Research map / Vision-language and multimodal / Multimodal LLMs

Long Grounded Thoughts: Synthesizing Visual Problems and Reasoning Chains at Scale

ICML 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.

Synthesizes over 1M vision-centric MCQ problems with long reasoning traces, preference pairs, and RL prompts from DOCCI captions and Grounded-SAM boxes, then fine-tunes Qwen2.5-VL-7B via SFT, DPO, and GRPO.

Problem

Open-source multimodal reasoning lags behind text reasoning because there is no systematic way to synthesize large-scale, complex, vision-centric problems with rich long chain-of-thought traces; prior caption-only synthesis saturates quickly.

Approach

Ground MCQ synthesis on object-level bounding boxes and tags (not only dense captions) to avoid diversity collapse, then compose up to five Stage-1 questions per image into harder multi-hop questions, and expand VLM rationales with a stronger reasoning LLM to add cognitive behaviors (subgoal setting, backtracking, verification).

Key results

• Qwen2.5-VL-7B fine-tuned with multistage SFT+DPO reaches 83.25 on V* Bench, exceeding MiMo-VL-7B-RL (81.70) on that benchmark.
• SFT+GRPO reaches 83.80 on CV-Bench and 68.40 on MMStar-V, matching or surpassing MiMo-VL-7B-RL on these benchmarks.
• Composition hardening reduces the perfect-solve rate from 36.7% (LPT) to 3.3% on Stage 2 data, and CoT richness rises from 0.65 to 1.99 behaviors per trace.

Surprising findings

• Short CoT distillation from VLM traces degrades performance below no-CoT direct answers.
• Online GRPO from the base instruct model peaks early and falls below SFT on the same data, so cognitive-behavior SFT is needed before RL.

Limitations

• Relies on human-annotated DOCCI captions; automatically generated captions are deferred to future work.
• Gains on some benchmarks (e.g., MMVP, RealWorldQA) lag MiMo-VL-7B and the base model on several configurations.
• GRPO gains saturate beyond about 70K prompts.

Experimental setup

Datasets: DOCCI MMStar-V

Baselines: LongPerceptualThoughts VLAA-Thinker ReVisual-R1 Virgo MiMo-VL-7B-RL MiMo-VL-7B-SFT GPT-4o

Builds on: LongPerceptualThoughts DPO GRPO Grounded-SAM

Closest papers by meaning

• MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · similarity 0.93)
• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.92)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.91)
• Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning (ICLR 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)
• Latent Implicit Visual Reasoning (CVPR 2026 · similarity 0.91)