acceptodds
Under review as a conference paper at ICLR 2027

Reported Safe, Blind in Practice: A Cross-Modal Conflict Benchmark for the Action Layer of Multimodal Agents

Abstract

Multimodal agents increasingly act on values they read from images, so the critical failure is not a wrong caption but a wrong action taken when a textual instruction contradicts the visual evidence. Prior cross-modal conflict studies operate at the answer layer and report modality-preference or "safety" rates, which cannot tell a grounded correction apart from a mere refusal. We introduce an action-layer conflict benchmark of 1,500 programmatically constructed items from 750 TextVQA and ChartQA image-question pairs, split evenly into conflict and matched agreement conditions, and a grounding-certifying metric, Visual Accuracy, that credits an agent only when it overrides the mistaken instruction and reports the correct visible value. Evaluating six state-of-the-art models, we find that Visual Accuracy collapses to at most 0.40 even though blind-follow rates are near zero, that this collapse and its model ordering are preserved across both distributions, and that the high "safety" of standard metrics is overwhelmingly refusal rather than perception. We then build a verification ladder and show that an inference-time locate-read-compare-decide chain significantly improves Visual Accuracy for every model (GPT-5.5: 0.14 to 0.84) and rescues even the hardest model, while we document its cost of trading hollow safety-by-refusal for action. Our benchmark provides a faithful yardstick for multimodal agents that must act on what they see.

Primary area:
datasets and benchmarks
Keywords:
multimodal agents, cross-modal conflict, visual grounding, agent evaluation, datasets and benchmarks

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 4 reviews, above 0% of the others.
Reviewer_yVx8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies how multimodal agents respond when a textual instruction specifies a value that conflicts with visual evidence. Its central argument is that avoiding the instructed wrong action does not establish that an agent can correctly read and use the image. The authors construct 1,500 items from 450 TextVQA and 300 ChartQA image–question pairs, each with a conflict condition and a matched agreement control. They introduce Visual Accuracy (VisAcc), which requires selecting a visual override and reporting a matching value, alongside measures of blind following, escalation, and incorrect overrides. Experiments with six model families show low baseline VisAcc despite generally low rates of blind following.

Critical strengths
  • The paper studies correct execution from merely avoiding an instructed error is useful for evaluating multimodal agents.
  • The paper introduces a cross-modal conflict benchmark and evaluates six model families, revealing poor performance.
Critical weaknesses
  • I strongly suspect that the paper, particularly Figure 1, was generated using AI. Figure 2 still contains explicit placeholder text. It appears that no one thoroughly checked the entire paper.
  • The benchmark only includes conflicts involving words, numbers, or alphanumeric strings. It does not address higher-level semantic conflicts.
  • Table 4’s caption claims a monotonic ordering of Base < TwoStep < CoVe for every model, but Claude’s VisAcc decreases from 0.03 to 0.01 under TwoStep, and Qwen3-VL’s VisAcc decreases from 0.52 under TwoStep to 0.49 under CoVe.
Questions for authors
  • Did the authors review the paper even once?
Reviewer_vnJ5OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces a cross-modal conflict benchmark to evaluate whether multimodal agents correctly ground their actions in visual evidence when textual instructions conflict. However, the submission appears unfinished, with substantial presentation issues, including a blurry, poorly formatted Figure 1; explicit placeholders in Figure 2; and inconsistencies between the reported results and the written claims. These issues significantly undermine the paper's clarity, credibility, and overall quality.

Critical strengths
  1. The paper addresses a relevant problem in multimodal agent reliability, particularly the distinction between refusing conflicting instructions and taking correctly grounded actions.

  2. The evaluation covers multiple multimodal model families and includes both scene-text and chart-based inputs.

Critical weaknesses
  1. The manuscript appears unfinished and has substantial presentation issues. Figure 1 suffers from poor readability and layout problems, while Figure 2 contains explicit placeholders ("placeholder - final schematic rendered on Mac"). There are also unresolved references and inconsistent terminology throughout the manuscript. These problems significantly affect the presentation quality of the submission.

  2. Several numerical results and conclusions are internally inconsistent. The paper repeatedly claims that VisAcc increases monotonically from Base to Two-Step to CoVe across all models. However, Table 4 shows that Claude decreases from 0.03 to 0.01 under Two-Step, while Qwen3-VL decreases from 0.52 to 0.49 when moving from Two-Step to CoVe. Additionally, Table 2 reports BFR = 0.00 and SAFE = 0.80 for Claude on ChartQA, which appears inconsistent with the stated definition SAFE = 1 - BFR. Results and descriptions in the main text and appendix also contain discrepancies. These inconsistencies raise concerns about experimental reliability and require clarification.

Questions for authors
  1. Can the authors explain and correct the inconsistencies in Tables 2, 4, and 5, as well as the discrepancies between the main text and the appendix?

  2. How does the proposed action-layer benchmark fundamentally differ from cross-modal VQA with abstention?

Reviewer_tdmbOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
1(−1)
Presentation
2
Soundness
1(−1)
Read review
Paper summary

This paper observed that many multimodal agents avoid following incorrect instructions by refusing to act rather than by understanding the image correctly. To distinguish them, the authors proposed a benchmark that checks whether agents can correct misleading instructions and act on the actual visual evidence. Empirically, a simple locate–read–compare–decide process substantially improves their performance.

Critical strengths
  • This paper distinguishes visual understanding from mere refusal to follow conflicting instructions and designs quantitative metrics to measure it.

  • The benchmark includes paired conflict and agreement conditions, enabling the evaluation of both grounded correction and unnecessary abstention.

  • The verification interventions and read-then-use analysis provide useful findings on whether failures arise from misreading visual evidence or failing to act on correctly perceived information.

Critical weaknesses
  • Despite the proposed inference-time intervention, the experiments don't show consistent mitigation. For instance, in Tab.4, the simple two-step strategy works better on Qwen3-VL-235B, while a more complex approach, cove works better on the rest. Which solution is more recommended in practice?

  • I'm curious about whether the two proposed interventions introduce substantial inference latency. The paper does not report efficiency metrics, such as latency and additional token usage. Without an evaluation of the performance–efficiency trade-offs, the practical utility of these interventions remains unclear.

  • The paper mentions "hallucination" in the related work section but offers little analysis. For instance, what's the relationship of your study and hallucination? As a well-studied issue, hallucination can be mitigated by over-trust penalty (opera, cvpr'24), contrastive decoding (vcd, cvpr'24) and so on. Do they work for your case?

  • There exist many typos and potential placeholders, which influence the reading experience.

Questions for authors
  • Could you explain the rationale for selecting vision as ground truth? In practice, both visual and textual inputs could contain mistakes, while this paper assumes visual-only grounding is correct.

  • Why does the evaluation and basic dataset focus only on OCR-related VQA? As described, this paper uses TextVQA and ChartQA to construct the benchmark.

  • Typos: (i) L893 says "Table tab:probe", please carefully check what it refers to. (ii) L904–905 says "stays≤0.04" and "Doubao 0.07", which is inconsistent.

  • Figure 2: What's the meaning of the bottom words "[placeholder - final schematic rendered on Mac]"? Did you forget to delete them?

Reviewer_bjAwOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper evaluates cross-modal conflict at the action layer. An agent sees an image and an instruction asserting a wrong value, then chooses (A) follow the instruction, (B) use the image value, or (C) report a conflict. The benchmark has 1,500 items built from 450 TextVQA pairs ("C1") and 300 ChartQA pairs ("C1b"), split into conflict items and agreement controls. The main metric, VisAcc, credits only (B) with the correct value. Six VLMs almost never pick (A) but mostly pick (C), so VisAcc is at most 0.40. The paper reads this as "safety is refusal, not grounding". A four-step locate–read–compare–decide prompt ("CoVe") raises VisAcc, e.g. GPT-5.5 from 0.14 to 0.84.

Critical strengths
  1. Reusable setup. Matched conflict/agreement controls and deterministic per-image seeding make the benchmark easy to reproduce.
  2. Sound statistical testing. Intervention gains are tested with paired-bootstrap 95% CIs over the same items
Critical weaknesses
  1. Refusal is a tunable policy, not a lack of grounding, and the paper never tests the baseline that would tune it.
    • Escalating is a sensible default. Flagging a conflict instead of guessing is how models are trained to avoid hallucinated or incorrect answers. When text and image disagree, the agent cannot know which is wrong, and a wrong silent correction acts on a bad value and removes the alarm.
    • Which source wins is a deployment choice, not a measure of quality. The paper hard-codes "the image is right". In many settings the text should win: an instruction carrying a newer value makes a chart or screenshot stale. Treating (C) as "safe but useless"and crediting only (B) scores models against one arbitrary policy.
    • The missing baseline is a policy prompt. If escalation is a policy, default internalized by the model, the first test is to state the policy: give the agent a role in which it is authorized to override the text using the image, and measure VisAcc. The paper skips this. It goes straight to a four-step CoVe scaffold. Without that baseline, there is no evidence that the four steps are needed, or that CoVe "recovers a latent ability" (Sec. 5.1) rather than simply telling the model what policy it should be following.
    • The reported metrics do not support "CoVe improves every model". Blind-follow under CoVe is never reported. The reader can only infer it from SAFE = 1 − BFR : Gemini's SAFE falls from 1.00 to 0.54 (Sec. 5.3), which implies BFR ≈ 0.46, about as large as its VisAcc of 0.49.
  • SAFE also counts wrong overrides (FakeOverride) as safe. A VisAcc gain alone therefore cannot establish an improvement; BFR and FakeOverride under each condition are needed
  1. The dataset is stale and narrow: one image type, one question type.

    • Old sources. TextVQA [1] (2019) is largely OCR-solvable and has ambiguous answers; harder text-rich benchmarks such as ConTextual [3] were built to address this. ChartQA [2] overstates chart understanding, as CharXiv [4] shows.

    • Narrow image scenarios. Only scene-text photos and charts are used. There are no documents (DocVQA [5]), screens/GUIs (ScreenSpot [6]), natural scenes without text, or real-world noisy captures (VizWiz [7]; MME-RealWorld [8]), although the motivation is agents reading "a screen, a gauge, or a chart" (Sec. 1).

    • Narrow question types. Every conflict is a single extracted string. There are no conflicts over:

      • counting (TallyQA [9]);
      • spatial relations (What'sUp [10]);
      • colour, attributes and fine detail (V* Bench [11]);
      • low-level perception (MMVP [12]; CV-Bench [13]; BLINK [14]);
      • GUI state [6].

      The "difficulty stratification" (Table 6) stratifies by answer type (word/number/alphanumeric), not question type, and the cross-distribution claim (Sec. 4.2) varies only the image source.

    • Small scale despite trivial generation. Items are a template plus a sampled wrong value, yet the paper uses only 750 pairs. The gold audit is 70 items checked by Gemini-2.5-Pro, itself an evaluated model.

  2. The writing and reporting make the paper hard to follow and hard to trust.

    • Key figures are missing. Fig. 1 is incomplete / illegible and Fig2 has placeholders.
    • Too many metrics defined that obscure the reader from understanding how VisAcc works, their key proposal. Interplay between the metrics should also be mentioned up front. For instance VoR = VisAcc + FakeOverride, because at place The above relation is not followed. (Sec. 4.3 (0.16/0.16/0.14) ≠ VOR − VisAcc from Table 2 (0.02/0.05/0.05).
    • Inconsistencies in results, like Table 5 and Table 4 report different VisAcc for the same CoVe condition (Gemini 0.86 vs. 0.49). ?

References [1] Singh et al., "Towards VQA Models That Can Read," CVPR 2019. [2] Masry et al., "ChartQA: A Benchmark for Question Answering about Charts with Visual and Logical Reasoning," Findings of ACL 2022. [3] Wadhawan et al., "ConTextual: Evaluating Context-Sensitive Text-Rich Visual Reasoning in Large Multimodal Models," ICML 2024. [4] Wang et al., "CharXiv: Charting Gaps in Realistic Chart Understanding in Multimodal LLMs," NeurIPS 2024 (Datasets & Benchmarks). [5] Mathew et al., "DocVQA: A Dataset for VQA on Document Images," WACV 2021. [6] Cheng et al., "SeeClick: Harnessing GUI Grounding for Advanced Visual GUI Agents," ACL 2024 (ScreenSpot). [7] Gurari et al., "VizWiz Grand Challenge: Answering Visual Questions from Blind People," CVPR 2018. [8] Zhang et al., "MME-RealWorld: Could Your Multimodal LLM Challenge High-Resolution Real-World Scenarios that are Difficult for Humans?," ICLR 2025. [9] Acharya et al., "TallyQA: Answering Complex Counting Questions," AAAI 2019. [10] Kamath et al., "What's 'up' with Vision-Language Models? Investigating Their Struggle with Spatial Reasoning," EMNLP 2023. [11] Wu & Xie, "V*: Guided Visual Search as a Core Mechanism in Multimodal LLMs," CVPR 2024. [12] Tong et al., "Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs," CVPR 2024. [13] Tong et al., "Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs," NeurIPS 2024 (CV-Bench). [14] Fu et al., "BLINK: Multimodal Large Language Models Can See but Not Perceive," ECCV 2024.

Questions for authors
  1. With a base prompt or role that explicitly authorizes overriding the text when the image disagrees, how much of the CoVe gain in VisAcc remains?
  2. Please report BFR and FakeOverride for every model under base, TwoStep and CoVe.
  3. How were the 450/300 pairs selected from the validation sets?
  4. Do the findings hold for other image scenarios (documents, GUIs, natural scenes) and question types (count, colour, spatial, comparison)?
  5. Please complete the figures 1, 2 and other places which are with placeholders

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.