PaperlayerView exact revision on arXiv

Source-linked AI summary

When to Rethink: Learning Multi-Perspective Self-Verification for Vision-Language Models

Ziquan Zhu, Hanruo Zhu, Si-Yuan Lu, Morris Yu-Chao Huang, Yicheng Lin, Wei Han, Tianlong Chen, Mingyuan Wu, Hanchao Yu, Gaojie Jin, Lu Liu, Bo Sun, Tianjin Huang

arXiv:2610.07018v1cs.AI

TL;DR

VLM self-verification is constrained by reliance on a single criterion or fixed prompt, even though verifier capability and prompt choice affect reliability. MOTIVE learns a correctness-aligned score from complementary verification views and uses it to guide accept-or-rethink decisions. Experiments report that it outperforms strong self-verification and self-correction baselines across multimodal benchmarks and VLM backbones.

  • Problem

    Existing self-verification methods typically use a single criterion or fixed prompt, leaving the challenge of obtaining reliable, comprehensive verification signals unresolved.

    Source: §1 · pp. 1–2

  • Method

    MOTIVE aggregates complementary verification views, supervises view-level and aggregate reliability estimates against candidate correctness, and uses the learned score for accept-or-rethink decisions.

    Source: §1 · p. 2, §3 · p. 6

  • Results

    MOTIVE consistently outperforms strong self-verification and self-correction baselines, with experiments spanning five benchmarks and three VLM backbones.

    Source: §1 · p. 2

  • Takeaways & Limitations

    MOTIVE supports selective revision within the same backbone VLM without requiring an external judge.

    Source: §1 · p. 2, §6 · p. 13

Abstract

from arXiv · show

Vision-language models (VLMs) have achieved strong performance in multimodal reasoning, yet they remain prone to generating plausible but incorrect answers. Self-verification offers a practical way to improve answer reliability without relying on external judges, but existing methods typically depend on a single verification criterion or fixed prompt, resulting in incomplete and unstable reliability estimates. We first systematically analyze how verifier capability and prompt design affect verification performance. Our findings show that stronger verifiers provide more reliable judgments, while verification performance is highly sensitive to prompt choice, with no single prompt consistently dominating across tasks. Guided by these findings, we propose \texttt{MOTIVE}, a \textbf{M}ulti-View Self-Verificati\textbf{O}n wi\textbf{T}h Rel\textbf{I}ability-Guided Selecti\textbf{VE} Rethinking framework for reliable multimodal reasoning. \texttt{MOTIVE} evaluates each candidate answer from complementary verification perspectives and learns a correctness-aligned reliability score through correctness-grounded multi-view verification learning. During inference, this score governs an accept-or-rethink decision, allowing reliable answers to be returned directly while uncertain ones trigger history-guided rethinking. Extensive experiments across diverse multimodal benchmarks and VLM backbones demonstrate that \texttt{MOTIVE} consistently outperforms strong self-verification and self-correction baselines. Further results show that reliable verification improves accept-or-rethink decisions and reduces unnecessary reasoning turns, enabling more reliable and efficient self-verification without an external judge.

1 Introduction

VLM self-verification is limited by single criteria and fixed prompts, while verifier capability and prompt choice affect reliability. MOTIVE combines complementary verification views into a correctness-aligned score that guides accept-or-rethink decisions; experiments report gains over strong baselines.

1 Introduction

  • Stronger verifiers provide more reliable assessments, but verification is prompt-sensitive and no single prompt consistently dominates across tasks.Different prompts have complementary strengths, motivating aggregation of diverse verification signals.Source: §1 · p. 2
  • MOTIVE evaluates candidate answers from complementary views and learns a unified reliability score using correctness-grounded supervision.The supervision aligns both view-specific judgments and the aggregated score with candidate correctness.Source: §1 · p. 2
  • MOTIVE consistently outperforms strong self-verification and self-correction baselines across five benchmarks and three VLM backbones.The experiments assess its performance across diverse tasks and model scales.Source: §1 · p. 2
  • Reliable verification supports earlier stopping with fewer unnecessary rethinking turns, while the best individual view varies across datasets and backbones.The varying best view indicates task-dependent complementarity among verification perspectives.Source: §1 · p. 2

2 Preliminary Analysis

The preliminary analysis finds that verification outcomes depend on judge capability and prompt design. Stronger external judges generally improve final-answer performance, while self-verification prompts produce task-dependent results.

2.1 Verification Capability

  • Larger InternVL3 judges generally achieve better final performance across Qwen backbones and multimodal benchmarks, and verification improves over direct answering.Figure 1 compares accuracy across three benchmarks and Qwen2.5-VL backbones, with dashed lines showing performance without verification.Source: §2.1 · p. 3, Figure 1 · p. 3

2 Preliminary Analysis

  • External verification requires an auxiliary judge and adds computational cost, motivating self-verification within the backbone VLM.The analysis uses the same backbone for answer generation and verification, alongside five direct prompts and three generated-question variants.Source: §2.1 · p. 3, §2.2 · p. 3

2.2 Verification Prompt Design

  • Changing only the self-verification prompt leads to clear performance variation, so an arbitrary checking instruction may not reliably elicit verification.Table 1 reports accuracy for DVP and GVQ prompts on VQA, ScienceQA, and TableVQA with Qwen2.5-VL-3B.Source: §2.2 · p. 3, Table 1 · p. 4

3 Method

MOTIVE jointly optimizes answer generation, correctness-grounded multi-view verification, and selective rethinking. Its learned reliability score guides whether to return a candidate or update the interaction history and generate a revision.

Method overview

  • MOTIVE jointly optimizes answer generation, correctness-grounded multi-view verification, and reliability-guided selective rethinking.The verification signal guides inference-time decisions about whether to accept a candidate or allocate more computation to rethink.Source: §3 · p. 4

Multi-View Verification

  • Each candidate is assessed from complementary verification views, whose reliability estimates are aggregated into one score for selective rethinking.The implementation averages five predefined views motivated by prompt-sensitivity analysis.Source: §3 · p. 5

Correctness-Grounded Multi-View Verification Learning

  • Candidate-correctness labels supervise both view-specific judgments and the aggregated reliability score through binary cross-entropy objectives.View-level supervision encourages each view to provide a correctness-sensitive signal.Source: §3 · p. 6
  • A Brier-style squared-error regularizer constrains the aggregate score’s numerical scale for threshold-based selective inference.It complements, rather than replaces, binary cross-entropy supervision.Source: §3 · p. 6

Reliability-Guided Selective Rethinking

  • MOTIVE accepts a candidate when its score reaches threshold τ or the reasoning budget ends; otherwise, it triggers another rethink step.If no candidate reaches the threshold before the maximum budget, the final candidate is returned.Source: §3 · p. 6
  • Rethinking appends the candidate, verification outputs, and a rethink instruction to history, which then produces a revised candidate.A computation-aware reward penalizes additional rethinking turns to discourage excessive computation.Source: §3 · pp. 6–7

4 Experiments

Experiments evaluate MOTIVE across five visual question answering benchmarks, three VLM backbones, and comparisons with self-verification, self-correction, and vanilla-model baselines. Results assess final-answer accuracy, verification decisions, reasoning efficiency, prompt diversity, and training and aggregation choices.

4 Experiments

  • Experiments use five benchmarks spanning general, structured-image, medical, and scientific VQA, and evaluate three VLM backbones against vanilla, prompt-level, and multimodal self-improvement baselines.Methods share dataset splits, answer-matching protocols, and decoding budgets unless otherwise specified.Source: §4 · pp. 7–8

4.1 Main Results

  • MOTIVE achieves the best performance across five VQA benchmarks and three VLM backbones, outperforming prompt-level and multimodal self-improvement baselines.The comparison covers different model scales and task domains; prompt-level methods improve over vanilla VLMs but provide limited gains.Source: Table 2 · p. 8, §4.1 · p. 8

4.2 Reliability of Accept-or-Rethink Decisions

  • MOTIVE achieves higher Self-Verification Accuracy than conventional Self-Verification across all evaluated benchmarks and backbones.The metric counts accepting a correct candidate or sending an incorrect one for rethinking as a correct decision.Source: Table 3 · p. 9, §4.2 · p. 9

4.3 Effectiveness of Multi-View Verification

  • MOTIVE has the highest and most stable overall accuracy, while the best single verification view varies across datasets and backbones.The comparison includes five single-view variants and full multi-view aggregation.Source: §4.3 · p. 9, Table 4 · p. 10

4.4 Efficient Reasoning via Reliable Early Stopping

  • MOTIVE consistently reduces average reasoning turns relative to Self-Verification across backbones and benchmarks under a maximum budget of T = 3.Turn reductions are more pronounced on VQA and SLaKE, and smaller on ScienceQA.Source: §4.4 · p. 9, Table 5 · p. 10

4.5 Effect of Prompt Redundancy and Diversity

  • Replacing repeated verification prompts with distinct views generally improves performance, and the complete five-view configuration achieves the best overall result despite task-specific fluctuations.The improvement gradually saturates, suggesting that added views retain some shared verification information.Source: §4.5 · p. 10, Figure 3 · p. 11

4.6 Ablation Study

  • View-level verification supervision provides the largest improvement, while aggregate-level supervision and calibration add further gains; combining all three loss components performs best.The losses respectively supervise individual views, align the aggregate reliability score with correctness, and regularize its numerical scale.Source: §4.6 · pp. 10–11, Table 6 · p. 11
  • Averaging continuous YES probabilities outperforms Majority Vote, and λver = 0.1 and βagg = 1.0 yield the best results among the tested weights.Removing verification or aggregate-level supervision lowers performance; larger λver and βagg slightly degrade results, while βcal = 0.1 gives the best calibration-weight result.Source: §4.6 · pp. 11–12, Table 7 · p. 12, Figure 4 · p. 12

5 Related Work

Prior work explores multimodal self-verification and correction at inference time, while verifiable-supervision research uses correctness outcomes as training signals. The paper identifies intermediate verification trajectories and judgments as underused training information.

Self-Verification in Multimodal Reasoning

  • Multimodal self-verification and self-correction use reflection, critique, verifier-guided assessment, or iterative correction, but are often treated mainly as inference-time output refinement.The cited approaches include methods addressing inconsistent predictions and errors from weak visual grounding or unstable multi-step inference.Source: §5 · pp. 12–13

Verifiable Supervision and Training Signals for VLMs

  • Verifiable-supervision methods use automatically checked outcomes for reward learning, process supervision, or intermediate guidance, but typically emphasize final-answer correctness.The paper notes that consistency between round-wise candidate answers and corresponding verification judgments is seldom modeled.Source: §5 · p. 13
  • The paper frames multi-turn candidate answers and verification judgments as training signals that existing methods have insufficiently exploited.This gap concerns intermediate verification trajectories, not only final-answer correctness.Source: §5 · p. 13

6 Conclusion

The conclusion presents MOTIVE as a multi-view framework that learns correctness-aligned reliability and uses it for selective answer revision without an external judge. Experiments span benchmarks covering general, structured-image, medical, and scientific visual question answering.

6 Conclusion

  • MOTIVE aggregates complementary verification signals, learns correctness-aligned reliability with view- and aggregate-level supervision, and uses the score for accept-or-rethink decisions without an external judge.Across five benchmarks and three VLM backbones, experiments show improvements in final-answer performance, verification decision quality, and early stopping.Source: §6 · p. 13

A.1 Additional Dataset Details

  • The benchmark suite covers natural-image VQA, structured charts and tables, radiology-based medical VQA, image-grounded science questions, and table-image reasoning.The experiments use English-language SLaKE questions and the image-based ScienceQA subset.Source: §A.1 · p. 21

A.2 Additional Implementation Details

The implementation fixes training and inference settings for MOTIVE, while its prompts define complementary verification views, a standardized initial answer, and a context-rich rethinking stage. These choices support consistent evaluation and selective refinement of uncertain answers.

A.2 Additional Implementation Details

  • MOTIVE averages YES probabilities from five verification views, uses τ = 0.5 and a maximum of three turns, and trains with GRPO and auxiliary verification objectives.The computation-aware rollout reward is used to compute group-relative advantages.Source: §A.2 · p. 21

B.1 Direct Verification Prompt Templates

  • Five Direct Verification Prompts assess correctness, visual grounding, backward evidence, contextual consistency, and potential errors through binary YES/NO judgments.The prompts provide complementary criteria for evaluating the same candidate answer.Source: §B.1 · p. 22, Figure 7 · p. 22

B.2 Initial Answer Generation Prompt

  • A fixed initial-answer prompt standardizes candidate generation and keeps the answer-generation protocol consistent across experiments.The resulting candidate is evaluated by verification modules and may be refined through rethinking.Source: §B.2 · pp. 22–23, Figure 8 · p. 23

B.3 Rethinking Prompt

  • When verification rejects an answer, the rethinking prompt includes the previous response, image, question, and initial prompt head to encourage revision rather than repetition.This provides a consistent way to refine uncertain answers while avoiding regeneration of accepted responses.Source: §B.3 · pp. 23–24, Figure 9 · p. 23

C.1 Sensitivity Analysis of Acceptance Threshold

MOTIVE’s accuracy remains relatively stable across acceptance thresholds, with τ = 0.5 performing best on all five benchmarks. Lower thresholds risk accepting unreliable answers, while higher thresholds may prompt unnecessary rethinking.

C.1 Sensitivity Analysis of Acceptance Threshold

  • τ = 0.5 achieves the best performance across all five benchmarks, while accuracy remains relatively stable across threshold values.The authors identify this default as a balance between reliable answer acceptance and selective rethinking.Source: §C.1 · p. 24, Figure 10 · p. 24
  • A lower threshold may accept unreliable answers, whereas a higher threshold may introduce unnecessary rethinking.Source: §C.1 · p. 24
  • The learned reliability score provides a decision signal for controlling verification through accept-or-rethink decisions.Source: §C.1 · p. 24

C.2 Computation-Aware Reward Analysis

Increasing the computation-aware reward coefficient reduces rethinking turns while accuracy remains largely stable, and a maximum budget of three turns captures most of the reported performance gains.

C.2 Computation-Aware Reward Analysis

  • Increasing c consistently reduces rethinking turns across benchmarks while final accuracy remains largely stable.Figure 11 reports both accuracy and rethinking turns; c = 0.05 is the selected trade-off.Source: §C.2 · p. 24, Figure 11 · p. 25
  • At c = 0.10, the larger penalty may cause premature stopping and slightly degrade accuracy; c = 0.05 is selected as the trade-off.Source: §C.2 · p. 24

C.3 Maximum Reasoning Turns Analysis

  • From T = 2 to T = 3, performance consistently improves on VQA, ScienceQA, and TableVQA, while T = 4 or T = 5 yields only marginal gains.The authors adopt T = 3 to capture most of the performance benefit without unnecessary extra computation.Source: §C.3 · p. 24, Figure 12 · p. 25

C.4 Sensitivity Analysis of Verification Views

Using more verification views improves accuracy across the reported benchmarks, with five views achieving the best overall performance and diminishing gains beyond three.

C.4 Sensitivity Analysis of Verification Views

C.5 Qualitative Analysis

A TableVQA example shows reliability-guided rethinking correcting successive table-grounding errors until the answer exceeds the acceptance threshold.

C.5 Qualitative Analysis

  • The initial 1.39% answer confuses two plan sections and receives reliability 0.27 < τ, triggering rethinking.Source: §C.5 · p. 26, Figure 13 · p. 27
  • The first rethinking step finds the correct Stock Option Plans section but selects the wrong year, producing 2.52% with reliability 0.44 < τ.Source: §C.5 · p. 26
  • The second rethinking step aligns the section, metric row, and 2004 column, yielding 3.14% with reliability 0.82 > τ and acceptance.Source: §C.5 · p. 26, Figure 13 · p. 27

Related papers

Loading 2610.07018v1…