PaperlayerView exact revision on arXiv

Source-linked AI summary

Think, Look, and Revise: Inconsistency-Aware Visual Self-Correction in MLLMs

Yu Cheng, Arushi Goel, Hakan Bilen

arXiv:2608.29374v1cs.CV

TL;DR

Tool-augmented multimodal reasoning often uses tool outputs without checking them, leaving limited support for detecting and correcting unreliable evidence. ReVISE trains models to verify and revise tool-derived evidence, and improves performance across perception-intensive benchmarks.

  • Problem

    Existing tool-augmented MLLMs rarely evaluate and revise unreliable tool outputs, while training data and rewards provide limited supervision for verification and correction.

    Source: §1 · p. 2

  • Method

    ReVISE trains MLLMs to check tool observations against the visual query and their own perception, using reflection-supervised trajectories and targeted rewards for self-correction and spatial grounding.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    ReVISE improves over base Qwen2.5-VL and its SFT and GRPO variants across perception-intensive benchmarks; versus Qwen2.5-VL-3B-Instruct, gains are 5.86% on CountBench, 11.28% on CVBench, and 9.94% on BLINK.

    Source: §5.2 · p. 6

  • Takeaways & Limitations

    The findings highlight the importance of error detection and correction in tool-augmented multimodal reasoning.

    Source: §6 · p. 8

    ReVISE relies on synthetic trajectories and bounding boxes from proprietary models and visually selective tools, which may limit robustness in open-domain scenarios.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Tool-augmented multimodal reasoning integrates external tools (e.g., object detection, depth estimation) into multimodal large language models (MLLMs) to address perceptual bottlenecks in complex visual tasks. However, existing approaches rarely verify tool outputs, limiting their ability to detect and recover from tool failures. We propose ReVISE, a framework that equips MLLMs with verification and dynamic error recovery for tool-augmented reasoning. ReVISE introduces (1) a curated training dataset that supervises reflective behaviors, enabling models to validate tool-derived evidence, reformulate queries when visual mismatches arise, and fall back to intrinsic grounding when external tools are unreliable; and (2) a reinforcement learning based targeted rewards that encourage internal reflection and penalize spatial misalignment. Experiments on several benchmarks demonstrate consistent improvements over existing methods, highlighting the importance of error detection and correction in tool-augmented multimodal reasoning.

1 Introduction

ReVISE addresses open-loop tool-augmented multimodal reasoning by teaching models to verify tool outputs against visual context and correct unreliable evidence. Its training combines reflection-supervised tool interactions with correction-oriented reinforcement learning.

Introduction

  • Existing tool-augmented methods use multi-step tool interactions but propagate outputs without verification, leaving models unable to detect mismatches or recover reliably.The paper attributes this open-loop behavior to limited SFT supervision for failure correction, RL objectives focused on final-answer accuracy, and tool over-reliance that weakens intrinsic visual reasoning.Source: §1 · pp. 1–2
  • ReVISE validates intermediate tool observations against direct image interpretation, then triggers self-correction when feedback is noisy, contradictory, or missing.It can reject noisy observations and refine tool execution before producing a final answer.Source: §1 · p. 2
  • ReVISE builds multi-turn verification and revision trajectories from SAT (Ray et al., 2024), TACO (Ma et al., 2025), and TallyQA (Acharya et al., 2019), then trains with GRPO and correction-oriented rewards.The rewards include a self-correction signal for effective revision and a grounding signal for spatial bounding-box refinement.Source: §1 · p. 2
  • ReVISE consistently improves Qwen2.5-VL performance on CountBench (Paiss et al., 2023), CVBench, and BLINK-HARD (Bigverdi et al., 2025).The paper presents these as challenging perception-heavy benchmarks.Source: §1 · p. 2

2 Related Work

MLLMs have advanced through instruction tuning and post-training, while tool-augmented visual reasoning addresses perceptual limitations. Existing approaches generally do not train models to revise reasoning when tool evidence contradicts initial predictions.

Multimodal Large Language Models

  • MLLMs combine image and text reasoning, and instruction-tuned systems such as LLaVA, LLaVA-OneVision, and NVILA improve diverse visual benchmark performance through scaled multimodal instruction data.The cited systems are LLaVA family (Liu et al., 2023, 2024b), LLaVA-OneVision (Li et al., 2024), and NVILA (Liu et al., 2025).Source: §2 · p. 2
  • MLLMs typically progress from pretraining to supervised fine-tuning and then RL or preference-based post-training, yet still struggle with precise visual grounding and evidence verification.DPO (Rafailov et al., 2023) and GRPO (Shao et al., 2024) are efficient alternatives to traditional RL pipelines for optimizing reasoning behavior.Source: §2 · p. 2

Tool-Augmented Visual Reasoning

  • Tool-augmented methods use compositional programs, learned tool invocation, distillation, and RL, but typically seek correct answers without modeling adaptation to contradictory tool evidence.Examples include ViperGPT and VisProg (Surís et al., 2023; Gupta and Kembhavi, 2023), learning-based invocation (Wang et al., 2025a; Zhou et al., 2025; Song et al., 2025; Hong et al., 2025; Wang et al., 2025b; Su et al., 2025), distillation (Hu et al., 2024), and RL for tool selection and reasoning (Zhao et al., 2025; Zhou et al., 2025; Zhang et al., 2025; Wang et al., 2025a; Song et al., 2025).Source: §2 · p. 3
  • Unlike iterative methods that rely on inference-time heuristics, this work supervises visual self-correction and object-level grounding when new evidence contradicts an initial hypothesis.CodeDance (Song et al., 2025) and AdaTooler-V (Wang et al., 2025a) revisit visual evidence, but do not explicitly supervise correction-oriented reasoning during training.Source: §2 · p. 3

3 Method

ReVISE frames tool-augmented multimodal reasoning as sequential decisions and trains a policy to verify tool-derived observations against the query and its own perception, then revise mismatches. Its method combines supervised tool-use and self-correction trajectories with reinforcement-learning rewards for reflection and spatial grounding.

3.1 Problem Formulation

  • At each step, the model uses the visual input, query, and trajectory history to choose a tool invocation or final answer, with tool observations updating the state.The process ends when the model answers or reaches the maximum step T.Source: §3.1 · p. 3

3.2 Cold-Start Supervised Fine-Tuning

  • Supervised fine-tuning on curated trajectories teaches tool use, verification, and self-correction, including detecting and revising inconsistent evidence, through next-token prediction over action sequences.The paper states that minimizing the supervised loss gives the model cold-start capability for sequential visual tool use and self-correction.Source: §3.2 · pp. 3–4

3.3 Policy Optimization via Reflection and Grounding Rewards

  • GRPO combines answer-accuracy, format, self-correction, and grounding rewards; grounding rewards mean IoU and positive localization improvement, while reflection is rewarded only when the final answer is correct.Format reward is 1 for prescribed tags and −1 otherwise; the self-correction reward is −1 when the model emits re-examine but answers incorrectly, and 0 otherwise. The improvement reward is 0 if no boxes are produced or updated.Source: §3.3 · p. 4

4 Training Data Curation

ReVISE’s reflection-rich supervised dataset teaches models to verify tool evidence, revise mismatches, and fall back to intrinsic perception. It combines tool-use and tool-free reasoning trajectories.

Toolset Collection

  • The tool suite combines an open-vocabulary detector, ZoomIn crop-and-resize, edge detection, and depth estimation for grounding, inspection, boundaries, and geometric reasoning.The four operations target MLLMs’ perceptual bottlenecks.Source: §4 · p. 4

Trajectory Design

  • Five trajectory types supervise positive and negative tool verification, tool retry, and positive and negative tool-free verification, including query revision and intrinsic fallback.Models detect mismatches between tool outputs and queried concepts, then refine tool arguments, switch tools, or rely on their own visual inspection.Source: §4 · pp. 4–5, Figure 2 · p. 5

Trajectory Generation and Training

  • Reasoning traces are synthesized with Gemini-2.5-pro using prompts tailored to each trajectory type, then manually filtered for errors.The synthesis process covers the trajectory types designed for cold-start supervision.Source: §4 · p. 5

5 Experiments

Across the evaluated benchmarks, ReVISE improves perception-intensive performance through explicit grounding, tool verification, and self-correction, while results are weaker on knowledge-intensive tasks. Ablations and qualitative examples show benefits from reflection data and complementary correction and grounding rewards.

5.1 Experiment Setup

  • Experiments use Qwen2.5-VL-3B and 7B backbones, difficult filtered data, and evaluations spanning counting, spatial reasoning, general multimodal tasks, and math reasoning.Training data comes from SAT, TACO, and TallyQA, retaining samples answered incorrectly by Qwen2.5-VL-7B-Instruct; comparisons include proprietary, open-source, tool-augmented, and internal SFT/GRPO baselines.Source: §5.1 · pp. 5–6

5.2 Results

  • ReVISE-3B gains 5.86% on CountBench, 11.28% on CVBench, and 9.94% on BLINK over Qwen2.5-VL-3B-Instruct.The paper attributes these gains over the Qwen2.5-VL-3B-SFT-GRPO baseline to explicit grounding and tool verification benefiting counting, spatial/depth, and relation understanding.Source: Table 1 · p. 6, §5.2 · p. 6
  • ReVISE-7B outperforms ReVPT and CodeDance on CountBench and several CVBench subtasks.ReVISE does not achieve the best results on MMMU or MathVista, which emphasize multimodal knowledge and abstract problem solving over fine-grained localization.Source: §5.2 · p. 6, Table 1 · p. 6

5.3 Further Analysis

  • Both model scales improve on most benchmarks, while RL diversifies tool use and enables recovery from tool-induced failures.The 7B model drops on BLINK (−4.04%), MMVP (−0.09%), and MMSTAR (−0.83%); recovery remains incomplete when failures are ambiguous or irrecoverable, tool coverage is limited, or interaction budgets prevent re-querying.Source: §5.3 · pp. 6–7, Figure 3 · p. 7

5.4 Ablation Study

  • ReVISE SFT reaches 77.89% on CVBench (+3% over REVPT) and 60.48% on MMVP (+2.06%), while MMSTAR drops from 52.61% to 47.53%.The curated cold-start data adds explicit verification and reflection traces absent from the REVPT dataset; the authors attribute the MMSTAR drop to different benchmark reasoning trajectories.Source: §5.4 · p. 7, Table 2 · p. 7
  • Adding self-correction improves BLINK-HARD from 56.80% to 61.20% and MMVP from 62.85% to 67.30%, while grounding improves CountBench from 73.15% to 74.45% and CVBench from 79.24% to 82.10%.Combining both rewards yields 83.83% on CVBench and 63.89% on BLINK-HARD; the training curves show stable reward increases during RL optimization.Source: Table 2 · p. 7, Figure 3 · p. 7, §5.4 · p. 8

5.5 Qualitative Results

  • ReVISE corrects failures caused by unreliable tool outputs by rejecting ambiguous detections, refining zoom regions, and adjusting bounding boxes to capture missing context.Examples include zooming into a sink area to locate a vase and expanding a traffic-scene crop to include all target vehicles.Source: §5.5 · p. 8, Figure 4 · p. 8

6 Conclusion

ReVISE improves the reliability of tool-assisted MLLM reasoning by having models verify and revise predictions using visual evidence. Experiments show greater robustness across perception-intensive benchmarks, while recovery remains limited by visual understanding and tool-output quality.

6 Conclusion

  • ReVISE uses visual evidence to reassess initial predictions, detect inconsistencies, and revise incorrect intermediate conclusions, improving robustness across perception-intensive benchmarks.The framework aims to make tool-assisted reasoning more reliable rather than relying solely on tools to produce answers.Source: §6 · p. 8
  • ReVISE’s revision stage may struggle when the underlying model’s visual understanding is weak or external tools provide inaccurate signals.Ambiguous visual evidence can also hinder recovery from incorrect initial predictions.Source: §6 · p. 8
  • Future work could explore iterative self-correction, uncertainty-guided adaptive revision, stronger perception modules, and applications to video understanding and interactive multimodal agents.Source: §6 · p. 8

Limitations

ReVISE has limitations in its training data and tool scope, performance on abstract tasks, dependence on external vision tools, and evaluation of model scaling.

Limitations

  • Synthetic trajectories and bounding boxes from proprietary models, plus tools designed for visual-centric datasets, may inadequately capture real-world reasoning and limit open-domain robustness.The framework relies on these data and tools for both SFT and RL rewards.Source: Limitations · p. 9
  • Fine-grained spatial-grounding optimization slightly degrades performance on knowledge-intensive tasks such as MMMU and MathVista, which rely less on object localization.The passage cites Yue et al. (2024) for MMMU and Lu et al. (2023) for MathVista.Source: Limitations · p. 9
  • Although ReVISE mitigates blind tool-trust, ambiguous or catastrophic failures in predefined external vision tools can still derail revision.The method remains bounded by the capabilities of those tools.Source: Limitations · p. 9
  • RL evaluations primarily cover 3B and 7B models, leaving how verification behaviors scale to or conflict with larger MLLMs for future study.The limited evaluation scope reflects compute constraints.Source: Limitations · p. 9

Broader Impacts and Ethical Considerations

The work presents foundational methods for visual self-correction and tool-use verification in MLLMs, using existing open-source models and public benchmarks. It introduces no direct societal, ethical, or safety risks, so it omits a dedicated risk discussion.

Broader Impacts and Ethical Considerations

  • The work focuses on foundational visual self-correction and tool-use verification for MLLMs, using existing open-source models and standard public benchmarks.Source: Broader Impacts and Ethical… · p. 9
  • It introduces no sensitive applications, deployable end-user systems, or datasets posing direct societal, ethical, or safety risks, so it omits a dedicated risk discussion.Source: Broader Impacts and Ethical… · p. 9

A Additional implementation details · B The details of applied tools

The appendix specifies the post-training configurations and describes the visual tools used to supply localization, detail, boundary, and depth evidence. It reports compute and training settings alongside each tool’s role and implementation.

A Additional implementation details

B The details of applied tools

  • LLMDet takes an image and text object queries and returns bounding boxes with confidence scores for object localization in grounding and spatial-relation questions.Source: B The details of applied tools · p. 12
  • ZoomIn examines a specified image region using a target box and zoom factor, returning high-resolution crops centered on the box.The crop is clipped to image boundaries, adjusted to keep aspect ratio within 4:1, and upsampled with Lanczos interpolation; zoom is capped at 4.0, with optional scaling for minimum resolution.Source: B The details of applied tools · p. 12
  • Edge detection uses OpenCV’s Scharr operator on grayscale images to produce normalized gradient-magnitude edge maps scaled to [0, 255].The maps provide boundary and shape cues for spatial reasoning.Source: B The details of applied tools · p. 12
  • Depth Anything V2 estimates scene depth to help interpret spatial relationships and occlusions in complex layouts.Source: B The details of applied tools · p. 12

C The details of our curated dataset … E Additional Ablations and Analysis

The supplementary sections document ReVISE’s curated SFT and RL data, prompt and training formulations, and additional analyses of robustness and training design.

C The details of our curated dataset

C.1 Prompt templates

  • SFT trajectories cover Tool-Retry, Negative Tool-Verify, and positive and negative Tool-Free Verify, with prompts guiding verification or revision of answers through visual grounding.Source: §C.1 · p. 12
  • The RL prompt specifies available vision tools and structured reasoning, requires evidence verification and parameter refinement, and requests normalized [0, 1000] bounding boxes for IoU grounding rewards.Gemini-2.5-Pro synthesizes reference boxes, which are de-normalized to image pixel coordinates for reward computation.Source: §C.1 · p. 16

C.2 Dataset statistics

  • The SFT dataset contains 4k samples from correct-answer tool-use trajectories, while RL uses 50k challenging QA pairs with manually verified generated reference boxes.Source: §C.2 · p. 16

D Formulation of standard GRPO

  • GRPO samples candidate completions per image-question pair, scores them with task-specific rewards, and optimizes group-normalized advantages using a clipped objective with KL regularization.The fixed reference policy, KL strength, clipping range, and importance ratio stabilize updates while encouraging reward improvement within each group.Source: D Formulation of standard GRPO · pp. 16, 18

E Additional Ablations and Analysis

  • Additional analyses examine grounding-supervision noise, general-domain data mixing for MMMU and MathVista, and cold-start SFT’s role in structured tool use and self-correction.The supplied passage states the analysis questions but does not report their findings.Source: E Additional Ablations and Analysis · p. 18

E.1 Robustness to Noisy Grounding Supervision · E.2 Effect of Mixed-Domain Training · E.3 Effect of Cold-Start SFT

Across these sections, ReVISE is tested under noisy grounding supervision, mixed-domain training, and removal of cold-start SFT. The results report robustness to moderate box noise, recovery on general reasoning benchmarks with mixed-domain data, and complementary roles for SFT and RL.

E.1 Robustness to Noisy Grounding Supervision

  • Approximately 98% of 1,000 randomly sampled RL-instance bounding boxes are correct, with 2% containing noticeable localization errors.The paper concludes that the grounding reward is generally based on reliable spatial supervision.Source: §E.1 · p. 18
  • GRPO clipping limits excessively large policy updates from high-variance rewards but cannot determine semantic box correctness or eliminate systematic teacher bias.ReVISE also relies on the low observed noise rate and combines grounding with answer-accuracy, format, and self-correction rewards.Source: §E.1 · p. 18

E.2 Effect of Mixed-Domain Training

  • ReVISE’s perception-heavy specialization improves CVBench but lowers MMMU and MathVista performance versus the original Qwen2.5-VL-3B-Instruct model.The authors hypothesize that this degradation stems mainly from training-data distribution rather than an inherent conflict between self-correction and abstract reasoning.Source: §E.2 · p. 18
  • Mixed-domain training improves MMMU by 1.71 points and MathVista by 4.35 points while preserving ReVISE’s strong CVBench performance.MMMU returns to approximately the original instruction-tuned model’s performance, and the CVBench gain remains intact.Source: Table 5 · p. 19, §E.2 · p. 19
  • Mixed-domain training adds math and science reasoning data, with grounding rewards disabled for those samples while answer-accuracy, format, and self-correction rewards remain.SFT adds 1K tool-free verification trajectories (500 MAVIS-Function math and 500 ScienceQA), while RL adds 10K samples with final-answer supervision; perception-heavy samples retain λ_ground = 1.Source: §E.2 · p. 19

E.3 Effect of Cold-Start SFT

  • Cold-start SFT initializes structured interaction behaviors before GRPO, and subsequent RL improves the policy once the model can consistently follow that protocol.The protocol includes reasoning steps, valid tool calls, incorporating tool observations, and self-correction when necessary.Source: §E.3 · p. 19
  • 72.41 to 74.92 on CountBench and 77.89 to 83.83 on CVBench after RL optimization, following SFT initialization.Direct RL without cold-start SFT performs substantially worse than the full pipeline, particularly on CVBench; SFT alone establishes useful structured tool-use and verification behaviors.Source: Table 6 · p. 19, §E.3 · p. 19

F Additional Qualitative Examples

The qualitative examples show that ReVISE can fail when tasks require abstract knowledge or computation beyond its tools, and that verification may not correct those failures. They also show how zoom-based recovery can lose global context and compound spatial errors.

F Additional Qualitative Examples

  • The model labels Neem Oil Extract “Organic” instead of “Chemical” and fails to correct the answer during verification.It reads the image label correctly but lacks the agricultural expertise needed to interpret it.Source: F Additional Qualitative Examples · pp. 19–20
  • The failures reflect a mismatch between spatial-perceptual tools and abstract tasks, alongside knowledge gaps that let verification reinforce mistaken conclusions.The paper proposes expanding the tool library with knowledge retrieval and programmatic execution, such as a Python interpreter.Source: F Additional Qualitative Examples · p. 20
  • The model abandons object_detection and parses a data table, but hallucinates a correlation tool and misses that tr = tb.It lacks the statistical knowledge that the Pearson-correlation test statistic equals the simple-regression slope test statistic.Source: F Additional Qualitative Examples · pp. 19–20, Figure 11 · p. 21
  • In a cluttered counting task, noisy overlapping detections prompt zooming that loses global context, double-counts objects across crop boundaries, and leaves evidence unreconciled.The agent cannot reconcile the localized zoomed evidence with the full scene.Source: F Additional Qualitative Examples · pp. 19–20, Figure 12 · p. 22

Related papers

Loading 2608.29374v1…