acceptodds
Under review as a conference paper at ICLR 2027

AnswerMap: Faithful Spatial Interpretability of VLMs from Answer Posteriors

Abstract

When a VLM answers a visual query, current interpretability tools rely on text rationales, which use a mismatched modality, or on internal read-outs, which originate too early to reflect the final output and require white-box access to the model. We introduce AnswerMap, a training-free, task-agnostic, black-box visual rationale constructed from the output head. The image is cut into row and column bands, each shown alone to the frozen model along with the query in the format of a yes/no relevance question. The outer product of the row and column “yes” posteriors gives the query-conditioned spatial map. Crucially, by defining a fixed read-out (e.g., expectation, maximum) on top of AnswerMap, we can derive continuous outputs like location natively. This bypasses the reliance on discrete text tokens for continuous-output tasks and guarantees an image-dependent answer by construction. However, a rationale can be confabulated, so we validate AnswerMap across four models and three query distributions with two tests: (a) agreement with the model's own generated point and (b) deletion of the map's region. The map lands where the model points (AUC 0.85 against 0.38 for attention), and deleting its region flips 53% of correct answers (against 19% for attention's). Beyond establishing faithfulness, we demonstrate the map's task-agnostic utility through three distinct read-outs: its maximum flags hallucinated objects without generation, its expectation localizes correctly when the model's own pointing fails, and its top-mass region, fed back as a crop, fixes half of the model's wrong answers. AnswerMap thus offers a new lens on VLM interpretability and, through its read-outs, a new output interface for visual tasks beyond text tokens.

Primary area:
interpretability and explainable AI
Keywords:
interpretability, explainability, vision-language models, multimodal large language models, faithfulness, black-box explanation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XX3XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Aiming to investigate the problem of mismatched VLM answers-text rationales, and early internal read-outs cannot reflect the final output. This paper introduces AnswerMap, a training-free/task-agnostic/black-box visual rationale constructed from the output head, which bypasses the reliance on discrete text tokens, providing a new lens on VLM interpretability.

Critical strengths
  • The method requires no training and relies only on output token probabilities, making it easy to apply across different VLM architectures without adapting to their internal visual token representations.

  • Evaluation across models with useful controls. The experiments cover multiple model sizes and two model families, and assess both spatial agreement and the effects of deleting selected regions. Matched random deletion, blurred masking, and blank or swapped image controls help characterize the method’s behavior, even though they do not fully establish faithfulness to the original decision process.

Critical weaknesses

My main concern is whether the paper sufficiently demonstrates faithfulness to the original model’s decision process.

  1. Mismatch between probing and original answering. AnswerMap probes the model using cropped images and reformulated questions, rather than directly examining how the original answer was produced. Since both the visual input and the question are modified, these probing responses may rely on different evidence from the original answer. The connection between the resulting map and the original decision therefore requires stronger justification.

  2. Image dependence does not establish faithfulness. Ensuring that a map depends on visual input does not necessarily mean that it identifies the evidence responsible for the original answer. The current results may demonstrate localization of semantically relevant content rather than faithful attribution of the model’s decision. Additional controls are needed to distinguish these interpretations.

Questions for authors
  1. Compare AnswerMap with a baseline that directly probes individual image cells using the same Yes/No prompt. Although more expensive, this comparison would quantify the spatial information lost through the row and column factorization. Comparisons with coarse grids, sliding windows, or hierarchical cropping under the same query budget would also help establish whether AnswerMap offers advantages over simpler ways of allocating computation.

  2. Apply the same deletion test to regions identified by another VLM or an independent grounding method. If these regions produce similar answer change rates, the results would suggest that AnswerMap primarily identifies generally relevant semantic content, rather than evidence specific to the original model’s decision.

I'm willing to consider revising the rating if my concerns are carefully addressed.

4 comments

We thank the reviewer for the constructive comments. Below are the new experiments they led to: a semantic-matching control (W2, Q2), and a per-cell probe with three other ways to spend the same query budget (Q1).

Link between the band queries and the original answer (W1). The two tests check this connection, and both compare the map with the answer the VLM gives on the original image with the original question. The intuition is that AnswerMap is a second, interpretable path to the answer. If this path agrees with standard inference, that suggests it reflects how the original answer was reached.

  • Test 1 checks that the two paths agree. The map is scored at the point the VLM itself generates on the full image. The agreement AUC is 0.853 on RefCOCOg, 0.840 on RefCOCO+ and 0.845 on CAVE (Table 2), so the two paths agree.
  • Test 2 checks that the two paths use the same evidence. Agreement alone leaves open that the band queries rely on different evidence. If they did, deleting the 8 cells that AnswerMap marks would change the original answer about as often as deleting 8 random cells. Deleting the 8 cells of AnswerMap changes 53.4% of the original answers, against 8.2% for random cells (Table 3). On POPE, in the 223 cases where the VLM answers yes although the named object is not in the image, deleting the 8 cells of AnswerMap changes the wrong yes 69.5% of the time, against 11.2% for random cells (Table 6).

While the band queries use a different input, the deletion results show that the original answer depends on the region they identify. We treat this as evidence that the band queries capture the relevant evidence, and not an unrelated signal.

Semantic-matching control (W2, Q2). The reviewer asks whether regions identified by an independent grounding method would give similar answer change rates, which would mean that AnswerMap identifies generally relevant semantic content and not evidence specific to the VLM's decision. The results below show that AnswerMap follows the VLM and not the semantics of the query.

An independent grounding method only matches the query to the image, with no access to the VLM. We call such a method a semantic matcher. When the VLM answers correctly, the part of the image it answered from is where the queried object is, and a semantic matcher marks that part too. On correct answers a faithful map and a semantic matcher are therefore expected to give similar answer change rates, so correct answers alone cannot separate the two. However, the VLM's wrong answers can provide this separation. The clearest case is a yes to an object that is not in the image. Here the VLM cannot have answered from the queried object, because there is no such object in the image, so a faithful map and a semantic matcher mark different regions. If AnswerMap were only a semantic matcher, deleting its region in this case would change the wrong answer no more often than deleting the matcher's region. If AnswerMap is faithful, deleting its region changes the wrong answer more often. The table below shows this test on POPE, the benchmark of Table 6, which asks whether a named object is in the image. We delete the region marked by AnswerMap, by the semantic matcher (the best box of the detector GroundingDINO for the named object) or at random, and count how often the VLM's yes changes. The test is run on two VLMs, each explained by its own AnswerMap, on the same questions: 800 where the named object is in the image and 4,500 where it is not. Each column holds the questions that its VLM answers yes to.

Deleted regionQwen3-VL-4B: yes, and the object is in the image (correct, n=672)Qwen3-VL-4B: yes, and the object is not in the image (wrong, n=223)InternVL3.5-8B: correct (n=726)InternVL3.5-8B: wrong (n=512)
AnswerMap38.269.531.558.0
Semantic matcher (detector's box)35.357.430.348.4
Random contiguous block5.718.84.813.3
Random cells1.211.21.010.9

All entries are the share of answers changed (%). On the correct answers, deleting the detector's box changes nearly as many answers as deleting AnswerMap's region, for both VLMs (35.3% against 38.2%, and 30.3% against 31.5%), as expected. On the wrong answers, deleting AnswerMap's region changes the wrong yes more often than deleting the detector's box, for both VLMs: 69.5% against 57.4% for Qwen3-VL-4B (gap 12.1 points, 95% CI 5.4 to 18.8), and 58.0% against 48.4% for InternVL3.5-8B (gap 9.6 points, 95% CI 4.5 to 14.6). The same detector that nearly matches AnswerMap on correct answers falls behind on wrong answers. On TextVQA, the benchmark of Table 3, the detector's box changes 27.5% of the answers against 53.4% for AnswerMap.

Conclusion. On the VLM's wrong answers AnswerMap follows the VLM and the semantic matcher falls behind, so AnswerMap identifies evidence specific to the VLM's decision, and not only generally relevant semantic content.

Cross-model control (Q2). The reviewer also asks for regions identified by another VLM. We ran this with InternVL3.5-8B as the second VLM. We argue that this control cannot isolate faithfulness, because its result depends on two things at once: whether the map is faithful, and how often the two VLMs give the same answer to the same question. The detector control above does not have this problem. A detector returns a box for the named object. It does not answer the question the VLM is asked (is the object in the image?), so it cannot give the same wrong answer as the VLM. A second VLM is asked the same question, and when it gives the same wrong answer, its map marks the evidence for that same wrong answer. This happens often: InternVL gives the same wrong yes in 189 of Qwen's 223 wrong cases, where about 25 would be shared if the two VLMs made their wrong answers independently. Deleting the region of InternVL's AnswerMap changes Qwen's wrong yes 64.1% of the time, against 69.5% for Qwen's own AnswerMap (gap 5.4 points, 95% CI -0.4 to 11.2). On TextVQA the two rates are 50.1% and 53.4%, and the two VLMs give the same answer on three in four of the questions. The two VLMs often give the same answer, so close rates are expected here, and we therefore rely on the detector control above to isolate faithfulness.

To accommodate this important control in the paper, we will extend Test 2 to cover both the VLM's correct answers, which are already in the paper, and its wrong answers. We believe Test 2 then evaluates faithfulness more comprehensively.

Per-cell probe and other ways to spend the query budget (Q1). The row and column factorization loses nothing on the two tests: probing each of the 64 cells alone with the same Yes/No prompt uses four times the queries and scores lower on both. AnswerMap also scores above a coarse grid, sliding windows and hierarchical cropping at the same query budget. All four were run on the examples of Tables 2 and 3.

MethodQueriesRefCOCOg, agreement (AUC)TextVQA, answers changed (%)
AnswerMap160.85353.4
Per-cell probe, 8x8 cells, each shown alone640.73442.8
Coarse grid, 4x4 cells, each shown alone160.78312.0
Sliding windows, 16 half-size windows160.76828.3
Hierarchical cropping, three rounds of four crops120.62420.4
Random cells0.4888.4

The per-cell probe is AnswerMap without the factorization, and it is 0.119 AUC behind on agreement and 10.6 points behind on deletion (95% CI 5.4 to 15.8). The coarse grid, the sliding windows and the hierarchical cropping are also below AnswerMap on both tests. AnswerMap therefore has an advantage over each of these simpler ways to spend the queries, and the comparison will be added to the ablation study (Section 6).

Reviewer_XX3XOpenReview ↗

Official Comment by XX3X

I thank the authors for the comprehensive rebuttal, which addresses my concerns regarding the baseline comparisons and stronger empirical support for the proposed method.

However, I remain concerned about whether the results sufficiently establish faithfulness to the original model's decision process. While the deletion experiments demonstrate that the identified regions influence the original predictions, such behavioral sensitivity does not necessarily establish that the probing procedure faithfully captures the evidence underlying the original inference.

Therefore, I'd like to keep my original score.

We thank the reviewer for the reply, and we are glad that the rebuttal addressed the concerns about the baseline comparisons and the empirical support.

We would like to clarify that AnswerMap is, in essence, an interpretable approximation of the VLM's decision process rather than a direct reading of that process, and the two tests measure the error of this approximation. Because the exact path of a decision inside the model is hard to interpret, AnswerMap builds a second, independent path to the same answer, one that is interpretable, through a different input and output. We see this as faithfulness by proxy, a new lens for interpreting a model. The error of the approximation is the divergence between the two paths. Test 1 measures this divergence on the answer and Test 2 on the evidence, with controls that rule out random regions and semantic matching (Tables 2, 3 and 6). When an interpretable path reaches the same answer from the same evidence as the original inference, this suggests that it reflects how that answer was reached, and the results of both tests and their controls support this for AnswerMap. Seen from this perspective, we believe AnswerMap is a first step in a new direction for interpreting VLMs. We will reflect this framing in the final version of the paper.

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

The paper introduces AnswerMap, a training-free, black-box spatial rationale method designed to interpret Vision-Language Model (VLM) predictions. The authors critique existing interpretability tools: text rationales suffer from modality mismatch, internal read-outs (like attention or gradients) originate too early and require white-box access, and traditional occlusion perturbation is prohibitively expensive. AnswerMap circumvents these issues by slicing an image into K row and K column bands, querying the frozen VLM with a yes/no relevance question for each band in isolation, and computing the outer product of the row and column "yes" posteriors to form a KxK heatmap. This reduces the cost to 2K forward passes.

To validate faithfulness without relying on human ground-truth labels, the authors propose a two-test protocol: Test 1 (Agreement) checks if the map's peak aligns with the model's own generated point, and Test 2 (Deletion) checks if deleting the map's high-mass region flips the model's answer. Across multiple VLMs (Qwen3-VL, InternVL), AnswerMap significantly outperforms attention, gradient-based methods (T-MM), and occlusion. Furthermore, the authors demonstrate that fixed read-outs applied to AnswerMap provide actionable utility: taking the expectation yields precise coordinates when the model's own pointing fails, taking the maximum reliably flags hallucinated objects, and cropping the top-mass region allows the model to self-correct 50% of its wrong answers.

Critical strengths
  1. Formulating the 2D spatial rationale as the outer product of 1D row and column posteriors is an innovative and cost-effective factorization.
  2. The multi-grid approach is interesting and it forces the rationale to be supported by both broad global context (wide bands) and sharp local detail (thin bands).
  3. The paper successfully elevates the rationale from a passive diagnostic tool to an active output interface.
Critical weaknesses
  1. As the authors briefly acknowledge in the conclusion, the outer product formulation fundamentally breaks down when the query matches two or more spatially separated objects. If an object exists in row 2, column 2, and another in row 6, column 6, the outer product will automatically hallucinate high mass at the empty intersections (row 2, col 6 and row 6, col 2). This structural flaw limits AnswerMap's applicability in dense, multi-instance reasoning tasks.
  2. The narrow 1D band strips out a lot of contextual cues from neighboring parts.
  3. The method relies entirely on the log-probabilities of the "Yes" and "No" tokens to isolated image slices. This assumes the VLM is well-calibrated for binary classification on abstract.
Questions for authors
  1. Could you add some qualitative examples? Additionally, could human evals also be added?
  2. How sensitive is the yes/no posterior to the exact phrasing of the band probe prompt?
3 comments

We thank the reviewer for the constructive comments. Below are the new experiments they led to: a count of how often crossing cells appear (W1), a per-cell probe that shows each cell alone (W2), and rewordings of the band prompt on both tests (Q2).

Two separated objects (W1). This is the limitation our conclusion states. It is real, however it is rare. On RefCOCOg, taking the ground-truth box and a detector's other candidates for the query as the objects, the top cell of the default 8x8 map is an empty crossing in only 1.0% of the maps (95% CI 0.4 to 1.8), and empty crossings are 1.8% of its 8 top cells. On the 197 of the 800 images (25%) that contain two candidate objects in different rows and columns, which is the case the reviewer describes, the top cell is an empty crossing in 4.1% of the maps, and empty crossings are 7.2% of the 8 top cells. It is therefore not a large problem on the tasks of the paper. For queries that match many separate objects we agree that it remains, and we keep it as a stated limitation. The revision will report these counts.

Context in a band (W2). The bands used in the paper keep the context the answer needs. The multigrid product, which the reviewer lists among the strengths, is built for this: wide bands carry the context, thin bands carry the detail, and a region survives only if both support it (Section 3.2). The paper also measures how much context a band needs. In Table 8, agreement collapses only when the bands become very thin (NSS 1.46 at K=6, 0.22 at K=40), and the grids {3,5} reach 1.50 with the same 16 queries as K=8 (1.14). A band also keeps more context than a cell, because it spans the full width or height of the image: in a new run, each of the 64 cells shown alone reaches an agreement AUC of 0.734 against 0.853 for AnswerMap, with four times the queries. Context is therefore measured in the paper, and the band width is set where it is kept.

Calibration of the yes/no answer (W3). The method does not assume a calibrated VLM, because its scoring is relative. It only needs a band that contains the queried content to get a higher P(yes) than a band that does not, in the same image. The read-outs (peak, expectation, top cells) use these relative scores and apply no threshold. A VLM that says yes too readily, or too rarely, therefore moves all bands together, and the peak of the map stays in place (Section 3.1). Two results in the paper check that the band answers carry this signal. On a blank image the map is flat and both agreement metrics land exactly at chance (Table 2). And the map's maximum separates hallucinated from correct claims at ROC-AUC 0.833 with no calibration step (Section 5). The band answers are therefore informative without any calibration.

Qualitative examples and human evaluation (Q1). Qualitative examples are in Appendix A.3 (Figure 7), and the revision will add more. Human evaluation is already in the paper where humans can judge the map: Table 5 scores it against human-annotated boxes, and the map's expectation lands inside the box 67.5% of the time on RefCOCOg, against 40.5% for the image centre. Faithfulness is different, and how to judge it is part of the paper's problem statement. It asks whether the map marks what the VLM used, which is hard for a human to judge, because a human sees the image and the VLM's answer, and not the steps in between that show how the VLM reached that answer. The two tests, the third contribution of the paper, are our solution. AnswerMap is a second, interpretable path to the VLM's answer, and the two tests check it against the answer the VLM gives by standard inference. Table 2 shows that AnswerMap is the explanation closest to that answer (agreement AUC 0.853), and Table 3 that deleting its region changes that answer (53.4%, against 8.2% for random cells). Table 6 gives the strongest evidence. It uses the POPE benchmark, which asks whether a named object is in the image. In 223 cases the VLM answers yes although the object is not in the image, so the image has no correct region for a human to mark. If AnswerMap follows what the VLM relied on for this wrong yes, deleting its region should change the answer often. If it does not, deleting its region should change the answer about as often as deleting random cells. Deleting AnswerMap's region changes the wrong yes in 69.5% of the 223 cases, against 11.2% for random cells. This shows that AnswerMap is faithful to the VLM, with no human labels.

Sensitivity to the band prompt (Q2). The results hold when the band prompt is reworded. We ran two rewordings on each test, written for its kind of query (an expression in Test 1, a question in Test 2).

TestBand promptScore
Agreement on RefCOCOg (AUC)Paper: "Is the following present in ANY of these tiles?"0.853
"Does any of these tiles show the following?"0.852
"This is a strip cut out of a bigger picture. Can the following be found anywhere in this strip?"0.848
Deletion on TextVQA (answers changed)Paper: "Is the following present in ANY of these tiles?"53.4%
"Is the answer to the following question visible in ANY of these tiles?"57.2%
"Do these tiles contain what the following question asks about?"58.3%

Agreement moves by at most 0.005. On deletion the rewordings give the same picture, a few points higher (random cells: 8.2%), so the paper's wording is the conservative one. Appendix A.2 will report this table.

Thank you for the detailed response and additional results! I appreciate the clarification on crossing-cell counts, context in a band, and prompt-sensitivity. The response adequately resolved my concerns and hence I have increased my score.

Reviewer_8mLWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AnswerMap, a training-free method that derives a spatial rationale for VLM answers from the first-token logits of a yes/no presence question. The image is cut into K horizontal and K vertical bands; each band is shown alone to the frozen model, and the outer product of the row and column posteriors yields a K \times K map at 2K forward passes. The authors assess faithfulness with two tests that require no ground truth: agreement between the map and the point the model generates, and changes in answers after deleting the highest-scoring cells. On four open-weight models, AnswerMap scores above attention, relevancy, and occlusion baselines on both tests. Lastly, the paper applies map readouts to hallucination detection, medical localization, and crop-based self-conditioning.

Critical strengths
  1. The method is simple, inexpensive, and independent of the architecture. Because the bands are cut in pixel space and the probe reads only label probabilities, it does not need to handle dynamic resolution, tiling, or token merging, and it costs 2K forward passes per map (Table 9). In addition, the paper openly discusses its limits, e.g., that the rank-one structure assigns mass to empty cells when the query matches two objects.

  2. The evaluation includes controls that many attribution papers omit. The blank-image and swapped-image probes show that the map follows the image it is shown (Tables 2 and 11), and the main pattern holds on three additional models without per-model tuning (Table 4).

Critical weaknesses
  1. The two tests do not yet separate faithfulness to the explained model from good localization of the query. On RefCOCOg, the point that the model generates lies in the ground-truth box 94.5% of the time (Table 5), so Test 1 is close to a pointing game against ground truth. Similarly, deleting the queried object in Test 2 breaks many answers regardless of which model produced the map. Hence, a map without access to the VLM, such as an open-vocabulary detector or the ground-truth box, or a map computed with a different VLM might pass both tests equally well. The paper reports none of these sanity checks (Adebayo et al., 2018). In addition, the deletion test uses a single budget and a scattered random control, whereas the top cells of the rank-one map form a contiguous block. Lastly, standard baselines such as Grad-CAM, integrated gradients, and RISE are missing, and the selected attention-layer scores are below chance (AUC 0.380), making it a weak reference for the comparison in the abstract.

  2. The paper reports no measure of uncertainty, and several claims exceed the evidence. No table reports confidence intervals or paired tests, although some margins are small, e.g., 45.8% against 45.0% for occlusion at 30B (Table 4). The abstract states that the crop "fixes half of the model's wrong answers", whereas the net gain in Table 7 is 0.8 points. In addition, the 11-point improvement with Lingshu-7B is confounded by the 30% of generated points that fail to parse and count as misses. Because the operator closely follows GridProbe (Eltahir et al., 2026), its main contribution lies in evaluation, which therefore needs to be more rigorous.

Questions for authors
  1. Do a model-independent map (e.g., an open-vocabulary detector or the ground-truth box) and a cross-model map (e.g., the probe computed with InternVL3.5-8B and evaluated against the points and answers of Qwen3-VL-4B) pass Test 1 and Test 2 at similar levels? If the probe clearly outperforms both, this would be the strongest evidence for the faithfulness claim and would raise my score.

  2. How do the deletion results change with curves over several budgets and with a shape-matched contiguous random region as the control?

  3. How does the probe compare with Grad-CAM, integrated gradients, and RISE at a matched number of forward passes?

  4. Can the authors report confidence intervals, e.g., based on a paired bootstrap over examples, for Tables 2 to 8?

  5. How is the presence condition formed for TextVQA and GQA? Appendix A.2 suggests inserting the benchmark question itself, which would ask whether, e.g., "what is written on the sign?" is present in a band.

3 comments

We thank the reviewer for the sharp and constructive comments. Below are the new experiments they led to: a semantic-matching control (Q1), deletion curves with a contiguous control (Q2), Grad-CAM, integrated gradients and RISE at matched passes (Q3), confidence intervals (Q4), and prompt rewordings (Q5).

Semantic-matching control (W1, Q1). The reviewer asks whether a map without access to the VLM, such as an open-vocabulary detector or the ground-truth box, or a map computed with a different VLM, would pass both tests equally well. The results below show that AnswerMap follows the VLM and not the semantics of the query.

A map without access to the VLM comes from a method that only matches the query to the image. We call such a method a semantic matcher. When the VLM answers correctly, the part of the image it answered from is where the queried object is, and a semantic matcher marks that part too. On correct answers a semantic matcher is therefore expected to score as well as a faithful map on both tests, so correct answers alone cannot separate the two. However, the VLM's wrong answers can provide this separation. The clearest case is a yes to an object that is not in the image. Here the VLM cannot have answered from the queried object, because there is no such object in the image, so a faithful map and a semantic matcher mark different regions. If AnswerMap were only a semantic matcher, deleting its region in this case would change the wrong answer no more often than deleting the matcher's region. If AnswerMap is faithful, deleting its region changes the wrong answer more often. The table below shows this test on POPE, the benchmark of Table 6, which asks whether a named object is in the image. We delete the region marked by AnswerMap, by the semantic matcher (the best box of the detector GroundingDINO for the named object) or at random, and count how often the VLM's yes changes. The test is run on two VLMs, each explained by its own AnswerMap, on the same questions: 800 where the named object is in the image and 4,500 where it is not. Each column holds the questions that its VLM answers yes to.

Deleted regionQwen3-VL-4B: yes, and the object is in the image (correct, n=672)Qwen3-VL-4B: yes, and the object is not in the image (wrong, n=223)InternVL3.5-8B: correct (n=726)InternVL3.5-8B: wrong (n=512)
AnswerMap38.269.531.558.0
Semantic matcher (detector's box)35.357.430.348.4
Random contiguous block5.718.84.813.3
Random cells1.211.21.010.9

All entries are the share of answers changed (%). On the correct answers, deleting the detector's box changes nearly as many answers as deleting AnswerMap's region, for both VLMs (35.3% against 38.2%, and 30.3% against 31.5%), as expected. On the wrong answers, deleting AnswerMap's region changes the wrong yes more often than deleting the detector's box, for both VLMs: 69.5% against 57.4% for Qwen3-VL-4B (gap 12.1 points, 95% CI 5.4 to 18.8), and 58.0% against 48.4% for InternVL3.5-8B (gap 9.6 points, 95% CI 4.5 to 14.6). The same detector that nearly matches AnswerMap on correct answers falls behind on wrong answers. On RefCOCOg (Test 1), where the VLM's point is inside the ground-truth box 94.5% of the time, the ground-truth box and the detector reach an agreement AUC of 0.914 and 0.879, against 0.853 for AnswerMap: on correct answers the semantic matchers score as high as AnswerMap or higher, as expected. On TextVQA (Test 2, Table 3), the detector's box changes 27.5% of the answers against 53.4% for AnswerMap.

Conclusion. On the VLM's wrong answers AnswerMap follows the VLM and the semantic matcher falls behind, so the two tests are not reducible to semantic matching.

Cross-model control (Q1). The reviewer also asks whether a map computed with a different VLM would pass both tests. We ran this with InternVL3.5-8B as the second VLM. We argue that this control cannot isolate faithfulness, because its result depends on two things at once: whether the map is faithful, and how often the two VLMs give the same answer to the same question. The detector control above does not have this problem. A detector returns a box for the named object. It does not answer the question the VLM is asked (is the object in the image?), so it cannot give the same wrong answer as the VLM. A second VLM is asked the same question, and when it gives the same wrong answer, its map marks the evidence for that same wrong answer. This happens often: InternVL gives the same wrong yes in 189 of Qwen's 223 wrong cases, where about 25 would be shared if the two VLMs made their wrong answers independently. Deleting the region of InternVL's AnswerMap changes Qwen's wrong yes 64.1% of the time, against 69.5% for Qwen's own AnswerMap (gap 5.4 points, 95% CI -0.4 to 11.2). On TextVQA the two rates are 50.1% and 53.4%, and the two VLMs give the same answer on three in four of the questions. On RefCOCOg (Test 1), InternVL's AnswerMap reaches an agreement AUC of 0.839 against 0.853 for Qwen's own (gap 0.015, 95% CI 0.004 to 0.026). The two VLMs often give the same answer, so close rates are expected here, and we therefore rely on the detector control above to isolate faithfulness.

To accommodate this important control in the paper, we will extend Test 2 to cover both the VLM's correct answers, which are already in the paper, and its wrong answers. We believe Test 2 then evaluates faithfulness more comprehensively.

Deletion curves and a contiguous control (Q2). The deletion result holds at every budget and against a shape-matched contiguous control. We ran four budgets on the questions of Table 3, with a random contiguous block of the same number of cells next to the scattered random cells.

Cells deleted (of 64)48 (paper)1624
AnswerMap, answers changed (%)40.953.465.971.7
Random contiguous block6.58.224.533.5
Random cells5.28.418.827.8
AnswerMap minus contiguous block (95% CI)34.3 (28.9 to 39.8)45.2 (39.5 to 50.7)41.4 (35.4 to 47.1)38.1 (32.2 to 43.9)

Deleting AnswerMap's region changes far more answers than deleting a contiguous block at every budget, and at the paper's budget the contiguous block changes no more answers than scattered cells (8.2 against 8.4). The result therefore does not depend on the budget or on the shape of the control.

Grad-CAM, integrated gradients and RISE at matched passes (Q3). AnswerMap leads all three on both tests at a matched number of forward passes. Each method gets 16 forward passes, the cost of one AnswerMap. Integrated gradients also uses 16 backward passes, and Grad-CAM needs a single forward and backward pass.

MethodReadsRefCOCOg, agreement (AUC)TextVQA, answers changed (%)
AnswerMapoutput probabilities0.853 (0.841 to 0.866)53.4 (48.2 to 58.6)
RISE, 16 masksoutput probabilities0.693 (0.673 to 0.712)39.2 (34.3 to 44.1)
Integrated gradients, 16 stepsgradients0.606 (0.584 to 0.628)35.7 (31.1 to 40.6)
Grad-CAMgradients0.497 (0.477 to 0.516)10.1 (7.1 to 13.4)
Random0.4888.4

The intervals of AnswerMap and of the three methods do not overlap. RISE, the other method that reads only output probabilities, comes second on both. Grad-CAM stays at chance on the visual tokens of this VLM, and the gradient-based relevancy method T-MM, already in Tables 2 and 3, remains the strongest gradient baseline (0.656 and 36.8).

Confidence intervals (Q4). In the appendix of the revision we will add 95% confidence intervals for Tables 2 to 8. The main rows:

AnswerMapOcclusion (65 queries)Random
Table 2, RefCOCOg, agreement (AUC)0.853 (0.841 to 0.866)0.649 (0.628 to 0.669)0.486 (0.466 to 0.507)
Table 3, TextVQA, answers changed (%)53.4 (48.2 to 58.6)48.8 (43.6 to 54.0)8.2 (5.4 to 10.9)

The intervals confirm the paper's main comparisons: AnswerMap is clearly above occlusion and random on agreement, and clearly above random on deletion. On deletion against occlusion, the reviewer is right that the margin is small: the two intervals overlap for Qwen3-VL-4B (53.4 against 48.8), and 45.8% against 45.0% for Qwen3-VL-30B is within noise. What this comparison shows is the cost: AnswerMap reaches the deletion rate of occlusion with 16 queries per map, where occlusion needs 65. We will revise the conclusion about occlusion in the paper to reflect this.

Presence condition for TextVQA and GQA (Q5). The benchmark question is inserted as is, as the reviewer reads Appendix A.2: the band prompt asks whether the quoted question "is present in any of these tiles". To check that the result does not rest on this loose wording, we ran two wordings that state the condition explicitly.

Band promptTextVQA, answers changed (%)Random cells
Paper: "Is the following present in ANY of these tiles?"53.48.2
"Is the answer to the following question visible in ANY of these tiles?"57.28.2
"Do these tiles contain what the following question asks about?"58.38.2

The explicit wordings give the same picture, a few points higher, so the paper's wording is the conservative one. On RefCOCOg, two rewordings of the prompt move the agreement AUC by at most 0.005 (0.853, 0.852, 0.848). Appendix A.2 will state the condition and report this table.

Abstract (W1, W2). We agree with both points. In the abstract we will replace attention with occlusion, the strongest baseline, as the reference, and give the net gain of the crop (0.8 points) and its margin over a random crop (3.4 points, 95% CI 1.4 to 5.6) next to the half of the wrong answers it fixes.

Lingshu-7B and parse failures (W2). The gain of the map over the generated point remains when parse failures are set aside. We recomputed both read-outs on the expressions where Lingshu's point parses (n=322). The map's expectation lands in the ground-truth box on 68.6% of them, and the generated point on 58.7%. The revision will report this view.

Comments on the paper

1 comment

We thank the reviewers for their constructive comments. Each question is answered in the reply under its review, and the new results are now in the revised PDF (changes in blue).

Semantic-matching control (8mLW, XX3X). The main concern of the reviews was that AnswerMap may only match the query to the image and not follow the VLM. To address it, we added the control the reviewers suggested, a semantic matcher with no access to the VLM, and AnswerMap passes it. We added this control to Test 2. (Section 4.5, Table 6)

Faithfulness by proxy (XX3X). The reviews raised that AnswerMap asks the VLM with a different input and output than the original inference, so it may not follow the original decision process exactly. To address it, we clarified that AnswerMap is, in essence, an interpretable approximation of the VLM's decision process rather than a direct reading of that process. It builds a second, independent path to the same answer, and the two tests measure the error of this approximation, the divergence between the two paths, on the answer (Test 1) and on the evidence (Test 2). We see this as faithfulness by proxy, a new lens for interpreting VLMs, and we will reflect this framing in the final version.

Robustness (8mLW, sALy, XX3X). The results hold across deletion budgets and against a contiguous random control, against Grad-CAM, integrated gradients and RISE at the same number of passes, against a per-cell probe and other ways to spend the queries, and under rewordings of the band prompt. (Sections 4.2 and 6, Appendix A.1, A.2, A.4)

Statistical reporting (8mLW). The reviews raised missing confidence intervals and the reference baseline chosen in the abstract. To address them, we added 95% confidence intervals for all main tables, which confirm the main comparisons, and revised the abstract and the occlusion comparison to match them. (Abstract, Section 4.3, Appendix A.6)

Scope (sALy, 8mLW). The reviews raised false peaks where two separated objects cross, parse failures in the Lingshu result, and the number of qualitative examples. To address them, we added new results showing that such peaks are rare and that the Lingshu gain holds on the parsed outputs alone, and a new qualitative example on a closed model. (Section 5, Conclusion, Appendix A.3, A.4)

We hope these results address the reviewers’ concerns, and we thank the AC and all reviewers again for their time and effort.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.