acceptodds
Under review as a conference paper at ICLR 2027

Visual Evidence Under Cross-Examination: Evaluating and Controlling Decision-Level Evidence Use in Vision-Language Models

Abstract

Vision-language models increasingly reason through crops, regions, and tool-produced observations. Yet an observation can influence the answer without benefiting the candidate it supports. We study candidate-bound visual contribution: valid evidence should help, invalidating its supporting relation should remove its additional effect, and valid rebinding should redirect that effect to the newly supported candidate. We introduce CROSS-Bench, a benchmark of 28,000 decision problems with matched invalidation and rebinding tests. Our RIVET interface preserves evidence identity and uncertainty, composes a candidate-conditioned response, and separately controls its strength. Shared-evidence experiments show that task accuracy and evidence ownership can diverge. Under matched capacity and training, RIVET increases normalized effect transfer from 0.512 to 0.651 where clean evidence has a positive effect. The advantage persists on common evaluation examples and across repeated decision-layer fits. With evidence predicted from raw inputs, RIVET improves CROSS-Bench accuracy by an average of 5.70 pp across four frozen backbones, relative to the same models without auxiliary evidence. These results separate the utility of visual evidence from the candidate-specific destination of its effect.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Vision-Language Models, Visual Reasoning, Counterfactual Evaluation, Evidence Attribution, Faithfulness

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NfbROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether visual evidence in vision-language models contributes to the answer candidate it actually supports, introducing the concept of candidate-bound visual contribution. The authors construct CROSS-Bench, a benchmark comprising 28,000 decision problems with matched evidence invalidation and rebinding tests, to evaluate evidence utility, specificity, and decision ownership. They further propose RIVET, a structured evidence-to-decision interface that preserves evidence identity, semantic bindings, and uncertainty through candidate-conditioned composition and scalar response control. Experiments demonstrate that prediction accuracy and evidence ownership can diverge. Compared with a capacity-matched baseline, RIVET improves normalized evidence-effect transfer from 0.512 to 0.651, while achieving an average accuracy improvement of 5.70 percentage points over frozen base models on CROSS-Bench. Additional evaluations on external VQA benchmarks demonstrate improvements in task accuracy.

Critical strengths
  1. The paper introduces a meaningful and underexplored perspective on visual evidence utilization. By distinguishing evidence utility from decision ownership, the work identifies a limitation of conventional accuracy-based evaluations: a model may produce correct predictions without assigning evidence contributions to the candidates they actually support. The proposed invalidation and rebinding criteria provide a useful framework for examining this issue.
  2. CROSS-Bench provides a structured evaluation framework specifically designed for evidence ownership. The benchmark covers eight visual evidence families and constructs matched clean, invalid, and valid-rebinding conditions. Its explicit source–entity–role–state relations enable more targeted evaluation of evidence contributions than conventional VQA benchmarks.
  3. The proposed architecture is well structured and supported by informative experimental controls. RIVET explicitly preserves evidence identity, semantic bindings, candidate alignment, and uncertainty, while separating evidence composition from response-strength control. Capacity-matched comparisons, component ablations, cross-backbone experiments, and evaluations on independent benchmarks provide useful evidence supporting the proposed design.
Critical weaknesses
  1. The benchmark's applicability is constrained by its predefined evidence structure and evaluation scope. Although CROSS-Bench contains 28,000 problems, the primary evidence-ownership evaluation relies on a subset of 960 examples, of which only 707 support valid rebinding. The benchmark requires explicit, annotatable source–entity–role–state relations and a fixed set of 2–5 answer candidates.
  2. The paper demonstrates a dissociation between accuracy and ownership but provides limited insight into its underlying causes. The experiments show that highly accurate fusion models may exhibit poor evidence ownership. However, the analysis does not fully explain why such predictions remain correct, how much they depend on base-model priors versus misallocated evidence, or under which conditions incorrect ownership causes actual prediction failures. This limits the mechanistic understanding of the central observation.
  3. The practical benefits of improving evidence ownership are not sufficiently demonstrated. While RIVET achieves higher transfer scores, its matrix accuracy is slightly lower than Capacity-matched Fusion (83.33% versus 84.27%) and Unconstrained Fusion (84.58%). The paper does not sufficiently establish whether improved ownership translates into better robustness, reliability, or reduced reasoning errors in realistic downstream applications.
Questions for authors
  1. Mechanistic explanation of the accuracy–ownership dissociation. The paper shows that some models achieve high answer accuracy despite poor evidence ownership. However, what mechanisms actually enable these correct predictions? Can the authors distinguish the contributions of base-model priors, alternative visual cues, and incorrectly assigned evidence to explain when and why accuracy becomes disconnected from evidence ownership?
  2. Practical significance of improving evidence ownership. Beyond improving the proposed ownership metrics, what concrete benefits does correct evidence ownership provide in real-world VQA applications? Could the authors demonstrate that higher ownership leads to fewer evidence-grounding errors, stronger robustness under misleading visual information, or more reliable reasoning on distribution-shifted examples?
  3. Computational efficiency and inference overhead. RIVET introduces several additional stages, including evidence acquisition, binding, specialized visual reading, and ensemble-based composition. What are the additional end-to-end inference latency, GPU memory consumption, and computational costs compared with the frozen Base and capacity-matched fusion baselines?
  4. Necessity of the scalar Bound controller. The ablation results show that removing Bound increases evidence transfer from 0.651 to 0.676 and reduces original-owner leakage from 0.205 to 0.194, although accuracy decreases slightly. Could the authors further clarify why the learned scalar controller is necessary and whether its accuracy improvements justify the degradation in the primary ownership metrics?
Reviewer_ti6VOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a benchmark about the visual evidences and the results they support. They claim that the VLM models might wrongly leverage the visual evidences in unrelated questions. They also build a benchmark to test this phenomenon and find VLM models have this issue. They propose RIVET, a pipeline leveraging different model to solve this issue. RIVET defines a pipeline to explicitly detect the objects and the binding relationship, which mitigates the issue of current VLM models.

Critical strengths
  • This paper raises the problem that current VLM models might leverage wrong visual evidences to reason the answer.
  • The authors build a benchmark to test it with different models. They also design a pipeline to mitigate this issue by leveraging different models.
Critical weaknesses
  • The pipeline of RIVET is quite task specific and not really general. It is designed to solve specific tasks whose reasoning progress can be serialized as detecting the object, then recognize the objects and finally map the results to the answers. However, this pipeline isn't quite general to different tasks. Only a small set of tasks can be solved in this way. I'm not quite convinced this is a scalable pipeline that can be used for the next generation of VLMs.
  • Human decisions also usually rely on background information. For example, if we know the two women are twins and they wear the same clothes, then we may guess they wear scarfs with the same color. My concern is the metrics designed in the CROSS-Bench might be too simple to assess a VLM.
Questions for authors
  • Can you explain how you expect RIVET to play its role for the future VLM design?
  • As the second point in weaknesses, I'm wondering if a model works better on CROSS-Bench really achieves better performance in general. Can you show the correlation of performance on CROSS-Bench and other grounding benchmarks?
1 comment

Thank the authors for their detailed response. I appreciate the time and effort they have devoted to addressing my concerns.

  • "The transferable idea is that a visual observation should affect the entity or claim supported by its relation, rather than merely becoming available in a pooled visual representation."

I agree with this intuition. However, similar conclusions can be found in previous works [1] [2], and I think the intuition itself can't be hold as a contribution. Also, the role of RIVET is still unclear. Directly leveraging RIVET as the teacher model is basically amortizing this pipeline into a neural network. However, I still don't think RIVET is general and robust enough as a caption model.

  • "Across eight frozen checkpoints, CROSS-Bench answer accuracy shows a stronger rank association with external grounding performance than ownership transfer does on both evaluated benchmarks."

I think 0.643 and 0.548 can't show a strong correlation with only 8 data points. With these two numbers I'm not quite convinced that CROSS-Bench could play an important role for VLM evaluation.

Another suggestion is about the presentation. The current Figure 3 is quite complicated and hard to understand. A good paper should be easy to read and understand. Current figures are quite fancy, but really hard for an audience to read. Also, some terminologies require more illustration and examples (say utility, specificity and ownership) otherwise it would only increase the cognitive load of the reader.

Overall, I appreciate the efforts from the authors, but I will maintain my rating as 2.

[1] Yuxiao Qu et al. “POPE: Learning to Reason on Hard Problems via Privileged On-Policy Exploration.” in TMLR 2026

[2] Zhaoyang Li et al. "ORIC: Benchmarking Object Recognition under Contextual Incongruity in Large Vision-Language Models" in CVPR 2026

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

This paper purpose an important problem in vlm that correctly recognizing visual evidence does not necessarily imply that its contribution is assigned to the appropriate answer candidate. To deal with problem, the author first purpose a new benchmark designed to evaluate decision ownership performance on VLMs. Furthermore, the author proposes a new framework called RIVET, adopting a four stages of specify, preserve, compose and bound. Experiments across several frozen VLM backbones demonstrate improvements in task accuracy over the corresponding models without auxiliary evidence.

Critical strengths

The innovation of the topic is great. This paper highlights the distinction between correctly recognizing visual evidence and assigning its contribution to the correct entity. The route to solve such problem is also concrete. The author first purpose a benchmark demonstrating such problem reside in recent VLMs. Then purpose a novel 4 stage framework to deal with such problem. The experiment part is also comprehensive.

Critical weaknesses

My first concern raise from the solution novelty. Although the part of finding problem and building benchmark is great, the framework of solution remain likes combination of external tools, while nowadays with the development of agent and harness, I believe the correct entity problem could also and better be done by harness framework using external tools instead of training a new model. Secondly, I find the framework have quite dedicate process and stages, I think it may fail to some long tail cases. I hope the author could study about it more and show some cases. Lastly, the solution is also only able to solve close form problem like selection, which is quite limited.

Questions for authors
  1. Could the method scale to open-end problems?
  2. What the difference and pros and cons of the purposed framework and using agent/harness to solve such problem?
  3. Could the authors further distinguish whether the ownership errors originate from inaccurate visual grounding or incorrect decision-level evidence integration?
1 comment

Thank you for the detailed and convincing responses. I find that my main concerns have been adequately addressed, and the additional evidence strengthens the paper’s central claims. I will therefore maintain my rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.