- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper introduces VTRACE, a token-attribution framework for explaining the visual and textual evidence supporting responses generated by large vision-language models. The method combines three components: modality-centered pairwise attribution, closed-form aggregation of direct and indirect paths through preceding tokens, and cross-modal calibration using two-modality Shapley contributions estimated from response-likelihood changes.
Critical strengths
S1. A meaningful problem with a clear multimodal motivation. Identifying which image regions and prompt tokens support a generated reasoning trace is an important interpretability problem. The paper identifies two relevant challenges: visual evidence can be under-ranked in joint attribution, and its influence may pass through intermediate generated tokens. The proposed method directly addresses both issues.
S2. A coherent design supported by useful ablations. The three components have distinct roles, and Table 2 examines their individual contributions. The substantial degradation without multi-hop aggregation supports the importance of this component under the chosen evaluation. Appendix H.2 further applies multi-hop aggregation to IFR, suggesting that the idea may be useful beyond the specific VTRACE construction. Reusing the attribution matrix across target tokens is also a practical advantage, supported by the reported efficiency measurements.
S3. Broad and consistent empirical evidence under the stated evaluation protocol. The experiments include seven baselines, six benchmarks, image-only and joint perturbations, and multiple model families and sizes. Results are reported for both correct and incorrect predictions, avoiding evaluation restricted to successful responses. The GRPO experiment provides an additional, promising indication of downstream utility beyond attribution visualizations.
Critical weaknesses
W1. The scope of the faithfulness evidence should be distinguished more clearly from answer-level and mechanistic interpretations. The paper explicitly defines faithfulness through changes in the likelihood of a fixed generated response, and the evaluation is appropriate for that definition. However, scoring the complete response under teacher forcing does not directly establish which evidence determines the final answer when reasoning is regenerated. In particular, information removed from the image may remain available through the original reasoning prefix. The results therefore support attribution of a fixed response more directly than the broader interpretation of recovering the evidence responsible for the model’s answer and reasoning process. This distinction matters when interpreting the path visualizations and the paper’s conclusions. Relevant sections: Section 5; Appendices F and G.
W2. The interpretation of multi-hop scores as evidence-propagation paths needs more direct validation. Equation 6 correctly aggregates paths in the constructed attribution matrix, but the matrix itself averages layer-level scores after modality centering and positive clipping. Paths in this collapsed graph need not correspond to layer-consistent paths in the model’s computation. Moreover, the incoming-flow factor in the score can favor tokens lying on many paths. The ablation establishes that aggregation improves the perturbation metrics, but does not by itself determine whether the recovered intermediate paths faithfully represent how evidence is used. A targeted validation would strengthen this central interpretation. Relevant sections: Sections 4.1–4.2; Appendix C.
W3. Robustness to the perturbation and metric-processing choices is insufficiently established. Appendix G applies cumulative minima or maxima to enforce monotonic perturbation curves and clips the resulting curves before computing AUC. These transformations can affect method rankings when model responses are nonmonotonic, making results without these transformations important. In addition, modality calibration uses padding embeddings, whereas visual evaluation uses blurred image regions followed by re-encoding. This does not invalidate the evaluation, but leaves uncertainty about how dependent the gains are on the selected interventions. Applying the same calibration to strong baselines would also help isolate the contribution of the underlying token rankings. Relevant sections: Section 4.3; Appendices D and G.
W4. The attribution-guided GRPO experiment does not yet isolate the benefit of VTRACE’s token selection. The average improvement of 1.28 percentage points over GRPO is encouraging. However, the experiment assigns weight 1.5 to the top 40% of response tokens without comparing against alternative token-selection rules under the same weighting budget. Consequently, the improvement cannot yet be specifically attributed to VTRACE’s attribution quality. Table 4 also does not provide uncertainty across repeated training runs, and the gains are not uniform across tasks, with small decreases on MathVista and MMMU-Pro. Relevant sections: Table 4; Appendix F.1.
Questions for authors
-
Can the authors clarify which conclusions concern support for a fixed response and which concern the final answer decision? On a representative subset, do the advantages persist when attributing only the final-answer span and when regenerating the reasoning and answer after perturbation? Please distinguish final-answer scoring with the original reasoning prefix from evaluation with regenerated reasoning.
-
Can the authors validate selected high-scoring intermediate paths through targeted activation interventions or a controlled task with known evidence dependencies, comparing them with matched low-scoring or random paths? An ablation using outgoing attribution alone would also help determine what the incoming-flow factor adds beyond path aggregation.
-
Do the main rankings persist without monotonic enforcement and clipping, and under an alternative perturbation strategy? Please explain the handling of near-zero or reversed clean-to-perturbed endpoint differences and provide paired uncertainty estimates for the main AUC comparisons. How much of the joint-setting advantage remains if a strong baseline receives the same modality calibration?
-
Under matched training and token-weighting budgets, how does VTRACE-guided GRPO compare with random token selection and selection using a strong existing attribution method? Can repeated training runs establish the uncertainty of the reported gain and clarify which tasks benefit consistently?
6 comments
Response to Reviewer QDRz Part 2
Q2c. An ablation using outgoing attribution alone would also help determine what the incoming-flow factor adds beyond path aggregation.
Following the reviewer's suggestion, we evaluated VTrace using outgoing attribution alone. The gains over baselines come from path aggregation, while the incoming factor removes a position bias that matters for reasoning tokens.
VTrace scores token by , where sums paths from to the answer and sums paths arriving at . Outgoing attribution alone favours early tokens regardless of content because more tokens follow them and more paths originate from them. On a complete causal graph with equal edge weights and the answer at position , grows exponentially towards the sequence start. By contrast, at every position, cancelling this bias.
Input tokens. Joint-setting results are reported as insertion ↑ / deletion ↓:
| Readout | Metric | MMStar | MathVista | MMMU |
|---|---|---|---|---|
| Outgoing only | RISE | 0.583 / 0.182 | 0.647 / 0.177 | 0.660 / 0.160 |
| MAS | 0.398 / 0.247 | 0.496 / 0.244 | 0.500 / 0.215 | |
| In & Out | RISE | 0.581 / 0.180 | 0.647 / 0.178 | 0.660 / 0.159 |
| MAS | 0.397 / 0.246 | 0.497 / 0.246 | 0.503 / 0.215 |
The readouts differ by at most 0.006, as expected: input tokens occur near the sequence start, with few preceding tokens and little inflow. The gains in Table 1 therefore come from path aggregation.
Reasoning tokens. Inflow matters here because reasoning tokens differ in how much visual evidence reaches them. In the Q2b intervention, replacing tokens selected by outflow alone changes the answer about half as often as replacing tokens selected with the incoming factor:
| Five tokens replaced | MMStar | MathVista |
|---|---|---|
| Highest outflow only | 5.5% | 9.0% |
| Inflow × outflow | 10.1% | 20.3% |
Thus, the incoming factor identifies intermediate tokens that carry visual evidence to the answer.
W3 & Q3a. Robustness to the perturbation and metric-processing choices is insufficiently established. ... Do the main rankings persist without monotonic enforcement and clipping?
VTrace remains the best-performing method without monotonic enforcement or clipping.
These transformations reduce noise: adding patches can temporarily lower response likelihood, so the monotonic envelope smooths fluctuations, while clipping keeps normalised curves within . Since both can affect rankings, we removed all three transformations: (1) response-curve clipping to , (2) monotonic enforcement, and (3) MAS-penalised curve clipping to .
Each cell reports VTrace / strongest baseline; raw results also include the paired 95% confidence interval for their difference.
RISE:
| Benchmark | Protocol | Image del ↓ | Image ins ↑ | Joint del ↓ | Joint ins ↑ |
|---|---|---|---|---|---|
| MMStar | Original | 0.357 / 0.388 | 0.596 / 0.561 | 0.182 / 0.204 | 0.576 / 0.550 |
| Raw | 0.369 / 0.399 [−0.035, −0.025] | 0.583 / 0.547 [+0.030, +0.042] | 0.184 / 0.206 [−0.024, −0.020] | 0.571 / 0.545 [+0.023, +0.030] | |
| MathVista | Original | 0.312 / 0.354 | 0.658 / 0.598 | 0.179 / 0.243 | 0.643 / 0.570 |
| Raw | 0.327 / 0.368 [−0.048, −0.035] | 0.644 / 0.583 [+0.054, +0.068] | 0.183 / 0.243 [−0.065, −0.057] | 0.637 / 0.564 [+0.068, +0.079] | |
| MMMU | Original | 0.389 / 0.424 | 0.657 / 0.615 | 0.165 / 0.185 | 0.643 / 0.612 |
| Raw | 0.408 / 0.444 [−0.044, −0.027] | 0.639 / 0.597 [+0.032, +0.054] | 0.168 / 0.186 [−0.022, −0.015] | 0.639 / 0.608 [+0.026, +0.037] |
MAS:
| Benchmark | Protocol | Image del ↓ | Image ins ↑ | Joint del ↓ | Joint ins ↑ |
|---|---|---|---|---|---|
| MMStar | Original | 0.494 / 0.532 | 0.455 / 0.408 | 0.248 / 0.333 | 0.390 / 0.345 |
| Raw | 0.603 / 0.667 [−0.073, −0.055] | 0.330 / 0.244 [+0.075, +0.096] | 0.254 / 0.353 [−0.104, −0.094] | 0.299 / 0.150 [+0.141, +0.158] | |
| MathVista | Original | 0.447 / 0.482 | 0.530 / 0.460 | 0.247 / 0.381 | 0.492 / 0.373 |
| Raw | 0.528 / 0.587 [−0.072, −0.046] | 0.450 / 0.330 [+0.106, +0.134] | 0.256 / 0.413 [−0.164, −0.149] | 0.446 / 0.206 [+0.229, +0.252] | |
| MMMU | Original | 0.534 / 0.571 | 0.527 / 0.487 | 0.223 / 0.302 | 0.480 / 0.420 |
| Raw | 0.697 / 0.761 [−0.079, −0.050] | 0.423 / 0.362 [+0.039, +0.083] | 0.230 / 0.313 [−0.090, −0.076] | 0.418 / 0.288 [+0.113, +0.147] |
Removing the transformations changes individual RISE AUC scores by at most 0.02. MAS changes more because it penalises disagreement between retained attribution mass and model response. This penalty can push curves outside ; removing clipping exposes larger penalties, particularly for baselines. VTrace's advantage remains under both metrics.
Response to Reviewer QDRz Part 3
Q3b. Do the main rankings persist under an alternative perturbation strategy? ... Modality calibration uses padding embeddings, whereas visual evaluation uses blurred image regions.
Yes. With an alternative perturbation, VTrace still ranks first on every score.
Following the reviewer's suggestion, we replaced the blur with a mean colour fill: each removed patch is filled with the image's mean colour, and the image is encoded again. This removes all content of a patch, including the coarse structure that blur keeps. Attribution and calibration are unchanged.
Due to tight turnaround, we evaluated on a subset of MMStar traces (ins$\uparrow$ / del$\downarrow$):
| Setting | Metric | IFR | FlowTracer | FlashTrace | AttnLRP | VTrace |
|---|---|---|---|---|---|---|
| Image | RISE | 0.557 / 0.414 | 0.548 / 0.420 | 0.573 / 0.398 | 0.582 / 0.393 | 0.617 / 0.361 |
| MAS | 0.399 / 0.556 | 0.381 / 0.572 | 0.427 / 0.532 | 0.411 / 0.559 | 0.476 / 0.499 | |
| Joint | RISE | 0.515 / 0.224 | 0.514 / 0.230 | 0.561 / 0.207 | 0.538 / 0.222 | 0.591 / 0.182 |
| MAS | 0.315 / 0.368 | 0.308 / 0.379 | 0.358 / 0.338 | 0.326 / 0.347 | 0.410 / 0.250 |
Results: VTrace is first in all eight comparisons, with margins similar to those under blur (e.g., Image RISE insertion 0.617 vs. 0.582 here, and 0.600 vs. 0.563 under blur).
Why does calibration use padding embeddings while the evaluation uses blur?
- Calibration removes a whole modality to measure its total contribution, padding embeddings do this for image and text in the same way, within three forward passes and without encoding the image again.
- The evaluation removes individual patches, which is done in pixel space so that the vision encoder processes the perturbed image.
Does this mismatch matter? We recomputed the calibration from the blurred image instead of the padding embeddings. The resulting image share ranks traces almost identically:
| MMStar | MathVista | MMMU | |
|---|---|---|---|
| Spearman correlation of the image share (padding vs. blur) | 0.92 | 0.95 | 0.92 |
Thus, VTrace still ranks first under a different evaluation perturbation, and this shows that calibration is not tuned to the perturbation used in the evaluation.
Q3c. Please explain the handling of near-zero or reversed clean-to-perturbed endpoint differences.
Thank you for your attention to detail. We highlight why this is not a problem for VTrace as follows:
- They are rare (0.16% of the samples measured).
- The endpoints are the clean and the fully perturbed input, which are the same for every method on a trace. No method receives a more favourable denominator.
- VTrace performs best no matter if these curves are kept (Table 1) or excluded (Q3a).
How often? We counted the curves with on 2,736 samples:
| Count | |
|---|---|
| All curves (traces x settings x insertion and deletion) | 10942 |
| Near zero endpoint difference | 18 (0.16%) |
| of which reversed | 3 |
How do we handle them?
- Main Table 1 protocol: these curves are kept. A small constant in the denominator avoids division by zero, and clipping to keeps their normalised values bounded.
- Raw curve evaluation (Q3a): there is no clipping, so a near zero denominator would produce extreme values. We exclude these curves for every method.
Response to Reviewer QDRz Part 1
We thank the reviewer for the careful reading and concrete suggestions. We clarify your concerns below.
Scoring only the final answer under teacher forcing still conditions on the original reasoning tokens, which may already contain the answer or the visual evidence needed to derive it. Following prior work, we evaluate the complete generated response to assess support for both the reasoning and the answer, not just the answer given that reasoning. This protocol measures faithfulness to a fixed response; it does not establish answer-level dependence when reasoning is regenerated.
Yes. With regenerated reasoning, VTrace maintains the superiority against baselines. We kindly invite the reviewer to review our response to Reviewer rsBr's "W2b & Q2b" for detailed experiment results, where we evaluate VTrace under two complementary interventions of regeneration.
Since W aggregates contributions across layers, its paths represent composed attribution relationships among tokens. The purpose is to identify intermediate reasoning tokens through which input evidence supports the answer. To directly assess their behavioral relevance, we replace five reasoning tokens with the highest image-to-answer path scores: the answer changes on 10.1% of MMStar and 20.3% of MathVista traces, compared with 5.7% and 9.5% for the strongest of six baselines (Q2b). These interventions support the effectiveness of the path scores in identifying intermediate tokens that are relevant to the answer. We also isolate the incoming-flow factor: using outgoing attribution alone changes input-attribution RISE/MAS scores by at most 0.006 (Q2c), showing that the main gains are not driven by this factor. We will clarify that these paths capture token-level attribution relationships, and our interventions show that the identified intermediate tokens influence the final answer.
Following the reviewer's suggestion, we intervened on the intermediate reasoning tokens that VTrace scores highest, and compared them with matched,low scoring and random controls.
Which tokens? For each trace, we take the five reasoning tokens with the highest relay score, i.e., the image evidence flowing into the token times the evidence flowing from it to the answer.
Setup: We keep the first half of the original reasoning and blur the image, so the model must rely on what the reasoning has written. We replace five tokens in that half with the model's next most likely alternatives, and let the model regenerate the rest of the reasoning and the answer. We report how often the final answer changes, on the traces where the unmodified first half still gives the original answer.
What does each control rule out?
Results: On MMStar and MathVista, replacing VTrace's top relays changes the answer about twice as often as the strongest control, with MMMU showing a scenario where the answer rarely changes for any selection (at most 2.7%).
Thus, the high scoring intermediate tokens identified by VTrace are the ones the final answer depends on.