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

Yu Cheng; Arushi Goel; Hakan Bilen

cs.CVAug 29, 2026

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

Authors: Yu Cheng, Arushi Goel, Hakan Bilen

Organizations: University of Edinburgh · NVIDIA Research

Abstract

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.

Figures & tables

Figure 2: Five types of training trajectories to supervise verification and self-correction, illustrating the combinations of negative and positive verification with or without tool use.

Figure 3: Analysis of RL training dynamics and tool policy optimization. (left) Tool invocation frequency on CVBench, illustrating the difference of tool usage between SFT and RL. (center) Error recovery rate, quantifying the proportion of SFT-stage tool-induced failures that are successfully rectified by the RL model via self-correction. (right) Convergence curves of the three designed reward components during the RL stage.

Components	CountBench	CVBench	BLINK (sub)	MMVP	MMSTAR	BLINK-HARD	MMMU	MMBENCH	MathVista
REVPT Zhou et al. (2025) Cold-Start w/o RL*	70.84	74.89	61.17	58.42	52.61	53.78	42.92	78.45	53.09
Ours w/o RL	72.41	77.89	61.57	60.48	47.53	54.67	42.46	78.62	53.21
Ours w/o $r_{\text{correction}}$ , $r_{\text{grounding}}$	73.15	79.24	64.32	62.85	49.60	56.80	44.10	80.50	54.10
Ours w/o $r_{\text{grounding}}$	73.68	81.15	71.85	67.30	54.10	61.20	46.30	83.20	54.80
Ours w/o $r_{\text{correction}}$	74.45	82.10	67.15	64.90	51.50	58.60	45.20	81.90	54.50
Ours	74.92	83.83	74.28	69.41	56.23	63.89	47.31	85.74	55.37

Table 2: Ablation study of curated SFT data and proposed reward components. We evaluate the performance gains brought by curated reflect data, each reward including self-correction (

r_{\text{correction}}

) and the grounding reward (

r_{\text{grounding}}

) over the base RL formulation (

r_{\text{accuracy}}+r_{\text{format}}

). * denotes that the results for the REVPT cold-start baseline are reproduced locally based on settings they report in ReVPT Zhou et al. (2025) , as the official checkpoints are not publicly available. Underline denotes the performance is better than ReVPT Zhou et al. (2025) and Bold represents SoTA performance on reward comparison.

Figure 4: Qualitative comparison of reasoning trajectories. While previous tool-based MLLMs blindly trust flawed tool outputs and wrong answers, ReVISE actively verifies intermediate evidence and dynamically refines tool parameters ( e.g. , adjusting zoom regions) to achieve accurate spatial grounding.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Param Name	SFT	RL
bf16 / tf32	True	True
per_device_batch	8	8
grad_acc_steps	2	–
ppo_mini_batch	–	128
num_generation	–	8
kl_loss_coef	–	1e-3

Appendix

Table 3: Comparison of SFT and RL Parameters

Panel of Figure 5 — Figure 5: The detailed system prompt used for instructing gemini-2.5-pro to synthesize tool-based verification and self-correction trajectories.

Panel of Figure 6 — Figure 6: The detailed system prompt used for instructing gemini-2.5-pro to synthesize negative tool-free verify.

Panel of Figure 7 — Figure 7: The detailed system prompt used for instructing gemini-2.5-pro to synthesize negative tool-free verify.

System Prompt for Ground-Truth Bounding Box Generation (RL Reward)

You are an expert data annotator for Visual Question Answering.

Your goal is to identify and localize ALL objects that are necessary to reason about the user’s Question and arrive at the Correct Answer.

Rules for Object Selection:

1. Identify objects explicitly mentioned in the Question (e.g., ’Where is the [chair] relative to the [table]?’).

2. Identify objects mentioned in the Answer or Options .

3. If the question involves comparison (e.g., ’closer’, ’left of’, ’bigger’), you MUST localize both objects involved in the comparison.

Appendix

Figure 8: Prompt template for generating ground-truth bounding boxes. This prompt is used to query Gemini-2.5-Pro to synthesize the reference object coordinates required to calculate the Grounding Reward (

r_{\text{grounding}}

) during the RL optimization phase.

System Prompt for the ReVISE Agent (RL Training Phase)

You are a specialized multimodal agent. Your purpose is to solve visual question answering tasks by thinking step-by-step and using tools.

# Tools

You are provided with following tools:

1. object_detection : Detects objects in an image. parameters: image_id, objects (list of object names).

2. zoom_in : Zooms in on a specified bounding box in an image. parameters: image_id, bbox (bounding box coordinates), factor (zoom factor).

3. edge_detection : Detects edges in an image. parameters: image_id.

Appendix

Figure 9: System prompt for the ReVISE agent. This instruction template defines the action space (tools) and constraints for the model during the RL training phase, specifically enforcing visual verification and the final output of grounding coordinates.

Figure 10: Failure Case 1: Knowledge deficit and confirmation bias. The model successfully grounds the visual evidence (reading the “Neem Oil Extract” label) but misclassifies it due to a lack of specialized agricultural knowledge. During the verification phase, re-examining the image merely reinforces its flawed heuristic, demonstrating that visual self-correction struggles to overcome fundamental gaps in abstract domain knowledge.

Figure 11: Failure Case 2: Tool hallucination and theoretical reasoning limits. While the model exhibits strong self-reflection by abandoning an ineffective object_detection tool to manually parse the table, it ultimately fails. It hallucinates a non-existent correlation tool and relies on a flawed numerical heuristic, failing to recognize the mathematical equivalence of the test statistics ( $t_{r}=t_{b}$ ).

Figure 12: Failure Case 3: Inappropriate Tool and Cumulative Errors. In a cluttered spatial reasoning scenario, the model attempts to recover from noisy object_detection outputs by invoking the zoom_in tool. While this action provides localized clarity, it inadvertently strips away the global context. The agent consequently double-counts objects across crop boundaries and fails to reconcile the fragmented visual evidence.

Setting	CountBench	CVBench	BLINK-HARD	Avg.
w/o grounding reward	73.68	81.15	61.20	72.01
ReVISE-3B	74.92	83.83	63.89	74.21
5% jitter	74.30 (-0.62)	83.08 (-0.75)	63.37 (-0.52)	73.58 (-0.63)
10% jitter	71.27 (-3.65)	79.76 (-4.07)	60.52 (-3.37)	70.52 (-3.69)

Appendix

Table 4: Sensitivity of ReVISE-3B to noise in the bounding boxes used for grounding supervision during GRPO training. Numbers in parentheses denote the absolute change relative to ReVISE-3B.

Setting	CVBench	MMMU	MathVista
Qwen2.5-VL-3B-Instruct	72.55	49.00	61.20
ReVISE-3B	83.83	47.31	55.37
ReVISE-3B + mixed-domain	83.97 (+0.14)	49.02 (+1.71)	59.72 (+4.35)

Appendix

Table 5: Effect of mixed-domain training on perception-heavy and general reasoning benchmarks. Numbers in parentheses denote improvements over ReVISE-3B.

Setting	CountBench	CVBench
RL-only w/o cold-start SFT	71.98	75.34
ReVISE-3B w/o RL (SFT only)	72.41	77.89
ReVISE-3B (SFT + RL)	74.92	83.83

Appendix

Table 6: Effect of cold-start SFT and RL on ReVISE-3B.

Explore similar work

Reflect to Inform: Boosting Multimodal Reasoning via Information-Gain-Driven Verification
Mar 27, 2026Shuai Lv, Chang Liu, Feng Tang +5Multimodal Large Language Models Multimodal Reasoning
University of Science and Technology of China, Hefei, China · Huawei Foundation Model Department, Shanghai, China · The Chinese University of Hong Kong, Hong Kong, China +1
Are Tools Always Beneficial? Learning to Invoke Tools Adaptively for Dual-Mode Multimodal LLM Reasoning
May 19, 2026Qinghe Ma, Zhen Zhao, Yiming Wu +3Multimodal Large Language Models Multimodal Reasoning
The State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China · Shanghai Artificial Intelligence Laboratory +2
Visual Reasoning through Tool-supervised Reinforcement Learning
Apr 21, 2026Qihua Dong, Gozde Sahin, Pei Wang +4Tool Use in VLMs VLM Reasoning
1Northeastern University · 2Amazon AGI

Mar 27, 2026J/K move · Enter open · S save

Reflect to Inform: Boosting Multimodal Reasoning via Information-Gain-Driven Verification

Shuai Lv, Chang Liu, Feng Tang, Yujie Yuan+4

University of Science and Technology of China, Hefei, China · Huawei Foundation Model Department, Shanghai, China · The Chinese University of Hong Kong, Hong Kong, China+1

Multimodal Large Language Models Multimodal Reasoning VLM Hallucination Mitigation Visual Grounding

Open paper AskPDF arXiv

Multimodal Large Language Models (MLLMs) achieve strong multimodal reasoning performance, yet we identify a recurring failure mode in long-form generation: as outputs grow longer, models progressively drift away from image evidence and fall back on textual priors, resulting in ungrounded reasoning and hallucinations. Interestingly, Based on attention analysis, we find that MLLMs have a latent capability for late-stage visual verification that is present but not consistently activated. Motivated by this observation, we propose Visual Re-Examination (VRE), a self-evolving training framework that enables MLLMs to autonomously perform visual introspection during reasoning without additional visual inputs. Rather than distilling visual capabilities from a stronger teacher, VRE promotes iterative self-improvement by leveraging the model itself to generate reflection traces, making visual information actionable through information gain. Extensive experiments across diverse multimodal benchmarks demonstrate that VRE consistently improves reasoning accuracy and perceptual reliability, while substantially reducing hallucinations, especially in long-chain settings. Code is available at https://github.com/Xiaobu-USTC/VRE.

Abstract

Explore similar work

Reflect to Inform: Boosting Multimodal Reasoning via Information-Gain-Driven Verification

Are Tools Always Beneficial? Learning to Invoke Tools Adaptively for Dual-Mode Multimodal LLM Reasoning

Visual Reasoning through Tool-supervised Reinforcement Learning