# Bridging Modality Disconnect in Self-Reflection via Closed-Loop Visually Grounded Verification

Source: [Bridging Modality Disconnect in Self-Reflection via Closed-Loop Visually Grounded Verification](<https://www.sota2.com/research/paper/699d30996017c4460ffab4f1>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-06-18T05:00:06.089Z

Authors: Haoyu Zhang, Yuwei Wu, Pengxiang Li, Xintong Zhang, Zhi Gao, Rui Gao, Mingyang Gao, Che Sun, Yunde Jia

Published: 2026-02-21T07:56:59.000Z

## Abstract

In the era of Vision-Language Models (VLMs), enhancing multimodal reasoning capabilities remains a critical challenge, particularly in handling ambiguous or complex visual inputs, where initial inferences often lead to hallucinations or logic errors. Existing VLMs often produce plausible yet ungrounded answers, and even when prompted to "reflect", their corrections may remain detached from the image evidence. To address this, we propose the MIRROR framework for Multimodal Iterative Reasoning via Reflection On visual Regions. By embedding visual reflection as a core mechanism, MIRROR is formulated as a closed-loop process comprising draft, critique, region-based verification, and revision, which are repeated until the output is visually grounded. To facilitate training of this model, we construct \*\*ReflectV\*\*, a visual reflective dataset for multi-turn supervision that explicitly contains reflection triggers, region-based verification actions, and answer revision grounded in visual evidence. Experiments on both general vision-language benchmarks and representative vision-language reasoning benchmarks show that MIRROR improves correctness and reduces visual hallucinations, demonstrating the value of training reflection as an evidence-seeking, region-aware verification process rather than a purely textual revision step.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Object Hallucination Evaluation on POPE](<https://www.sota2.com/research/sota/object-hallucination-evaluation-on-pope.md>): Object Hallucination Evaluation; POPE; Accuracy: 94.42.

- [Hallucination Evaluation on POPE](<https://www.sota2.com/research/sota/hallucination-evaluation-on-pope.md>): Hallucination Evaluation; POPE; Accuracy: 94.42.

- [Hallucination Evaluation on HallusionBench](<https://www.sota2.com/research/sota/hallucination-evaluation-on-hallusionbench.md>): Hallucination Evaluation; HallusionBench; Accuracy: 82.02.

- [Multimodal Understanding on SEEDBench2 Plus](<https://www.sota2.com/research/sota/multimodal-understanding-on-seedbench2-plus.md>): Multimodal Understanding; SEEDBench2 Plus; Accuracy: 76.86.

- [OCR-related Understanding Tasks on TextVQA (val)](<https://www.sota2.com/research/sota/ocr-related-understanding-tasks-on-textvqa-val.md>): OCR-related Understanding Tasks; TextVQA (val); Accuracy: 86.62.

- [OCR VQA on ChartQA (test)](<https://www.sota2.com/research/sota/ocr-vqa-on-chartqa-test.md>): OCR VQA; ChartQA (test); Accuracy: 87.92.

- [Fine-grained Perception on VStarBench](<https://www.sota2.com/research/sota/fine-grained-perception-on-vstarbench.md>): Fine-Grained Perception; VStarBench; Score: 83.77.

- [Fine-grained Perception on MME-RealWorld Lite](<https://www.sota2.com/research/sota/fine-grained-perception-on-mme-realworld-lite.md>): Fine-Grained Perception; MME-RealWorld-Lite; Score: 51.49.

- [Multimodal Reasoning Efficiency on MM-Vet](<https://www.sota2.com/research/sota/multimodal-reasoning-efficiency-on-mm-vet.md>): Multimodal Reasoning Efficiency; MM-Vet; Time (s/sample): 3.73.

- [Fine-grained Visual Reasoning on V\* OOD](<https://www.sota2.com/research/sota/fine-grained-visual-reasoning-on-v-ood.md>): Fine-grained Visual Reasoning; V\* OOD; Overall Score: 83.8.

- [Fine-grained visual understanding on HR-Bench 4K](<https://www.sota2.com/research/sota/fine-grained-visual-understanding-on-hr-bench-4k.md>): Fine-grained visual understanding; HR-Bench-4K; Score: 72.88.

- [Fine-grained Visual Reasoning on MME-RW-Lite ID](<https://www.sota2.com/research/sota/fine-grained-visual-reasoning-on-mme-rw-lite-id.md>): Fine-grained Visual Reasoning; MME-RW-Lite ID; Overall Score: 51.5.

- [Fine-grained Visual Reasoning on HR-Bench-4K OOD](<https://www.sota2.com/research/sota/fine-grained-visual-reasoning-on-hr-bench-4k-ood.md>): Fine-grained Visual Reasoning; HR-Bench-4K OOD; Overall Score: 72.9.

- [OCR & Document Understanding on OCRBench](<https://www.sota2.com/research/sota/ocr-and-document-understanding-on-ocrbench.md>): OCR & Document Understanding; OCRBench; Score: 92.

- [Multimodal Understanding on MM-Vet](<https://www.sota2.com/research/sota/multimodal-understanding-on-mm-vet.md>): Multimodal Understanding; MM-Vet; MM-Vet Score: 66.7.

- [Multimodal Understanding on MMStar](<https://www.sota2.com/research/sota/multimodal-understanding-on-mmstar.md>): Multimodal Understanding; MMStar; Accuracy: 73.33.

- [Multimodal Reasoning on MM-Vet](<https://www.sota2.com/research/sota/multimodal-reasoning-on-mm-vet.md>): Multimodal Reasoning; MM-Vet; MM-Vet Score: 66.7.

- [Mathematical Reasoning on MathVision](<https://www.sota2.com/research/sota/mathematical-reasoning-on-mathvision.md>): Mathematical Reasoning; MathVision; Accuracy: 28.29.

- [Optical Character Recognition on OCRBench](<https://www.sota2.com/research/sota/optical-character-recognition-on-ocrbench.md>): Optical Character Recognition; OCRBench; Score: not recorded.

- [Multi-modal Evaluation on MME-RW](<https://www.sota2.com/research/sota/multi-modal-evaluation-on-mme-rw.md>): Multi-modal Evaluation; MME-RW; Mean Accuracy: not recorded.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2602.18746>)

[PDF](<https://arxiv.org/pdf/2602.18746.pdf>)
