
Under review as a conference paper at ICLR 2027

Models See More Than They Say: The Answer Bottleneck in Multimodal Language Models

Abstract

Multimodal language models can encode useful visual information and still answer incorrectly. We investigate this gap through the Guided Answer Space (GAS), an explicit set of candidate answers specifying the alternatives the model must distinguish. By varying this space while keeping the image, question, and target answer fixed, and intervening on intermediate representations, we provide evidence for an Answer Bottleneck in determining correct answers from encoded visual evidence. Crucially, intermediate visual representations already provide substantial support for accurate open-ended answers: they help recover accuracy lost to image corruption and retain much of their contribution even when their states are held fixed in later layers. Across eleven models on TextVQA and GQA, two-choice selection depends on visual tokens through fewer layers than open-ended answering, with a median difference equivalent to 42% of model depth; expanding the GAS extends this dependence to later layers. Building on these findings, we develop a two-stage inference procedure that constructs a GAS from other models' answers and then selects the final answer from it, without ground-truth access during either stage. Using three of the models, this procedure corrects 23.7% of original open-ended errors and improves overall accuracy by 3.56 percentage points. Our findings show that answer requirements shape the decoder's reliance on visual evidence and that the GAS can improve accuracy through two-stage inference.

Primary area:
interpretability and explainable AI
Keywords:
multimodal large language models, interpretability

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_CQR3OpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how answer requirements affect multimodal language models’ use of visual representations. It introduces the Guided Answer Space (GAS), an explicitly supplied set of candidate answers, and measures visual dependency depth through layer-wise interventions on visual token states. Across eleven models, converting TextVQA and GQA questions into two-choice selection reduces this depth by a median of 42% of model depth. Controlled experiments further show that increasing candidate count extends visual dependence under both option-label and answer-text outputs. Transfer and Freeze interventions demonstrate that intermediate visual states contribute substantially to open-ended answering, including when their values are held fixed in subsequent layers. Finally, a two-stage procedure constructs candidates from other models and selects an answer without ground-truth access. Across three target models and two datasets, it improves pooled accuracy by 3.56 percentage points and exceeds frequency voting on identical candidate pools by 2.56 points.

Critical strengths
  1. A broad and carefully controlled investigation of visual dependence. The central comparison holds images, questions, and target answers fixed across answering settings, and the depth ordering holds across all eleven models on both TextVQA and GQA. The harder-distractor control, alternative state replacements, unchanged-prompt likelihood analysis, and candidate-count-by-output-format experiment address several obvious alternative explanations (Sections 3–5; Appendices A–D). This makes the contribution more substantial than a behavioral observation that multiple-choice questions are easier than open-ended questions.

  2. Complementary evidence connecting internal representations to practical answering. Transfer and Freeze test different properties of intermediate visual states: their causal contribution to answering and their usefulness without subsequent state updates. The practical experiment also separates candidate generation from selection by comparing methods on identical pools. Reporting reverse errors, example-clustered uncertainty estimates, and image-replacement controls during selection strengthens the evidence for a real, if modest, inference benefit (Section 4; Appendix E.1). These analyses give the paper both scientific and practical value.

Critical weaknesses
  1. W1: The interventions do not directly establish the cause of the model’s original errors. Transfer measures recovery from artificial image corruption, while Freeze measures how much performance survives when intermediate visual states stop updating. These experiments establish that the states contain useful information, but their aggregate results do not show that the information needed to answer originally misanswered examples was already available and subsequently misused. Moreover, freezing visual states still permits substantial downstream computation through text states. The high noise-image selection accuracies in Table 8, particularly on GQA, also show that candidate-conditioned language cues contribute considerably. The harder-distractor and unchanged-prompt controls help, but do not fully connect the representation interventions to the original failure cases. This limits the strength of the Answer Bottleneck interpretation, although it does not invalidate the observed dependency-depth differences.

  2. W2: The answer-scoring protocol leaves the accuracy claims insufficiently validated. Appendix A.1 accepts a response whenever it contains the reference answer as a substring, and Appendix E.1 uses this rule for the practical inference results. This can accept incorrect numeric strings or responses containing incompatible answers; for example, a reference of “2” would match “12.” Consistent use of the rule does not ensure equal bias across free generation, candidate likelihood, voting, and selection, because these procedures can produce different response distributions. Appendix D.1 provides scoring sensitivity for the candidate-count experiment, but does not establish robustness of the headline practical gains. Rescoring with benchmark-standard metrics and auditing disagreements would materially strengthen confidence in the reported 3.56-point improvement. TextVQA’s benchmark evaluation follows the VQA accuracy protocol.

Questions for authors
  1. Clarification and supporting analysis for W1: Can you report the intervention results separately for examples answered correctly and incorrectly under the original clean open-ended prompt? What evidence specifically connects useful intermediate representations to the examples whose original errors are corrected by GAS, rather than demonstrating preservation of already successful answering?

  2. Clarification and robustness analysis for W1: On the same examples and with harder distractors, how much of the selection-versus-generation depth difference remains when evaluation is restricted to questions that require the image in both settings? Table 9 provides a related answer-text analysis, but not this paired comparison.

  3. Validation of W2: Do the practical gains over original generation and frequency voting persist under benchmark-standard scoring? Please report the rescored gains and paired confidence intervals, together with the frequency of disagreements with substring matching.

4 comments
AuthorsOpenReview ↗

Comment 1 of 2

Thank you for these constructive questions. Intervals are paired 95% CIs that resample questions, and pp denotes percentage points. Several analyses intervene at ℓ_sel, the selection dependency depth of each model and task (Section 3.2): in two-choice selection, zeroing visual token states from this layer onward leaves accuracy nearly unchanged.

[Q1] Do the interventions help originally incorrect answers, and what links them to the errors that a GAS corrects?

Sections 4.1–4.2 pool all questions, so, as the review notes, their gains could come mainly from questions that the model already answers correctly. To examine the original errors themselves, we restrict two analyses to the questions that each model originally answered incorrectly.

1. On original errors, the interventions still produce correct answers (new; Section 4.1, Appendix C.3, Table 15). If the Clean runs on these questions encoded evidence for the correct answer, supplying their intermediate visual states should produce correct answers that the original run did not give. Transfer (Section 4.1) runs the model on a noise image and, at ℓ_sel, replaces its visual token states with those of the clean-image run; the gain is measured over the noise-image run. Freeze (Section 4.2) holds the visual token states fixed from ℓ_sel onward; the gain is measured over zeroing them. On TextVQA, where answers must be read from the image, results pooled over eleven models are (n counts model–question cases):

InterventionOriginal errors (score 0): n, gainOriginally correct (score 1): n, gain
Transfer714, +8.28 [6.11, 10.58]2,122, +68.73 [64.93, 72.33]
Freeze783, +2.64 [0.30, 4.97]2,123, +49.64 [45.74, 53.42]

Because the clean run scored 0 on the original errors, every correct answer in the left column is one that the original run did not give, so these gains cannot come from preserving successful answers. Both interventions yield such gains, Transfer in all eleven models; these Clean runs had thus already encoded evidence for the correct answer at ℓ_sel.

2. When a GAS corrects these errors, it uses this intermediate evidence (new; Appendix C.3, Table 16). When the correct answer is offered as one of two options, selection corrects a median of 91% of open-ended errors across eleven models (Appendix E.2, Table 33). As the review notes for Table 8, such corrections could also rely on language cues from the candidates. To separate the two sources, we let the model select on the original errors while limiting its image evidence to the intermediate states of the original run, and we compare the result with images that carry no relevant evidence.

The model chooses between the reference answer and a distractor of the same answer type. Because the image precedes the question and options, causal attention keeps its visual token states identical to those of the original open-ended run. We zero them from ℓ_sel onward, leaving these intermediate states as the only image evidence. We then repeat the choice with the image replaced by a uniform gray image, which leaves only the question and candidates, or by the image of another TextVQA question, a natural image that does not answer the question. Each question is run with the reference as option A and as option B and counts as correct only if both runs choose it. Results on the original errors among 300 TextVQA questions for which our predefined answer-type rules provide a distractor:

TextVQA modelnReal image, %Real − gray, ppReal − other image, pp
LLaVA2974.1+21.7 [7.2, 37.2]+27.2 [12.4, 43.5]
Qwen1766.5+34.7 [14.7, 56.5]+36.5 [17.1, 58.2]
InternVL2780.4+30.0 [14.8, 46.7]+22.2 [9.3, 37.4]

The real image exceeds both replacements in all three models. When a GAS corrects these errors, it therefore draws on the intermediate visual evidence of the original Clean run, beyond what the question and candidates support on their own.

AuthorsOpenReview ↗

Comment 2 of 2

[Q2] Does the depth difference hold on the same questions when both settings require the image?

Because the question and options alone can suffice for some selections (Table 8), we follow your suggestion and compare both settings on the same questions, keeping only those that neither setting can answer without the image.

1. Questions that require the image in both settings. We keep TextVQA questions that both settings answer correctly with the real image and incorrectly when the image is replaced by a uniform gray image. Selection uses a distractor of the same answer type, and with the gray image the model must choose this distractor in both option orders. These distractors are thus plausible without the image, and the image makes both settings correct.

2. Depth on these questions. d is the sustained-crossing depth (Appendix A.2), which requires accuracy to stay within 5 points of the unmodified run at that layer and every later tested layer; we divide d by the number of decoder layers L (new; Section 3.2, Appendix B.2, Table 9):

TextVQA modelnSelection d/LOpen-ended d/LGap (fraction of L)
LLaVA240.440.940.50 [0.28, 0.58]
Qwen240.570.930.36 [0.07, 0.39]
InternVL190.570.960.39 [0.29, 0.39]

On the same image-dependent questions, selection depends on visual tokens through 0.36L–0.50L fewer layers than open-ended answering, consistent with the 0.42L median in Section 3.2.

3. The result does not depend on the gray image or the 5-point threshold. Because a gray image differs from natural images, it could mark some questions as image-dependent for the wrong reason. To rule this out, we also identify image-dependent questions with the image of another TextVQA question, a natural image that does not answer the question. Selection remains shallower when questions are identified with this image, when both images must lead to failure, and when the allowed drop is 2 or 10 points instead of 5.

[Q3] Practical gains under benchmark-standard scoring

1. Rescoring. As the review notes, substring matching can accept incorrect answers, and its bias may differ across methods. To test whether the gains depend on this rule, we rescored all saved outputs with the standard ten-annotator VQA accuracy for TextVQA and exact match for GQA, without rerunning any model (new; Section 5.3, Appendix E.1).

2. Gains. The gain persists for the lowest-cost configuration in Table 30, in which each of the other two models generates one answer and the target model selects between them. Accuracy increases from 73.52% to 76.01%.

Direct selection vs. (1,800 predictions)Gain, pp
Direct generation+2.49 [1.37, 3.62]
Frequency voting over identical candidates+2.73 [1.29, 4.23]
Self-consistency over 8 target samples+2.54 [1.29, 3.77]

3. The four-answer configuration benefits from other models' answers. The +3.56 pp in the submitted abstract comes from the configuration in which each of the other two models generates four answers. Under the standard metrics, frequency voting over these eight answers raises accuracy by 1.25 pp [0.08, 2.42], whereas voting over eight samples from the target leaves it unchanged (−0.05 pp [−0.48, 0.37]). There, answers that recur across samples indicate the common answer; with one answer per model, no answer recurs, and direct selection supplies the gain. In both configurations, the gain comes from correct answers that other models generate and the target does not, as the Answer Bottleneck predicts.

4. Selection uses the image. The image-replacement control noted in the review (Appendix E.1), which uses four answers per model, is reproduced under the standard metrics: replacing the selector's image with a gray image or an image from another question reduces accuracy by 3.51 pp [1.58, 5.41] and 4.17 pp [2.16, 6.19], respectively.

5. Disagreement with substring matching. Standard and substring scores disagree on a similar fraction of predictions for every method, 14.2–14.9%, owing to additional references, partial credit and substring false positives. The rescoring thus affects all methods similarly.

Revision (in red). The new analyses appear in Section 4.1 and Appendix C.3 (Q1), Section 3.2 and Appendix B.2 (Q2), and Appendices A.1 and E.1 (Q3); the abstract, Introduction, Section 5.3, Figure 7 and Table 1 report the configuration with one answer per model.

Reviewer_CQR3OpenReview ↗

Response to Authors

Thank you for the detailed response and the additional experiments. The analyses on originally incorrect examples provide a clearer connection between intermediate visual evidence and the errors corrected by GAS. I also appreciate the paired comparison on image-dependent questions and the benchmark-standard rescoring, which address my concerns about the depth comparison and the robustness of the practical gains.

These additions resolve my main concerns and strengthen the paper’s conclusions. I will maintain my rating and increase my confidence from 2 to 3.

Thanks a lot for the follow-up! Your suggestions really helped strengthen the paper.

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

The paper investigates the “Answer Bottleneck” in multimodal language models, where useful visual information is encoded but does not lead to a correct answer. The authors introduce the Guided Answer Space (GAS), an explicit set of candidate answers, and examine how changing this space affects the model’s dependence on visual-token representations. Across eleven models, converting open-ended questions into two-choice selection consistently reduces visual-token dependency depth, with a median difference equivalent to approximately 42% of model depth. Increasing the number of candidates generally extends this dependence to later layers. Through visual-state transfer and freezing experiments, the authors further show that intermediate visual representations provide substantial support for open-ended answering even without continued updates to those representations. This can be interpreted as evidence that answering performance depends on how effectively the decoder uses existing visual evidence. In addition, the authors also propose a two-stage inference procedure that constructs candidate sets from other models’ answers and selects the final answer without ground-truth access, improving pooled accuracy from 65.33% to 68.89% and correcting 23.7% of original errors.

Critical strengths
  1. The paper studies an important distinction between visual representation and answer production. Incorrect answers do not necessarily imply that the visual encoder failed to capture relevant evidence. The decoder must also identify, compare, and express the appropriate answer. Investigating these stages separately is useful for understanding multimodal benchmark performance and deciding whether improvements should target visual representations, their use by the decoder, or the answering procedure.
  2. The study covers a reasonably broad range of models and includes substantial robustness analysis. Eleven models from five families are evaluated, with different model sizes and decoder depths. The appendix examines threshold sensitivity, alternative replacement states, harder distractors, document questions, answer-text outputs, verification, and candidate likelihoods under the unchanged open-ended prompt. These checks make the broad dependence result more reliable.
  3. The candidate-count experiment separates candidate availability from response format. The factorial comparison varies the number of candidates while separately requiring option-label or answer-text output. Candidate sets are nested, and their order is kept fixed across output formats. The paired stable-depth analysis reports positive two-to-eight-candidate changes in all twelve model-task-format combinations, where the confidence intervals is above zero. This shows that the findings are not explained entirely by generating a short option label instead of an answer string.
Critical weaknesses
  1. The central mechanistic interpretation remains insufficiently isolated. Transfer shows that intermediate visual states help restore performance after image corruption. However, this recovery could mainly restore examples the clean model already answered correctly. It does not establish that the correct answer was available in the intermediate representations on the examples the clean model originally answered incorrectly.
  2. Providing candidates changes the information available to the model, not just the number of outputs it must consider. Keeping the image, question, and target answer fixed is a strong control, but the full input changes when the correct answer and distractor are provided. For example, supplying “5.00” and “5.20” tells the model exactly which visual distinction to inspect. It may enable targeted comparison or partial matching that the original prompt did not do. Selection can therefore succeed through a different evidence-extraction strategy. 3.Candidate count is confounded with prompt length and distractor content. Expanding from two to eight candidates adds text and potentially harder semantic distinctions. Controlling output format does not separate these factors. Length-matched comparisons and controlled distractor similarity would clarify whether the depth increase specifically reflects the number of answer alternatives.
  3. The entropy analysis combines different notions of answer diversity. For multiple-choice benchmarks, entropy is calculated over option labels; for open-ended benchmarks, it is calculated over lexical answers. These quantities are not directly comparable measures of the underlying visual discrimination problem.
Questions for authors
  1. Is it possible to directly examine originally incorrect clean-model predictions? For this subset, is there any evidence to show that intermediate representations computed under the unchanged open-ended prompt contain sufficient information to distinguish the correct answer from the model’s actual incorrect answer?
  2. How to distinguish improved use of already sufficient evidence from candidates prompting a different, more targeted visual computation?
  3. Does the headline dependency-depth gap still exist under stronger image-dependence controls? It would be more convincing to report paired open-ended and option-label results on the same image-dependent examples, ideally using plausible distractors with matching answer types and low no-image selection accuracy.
  4. How much of the two-to-eight-candidate effect remains after controlling prompt length and distractor similarity?
7 comments
AuthorsOpenReview ↗

Comment 1 of 3

Thank you for these precise questions. All four new analyses support the Answer Bottleneck. Intervals are paired 95% CIs, and pp denotes percentage points. Several analyses intervene at ℓ_sel, the selection dependency depth of each model and task (Section 3.2): in two-choice selection, zeroing visual token states from this layer onward leaves accuracy nearly unchanged.

[Q1] Do the representations of originally incorrect runs contain the evidence needed for the correct answer?

We study the TextVQA questions that each model originally answered incorrectly (VQA accuracy 0). The question has two parts: whether the Clean runs on these questions encoded evidence for the correct answer at all, and whether this evidence suffices to distinguish the correct answer from an incorrect one. We answer them in turn.

1. Clean runs that answer incorrectly already encode evidence for the correct answer. If such a Clean run encoded this evidence, supplying its intermediate visual states to a run that otherwise receives no image information should produce correct answers. Transfer (Section 4.1) performs this test. We run the model on a Gaussian-noise image with the unchanged open-ended prompt and, at ℓ_sel, replace its visual token states with those of the Clean run. Across 714 original errors from eleven models, accuracy rises by 8.28 pp [6.11, 10.58] over the noise-image run, with gains in every model (new; Section 4.1, Appendix C.3, Table 15). Because the clean run scored 0 on these questions, these correct answers come from evidence that the Clean run encoded but did not turn into its answer.

2. This evidence distinguishes the correct answer from a plausible incorrect one. Free generation does not reveal which answers the states can tell apart. We therefore present the reference answer together with an incorrect answer of the same answer type and ask the model to choose.

For this choice to rely on the states of the original open-ended run, the model must not obtain new image evidence once it sees the options. Two properties ensure this. First, in all three models the image precedes the question and options, and causal attention prevents visual token states from attending to later text, which keeps these states identical to those of the original open-ended run. Second, we zero them from ℓ_sel onward (Section 2.2), leaving the open-ended run's states below ℓ_sel as the only image evidence.

The model may also choose correctly from the text alone, for example when one option fits the question better. To measure what the image contributes, we repeat the choice with the image replaced by a uniform gray image, which carries no image content, or by the image of another TextVQA question, a natural image that does not answer the question. Each question is run with the reference as option A and as option B and counts as correct only if both runs choose it, so a preference for one option position cannot produce a correct score. Results on the original errors among 300 TextVQA questions for which our predefined answer-type rules provide a distractor (new; Appendix C.3, Table 16):

TextVQA modelnReal image, %Gray, %Real − gray, ppReal − other image, pp
LLaVA2974.152.4+21.7 [7.2, 37.2]+27.2 [12.4, 43.5]
Qwen1766.531.8+34.7 [14.7, 56.5]+36.5 [17.1, 58.2]
InternVL2780.450.4+30.0 [14.8, 46.7]+22.2 [9.3, 37.4]

Even on the model's own errors, the visual token states computed under the unchanged open-ended prompt thus distinguish the correct answer from a plausible alternative. This is the pattern that the Answer Bottleneck predicts: the evidence is encoded, and the error arises when the decoder determines the answer.

AuthorsOpenReview ↗

Comment 2 of 3

[Q2] Better use of existing evidence, or a new, more targeted visual computation?

The two accounts differ in where the evidence that separates the candidates comes from. In the review's example, the candidates "5.00" and "5.20" direct the model to one digit. If they prompt a new visual computation, the model extracts evidence about this digit only because the candidates ask for it. If the GAS instead improves the use of existing evidence, this evidence is already present in what the model computes from the image without candidates. Three observations support the second account.

1. Candidates cannot change how the image is encoded. In all three models, the image precedes the question and candidates, and causal attention prevents visual token states from attending to later text. The visual token states are therefore identical with or without candidates. The candidates can influence only the later text positions, which read these states through attention at every layer. Consequently, a targeted comparison would have to arise from text positions that, guided by the candidates, read details the open-ended run does not read. The two tests below examine whether the evidence that separates the candidates is available without such guidance.

2. Without candidates in the input, the model already prefers the correct answer (Appendix C.5). We give the model the unchanged open-ended prompt and compute, for each candidate separately, the probability that the model produces it as the answer. Because the candidates never appear in the input, they cannot guide what the model reads from the image. On the questions that each model originally answered incorrectly, the correct answer receives a higher probability than the distractor in 87–94% of cases, compared with 50% for chance (three models, two tasks; Table 19).

3. With candidates in the input, selection follows evidence computed without them (new; Appendix C.6, Table 20). Selection lists the candidates, so we also test it directly. The test must show the candidates in the prompt while preventing them from guiding how the model reads the real image.

We use 18 pairs of TextVQA questions that share the question text but have different images, A and B, and different reference answers, a and b. If an image's visual states carry evidence for its answer, inserting A's states should favor a, and inserting B's states should favor b. For each image, we record its visual token states at ℓ_sel under the original open-ended prompt, without candidates. We then build one fixed selection input: a gray image, the shared question, and a and b as options. As in Transfer (Section 4.1), we replace its visual token states at ℓ_sel with those recorded from A, or from B.

Below ℓ_sel, the question and candidates meet only the gray image, which prevents them from guiding what is read from the real image, and the two runs differ only in the inserted states. In each run, we measure the preference for a as the logit of a's option label minus that of b's, averaged over both option orders so that a preference for one option position cancels.

TextVQA modelPairs in which A's states yield a stronger preference for a than B's states
LLaVA18/18
Qwen18/18
InternVL17/18

The mean difference in preference is positive for all three models (CIs exclude zero).

Conclusion. In both tests, the evidence that separates the candidates is present in states that the model computed without them. A comparison prompted by "5.00" and "5.20" thus operates on existing evidence, and the GAS changes how the decoder uses this evidence to determine the answer, which is the Answer Bottleneck.

AuthorsOpenReview ↗

Comment 3 of 3

[Q3] Does the depth gap hold on the same questions when both settings require the image?

Section 3.2 compares the two settings on all questions. On some of them, the question and options alone may suffice for selection, as the noise-image accuracies in Table 8 indicate. The shallower selection depth could then reflect a smaller need for the image rather than a different use of visual evidence. Following your suggestion, we therefore compare both settings on the same questions and keep only those that neither setting can answer without the image.

1. Questions that require the image in both settings. We keep TextVQA questions that both settings answer correctly with the real image and incorrectly when the image is replaced by a uniform gray image. Selection uses a distractor of the same answer type, and with the gray image the model must choose this distractor in both option orders. The distractors are thus plausible enough to be chosen without the image, and the image is what makes both settings correct.

2. Depth on these questions. d is the visual token dependency depth of Section 2.2: the earliest tested layer from which visual token states can be zeroed while accuracy stays within 5 points of the unmodified run, required here at that layer and at every later tested layer. We divide d by the number of decoder layers L (new; Section 3.2, Appendix B.2, Table 9):

TextVQA modelnSelection d/LOpen-ended d/LGap (fraction of L)
LLaVA240.440.940.50 [0.28, 0.58]
Qwen240.570.930.36 [0.07, 0.39]
InternVL190.570.960.39 [0.29, 0.39]

On the same image-dependent questions, selection depends on visual tokens through 0.36L–0.50L fewer layers than open-ended answering, consistent with the 0.42L median in Section 3.2.

3. The result does not depend on the gray image or the 5-point threshold. Because a gray image differs from natural images, it could mark some questions as image-dependent for the wrong reason. To rule this out, we also identify image-dependent questions with the image of another TextVQA question, a natural image that does not answer the question. Selection remains shallower when questions are identified with this image, when both images must lead to failure, and when the allowed drop is 2 or 10 points instead of 5.

[Q4] Prompt length and distractor similarity

Expanding from 2 to 8 candidates adds text as well as alternatives, and the added alternatives may be harder to distinguish from the correct answer. Either change alone could deepen visual dependence, so we hold each fixed in turn.

1. The same prompt for every set size. Appendix C.5 does not list the candidates in the prompt. It keeps the unchanged open-ended prompt and scores each candidate by the probability that the model produces it as the answer, so the prompt is identical for sets of 2, 4 and 8, and any change in depth across set sizes reflects the candidates alone (Table 19).

2. Added distractors no closer to the correct answer (new; Section 5.1, Appendix D.1, Table 25). In this setting, we further keep only added distractors that are no more similar to the correct answer than the original distractor, measured by sentence-embedding similarity. We then compare each question's stable depth (Section 5.1), here the earliest layer from which zeroing visual token states at that or any later tested layer still ranks the correct answer first. From 2 to 8 candidates, it increases by 0.11L–0.26L in all six model–task pairs, with all CIs above zero.

3. Equal length when candidates are listed (new; Appendix D.1). In selection, listing 8 candidates lengthens the prompt. If length drove the effect, padding the 2-candidate prompt to the same length should deepen its dependence. With two padding schemes that reach the exact token length of the 8-candidate prompt, the 2-candidate depth changes by at most two layers in all six model–task pairs. Hence, prompt length alone does not deepen visual dependence.

Taken together, visual dependence grows with the number of alternatives the model must distinguish.

[W4] Section 5.2 now defines entropy as the diversity of each task's answers, computed over option labels for benchmarks with fixed answer choices and over answer strings for open-ended benchmarks.

Revision (in red). Section 4.1 and Appendix C.3 add the analyses of original errors (Q1). Appendix C.6 adds selection with visual states computed without candidates (Q2). Section 3.2 and Appendix B.2 add the depth comparison on questions that require the image in both settings (Q3). Appendix D.1 adds the controls for prompt length and distractor similarity, and Section 5.1 points to them (Q4). Section 5.2 revises the definition of entropy (W4).

Reviewer_d7XmOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates the discrepancy between visual information encoded in MLLMs and their final answers, proposing the Answer Bottleneck hypothesis. It introduces the Guided Answer Space (GAS) to study how answer requirements affect the decoder's dependence on visual representations. Through layer-wise visual token removal, transfer, and freezing experiments, the authors show that intermediate visual representations can substantially support open-ended answering, while answer selection requires access to visual tokens through fewer layers. Finally, a two-stage inference procedure uses answers generated by other models as candidates and improves overall accuracy by 3.56 percentage points.

Critical strengths

The paper investigates an important distinction between visual information encoding and answer determination. Understanding why MLLMs fail to produce correct answers despite having useful visual representations is relevant to both model interpretability and inference-time improvement.

Critical weaknesses

I think the proposed two-stage inference procedure offers a relatively modest practical contribution. Generating candidate answers from other models and selecting among them follows an established generate-then-select paradigm. Although the reported 3.56 percentage-point improvement is encouraging, it requires additional generations from multiple models, without a comprehensive comparison against inference-budget-matched ensemble or sampling baselines. Moreover, over half of the candidate pools contain only one distinct answer after deduplication, limiting the role of actual answer selection in many cases.

Questions for authors

None.

3 comments

Thank you for the careful review and for finding the reported improvement encouraging.

[C1] Contribution beyond generate-then-select

The primary contribution of the paper is mechanistic. It identifies how effectively the decoder uses encoded visual evidence as a limiting factor in multimodal answering, which we term the Answer Bottleneck. With the image, question and target answer held fixed, supplying a guided answer space (GAS) reduces visual token dependency depth by a median of 42% of model depth across eleven models (Section 3.2, Table 6). Moreover, intermediate visual representations computed under the original open-ended prompt already provide substantial causal support for correct answers (Section 4).

Generate-then-select is an established inference pattern (Fu et al., 2023; Section 6), and we use it to test a prediction of the Answer Bottleneck. The paper shows that models often recognize correct answers that they fail to generate: when the correct answer is offered as one of two options, selection corrects a median of 91% of open-ended errors across eleven models (Appendix E.2, Table 33). If the limitation lies in generating the answer rather than in recognizing it, additional samples from the same model should help little, because they come from the same generation process that produced the error. Answers generated by other models, in contrast, can contain the missing correct answer, which the target can then recognize. The comparisons below test this prediction.

[C2] Comparison with sampling and ensemble baselines

The review asks whether the gain holds against ensemble and sampling baselines with matched inference budgets. We therefore compare these baselines with the lowest-cost configuration in Table 30, in which each of the other two models generates one greedy answer and the target model (LLaVA-OV-1.5-8B, Qwen2.5-VL-7B or InternVL3-8B) selects between them conditioned on the image. All methods are evaluated on the same 1,800 predictions (300 TextVQA and 300 GQA questions per target) and scored with the standard VQA accuracy for TextVQA and exact match for GQA. Gains are reported in percentage points (pp) with paired 95% CIs (new; Section 5.3, Appendix E.1):

MethodCompute relative to oursGain over the target's open-ended answer
Ours: one answer from each other model + direct selection1×+2.49 [1.37, 3.62]
Ensemble: frequency voting over the same two answersslightly lower−0.24 [−1.36, 0.83]
Sampling: self-consistency over 8 target samplesmore than 2×−0.05 [−0.48, 0.37]

Direct selection thus outperforms voting over identical candidates by 2.73 pp [1.29, 4.23]. It also outperforms self-consistency by 2.54 pp [1.29, 3.77], although self-consistency requires more than twice the computation.

This pattern agrees with the prediction. Eight samples from the target mostly repeat its greedy answer, so voting over them leaves accuracy unchanged; Figure 7c likewise shows the low coverage of the target's own candidates. Answers from other models, by contrast, often contain a correct answer that the target did not generate.

[C3] The role of selection when proposed answers coincide

The 54% of single-answer pools noted in the review arises in the submitted paper's configuration, in which each other model generates four answers (Appendix E.1). With one answer from each model, the two answers coincide more often, on 72% of questions, consistent with the average of 1.28 distinct answers in Table 30. To examine the role of selection, we split the same 1,800 predictions by whether the two proposals coincide (new):

  • The two answers coincide (1,296 predictions). The common answer is returned without a selection step. Accuracy rises from 80.4% for the target's own answer to 85.3% (+4.84 pp [3.17, 6.55]).
  • The two answers differ (504 predictions). Frequency voting faces a one-to-one tie and falls back to the first-listed answer, whereas direct selection chooses with the image and is 9.76 pp [4.58, 15.06] more accurate. Because the two methods coincide on single-answer pools, these cases account for the entire 2.73 pp advantage over voting.

Agreement between the other models therefore supplies reliable answers at no selection cost, and selection resolves their conflicts more accurately than voting.

Revision (in red). The abstract, Introduction, Section 5.3, Figure 7 and Table 1 now present the configuration with one answer per model as the main practical result (C2–C3). Appendix E.1 reports all configurations under benchmark-standard scoring, together with the self-consistency baseline (Table 30).

Thank you for the detailed response and additional experiments. I think the authors have largely addressed my first two concerns . While the role of selection remains somewhat unclear, I do not consider this a major issue given the paper's primary focus. I will maintain my score of 3.

Thank you for the follow-up and for your constructive comments!

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank the reviewers and the Area Chair. The reviewers regard the distinction between encoding visual evidence and determining the answer as important (d7Xm, uJsy) and note the breadth and controls of the evaluation (uJsy, CQR3). Their shared questions concern the original errors, the depth difference on image-dependent questions, and the practical gain under benchmark-standard scoring. We address each with new analyses.

1. Intermediate visual representations contain evidence for the correct answer even on original errors. On TextVQA questions that the Clean run answers incorrectly, Transfer at ℓ_sel raises accuracy over Noise by 8.28 pp, with gains in all eleven models (Section 4.1, Appendix C.3; CQR3 Q1, uJsy Q1). Moreover, the evidence that selection uses is already encoded before the candidates appear: visual states computed under the open-ended prompt, once inserted into a two-choice prompt with a gray image, shift the selection toward the answer of their source image in 53 of 54 image pairs (Appendix C.6; uJsy Q2).

2. The difference in dependency depth persists on image-dependent questions. On TextVQA questions that open-ended answering and two-choice selection both answer correctly with the original image but not with a gray image, selection depends on visual tokens through 0.36–0.50 of model depth fewer layers than open-ended answering, consistent with the median difference of 0.42 across all questions (Section 3.2, Appendix B.2; uJsy Q3, CQR3 Q2).

3. The main conclusions hold under benchmark-standard scoring. As CQR3 suggested, two-stage inference is now scored with VQA accuracy for TextVQA and exact match for GQA (Appendix A.1). With only one answer from each of the other two models, which is the lowest-cost configuration in the submission, two-stage inference corrects 33.0% of original open-ended errors and improves accuracy by 2.49 pp [1.37, 3.62]. It also outperforms frequency voting on identical candidate pools and self-consistency over eight samples from the target model (Section 5.3, Appendix E.1; d7Xm C2–C3, CQR3 Q3). Because the analyses of visual dependence and intermediate representations in Sections 3 to 5.2 compare conditions on the same questions under a fixed rule, their results are unchanged. All new numbers are computed from saved outputs without rerunning any model.

Revision. Changes in the uploaded paper are shown in red, and section and table numbers in our responses refer to the revised version. In line with point 3, the abstract, Introduction, Figure 7 and Table 1 now report the configuration with one answer per model. In addition, Section 5.2 defines entropy as the diversity of each task's answers (uJsy W4).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.