Research map / Vision-language and multimodal / Multimodal LLMs

GThinker: Towards General Multimodal Reasoning via Cue-Guided Rethinking

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 GThinker, a 7B multimodal reasoning MLLM trained with a Cue-Rethinking pattern via pattern-guided SFT cold start and DAPO-based RL, plus a 11K-sample GThinker-11K dataset, targeting general-scenario multimodal reasoning.

Problem

Multimodal LLMs underperform on vision-centric reasoning in general scenarios, where initial visual perceptions can be misleading and models rarely revise them; existing slow-thinking and outcome-reward RL methods are tuned for math and science.

Approach

A free-form reasoning pattern that tags visual cues, triggers a rethinking prompt after the initial reasoning, and revises flawed cues before answering; trained via pattern-guided selective SFT cold start and then DAPO with a hybrid verifiable reward supporting multiple-choice, math, and short open-ended answers.

Key results

• GThinker-7B reaches 81.5% overall on M3CoT, on par with O4-mini (80.9%) and above most open-source reasoning models.
• GThinker-7B scores 66.4% on MMStar, 70.1% on RealWorldQA, 40.7% on MMMU-Pro, and 72.7% on MathVista.
• Ablation: removing incentive RL drops M3CoT overall from 81.5 to 73.6; removing pattern-guided cold start drops it to 61.5.

Surprising findings

• Pattern-guided selective formatting improved science and math but reduced commonsense accuracy by 2.7 points.
• DAPO alone offers limited gains on general scenarios, though it helps math and science.

Limitations

• Relies on scarce open-source multimodal reasoning data; the 11K dataset is small.
• Authors state code, model, and data will be released 'soon', not yet available at submission.
• Weaker on knowledge/perception-heavy benchmarks such as RealWorldQA, MMMU-Pro and MathVision compared with GPT-4o/Gemini.

Experimental setup

Datasets: M3CoT KiloGram ScienceQA TableWMP Math Sherlock

Benchmarks: M3CoT MMStar RealWorldQA MMMU-Pro MathVista MathVision

Metrics: accuracy

Baselines: O4-mini O3 Gemini-2.5 Pro GPT-4o Qwen2.5-VL-7B InternVL2.5-MPO-8B Kimi-VL-A3B-Thinking MM-Eureka-7B R1-OneVision-7B VLAA-Thinker-7B LLaVA-CoT-11B Ovis2-8B

Builds on: DAPO GRPO DeepSeek-R1 Qwen2.5-VL M3CoT

Closest papers by meaning

• Look-Back: Implicit Visual Re-focusing in MLLM Reasoning (AAAI 2026 · similarity 0.91)
• Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning (NeurIPS 2025 · similarity 0.91)
• DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (ICML 2026 · similarity 0.91)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.91)
• VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning (ICLR 2026 · similarity 0.91)
• LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.91)