Research map / Large language models / Reasoning and chain of thought
Latent Chain-of-Thought for Visual Reasoning
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.
Reformulates visual chain-of-thought as amortized posterior inference, training a GFlowNet-style rationale sampler (RGFN) on Qwen2.5-VL and selecting answers via a Bayesian marginal-likelihood scoring (BiN), reporting gains on seven visual reasoning benchmarks.
Problem
Existing training methods (SFT, PPO, GRPO) generalize poorly across unseen visual reasoning tasks and rely on biased reward models, while inference-time methods like Best-of-N are costly and reliant on critics.
Approach
Treat latent CoT as a variable sampled from the posterior P(Z|X,Y) and train an amortized sampler with a sub-trajectory balance (GFlowNet) objective using interpolated token-level sparse rewards and reference-guided filtering; at inference, rank rationale-answer pairs by length-normalized marginal likelihood instead of a verifier.
Key results
- LaCoT-Qwen-7B reaches 68.4 on MathVista vs 63.7 for the Qwen2.5-VL-7B base, and 62.6 for GRPO in the ablation (table reports GRPO 62.6 on MathVista).
- LaCoT-Qwen-3B reaches 40.0 on MathVerse vision-only, versus 26.1 for Qwen2.5-VL-3B.
- BiN outperforms Best-of-N on 3B and 7B (e.g., 7B MMMU 54.9 vs 47.3).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with LaCoT rather than the base model | Qwen2.5-VL-3B · MathVerse vision-only | accuracy (%): 26.1 -> 40.0 | Helped |
| Train with LaCoT rather than the base model | Qwen2.5-VL-7B · MMMU val | accuracy (%): 50.0 -> 54.9 | Helped |
| Use RGFN instead of zero-shot prompting (ablation) | Qwen2.5-VL-7B · MathVista | accuracy (%): 63.7 -> 68.4 | Helped |
| Use RGFN instead of SFT (ablation) | Qwen2.5-VL-7B · MathVerse | accuracy (%): 38.7 -> 43.3 | Helped |
| Use RGFN instead of GRPO (ablation) | Qwen2.5-VL-7B · MMMU | accuracy (%): 47.9 -> 54.9 | Helped |
| Replace BoN with BiN (ablation) | LaCoT-Qwen-3B · MathVerse | accuracy: 21.2 -> 40.0 | Helped |
| Replace BoN with BiN (ablation) | LaCoT-Qwen-7B · MathVista | accuracy: 62.2 -> 68.4 | Helped |
| Add BiN inference-time scaling to SFT (ablation) | Qwen2.5-VL-7B SFT · MMMU | accuracy: 50.6 -> 51.6 | Helped |
| Add BiN inference-time scaling to SFT (ablation) | Qwen2.5-VL-3B SFT · MathVerse | accuracy: 33.3 -> 35.2 | Helped |
| Increase interpolation skip length from 8 to 32 (ablation) | LaCoT-Qwen-7B · MathVision full | accuracy (%): 24.9 -> 23.0 | Hurt |
| Increase interpolation skip length from 8 to 32 (ablation) | LaCoT-Qwen-7B · MathVerse vision-only | accuracy (%): 39.7 -> 42.5 | Helped |
| Increase the number of rationale candidates N (ablation) | LaCoT-Qwen-3B with BiN · MMMU | accuracy | Helped |
Surprising findings
- GRPO fine-tuning underperforms the zero-shot base model on MathVista, MathVerse and MMMU in the ablation.
- SFT on 250k reasoning samples yields only marginal gains over zero-shot on several benchmarks.
Limitations
- Only models up to 7B parameters were tested due to resources.
- Exploration in long latent rationales remains challenging and memory-costly.
- Hallucination from internal knowledge is not addressed.
Experimental setup
Datasets: LLaVA-CoT R1-Onevision MathVista MathVision MathVerse MMMU MMMU-pro MMVet MME
Benchmarks: MathVista MathVision MathVerse MMMU MMMU-pro MMVet MME
Metrics: accuracy
Baselines: SFT GRPO BoN R1-Onevision LLaVA-CoT-11B LLaVA-OV-7B MiniCPM-V2.6 InternVL2-8B InternVL2-4B
Builds on: GFlowNets Sub-Trajectory Balance GRPO Qwen2.5-VL LLaVA-CoT
Closest papers by meaning
- Unveiling Chain of Step Reasoning for Vision-Language Models with Fine-grained Rewards (NeurIPS 2025 · similarity 0.92)
- Rationale-Enhanced Decoding for Multi-modal Chain-of-Thought (CVPR 2026 · similarity 0.92)
- HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning (NeurIPS 2026 · similarity 0.91)
- Unified Multimodal Chain-of-Thought Reward Model through Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.91)
- Let VLMs Grade Their Own Thoughts: A Self-Quantification Approach to Reasoning-Aware Reward Modeling (CVPR 2026 · similarity 0.90)
- Revisiting the Necessity of Lengthy Chain-of-Thought in Vision-centric Reasoning Generalization (CVPR 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.