Research map / Vision-language and multimodal / Multimodal LLMs

Visual-Aware CoT: Achieving High-Fidelity Visual Consistency in Unified Models

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VACoT, a visual-aware chain-of-thought for unified generation-understanding models (built on BAGEL) that plans visual consistency checklists, self-corrects iteratively, and is trained with SFT followed by flow-GRPO using visual consistency rewards.

Problem

Existing multimodal CoT methods for unified models check only text alignment with the prompt and ignore visual consistency with reference images in multi-reference generation, causing loss of identity, object attributes and style.

Approach

Generate a structured visual checklist (identity/style/attribute checks between source reference and output), then iteratively self-evaluate and edit the image against that checklist, and reinforce visual consistency via flow-GRPO with a DINO/CSD/GroundingDINO-based visual reward plus CLIP text reward.

Key results

• On OmniContext MULTIPLE, VACoT averages 8.44 vs 6.02 for BAGEL and 7.89 for Uni-CoT (ablation table).
• Overall OmniContext average of 8.26 vs Uni-CoT 7.89 and UiG 6.85; GPT-4o 8.75 remains higher on the SCENE/character setting.
• GenEval overall of 0.84, slightly above Uni-CoT (0.83) and UiG (0.82).

Surprising findings

• Adding PickScore to the reward degrades performance (average 7.49 vs 8.44).
• Additional refinement iterations beyond 2-3 can degrade output quality due to accumulated artifacts and over-correction.

Limitations

• Iterative correction yields diminishing or negative returns beyond 2-3 iterations.
• Reliance on a VLM (Gemini) for planning and evaluation annotations and on GroundingDINO for localization in rewards.

Experimental setup

Datasets: Echo-4o

Benchmarks: OmniContext GenEval

Metrics: OmniContext average score (PF/SC) GenEval overall score

Baselines: BAGEL UiG Uni-CoT GPT-4o Gemini-2.0-flash UNO OmniGen OmniGen2 Echo-4o

Builds on: BAGEL flow-GRPO Uni-CoT UiG

Closest papers by meaning

• Uni-CoT: Towards Unified Chain-of-Thought Reasoning Across Text and Vision (ICLR 2026 · similarity 0.89)
• Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning (ICLR 2026 · similarity 0.88)
• Revisiting the Necessity of Lengthy Chain-of-Thought in Vision-centric Reasoning Generalization (CVPR 2026 · similarity 0.87)
• When Visualizing is the First Step to Reasoning: MIRA, a Benchmark for Visual Chain-of-Thought (CVPR 2026 · similarity 0.87)
• TVI-CoT: Text-Visual Interleaved Chain-of-Thought Reasoning for Multimodal Understanding (ICML 2026 · similarity 0.87)
• Chain-of-Thought Guided Multi-Modal Object Re-Identification (CVPR 2026 · similarity 0.87)