Research map / Vision-language and multimodal / Multimodal LLMs
AtomThink: Multimodal Slow Thinking With Atomic Step Reasoning
Kun Xiang, Zhili Liu, Terry Jingchen Zhang, Yinya Huang, Yunshuang Nie, Kaixin Cai, Yiyang Yin, Runhui Huang, Hanhui Li, Yihan Zeng, Yu-Jie Yuan, Jianhua Han and 3 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune on AMATH-SFT and use SCoT inference (ablation) | Llama3.2-Vision-11B · MathVista | accuracy (%): 44.3 -> 57.1 | Improved |
| use Self-structured CoT inference instead of standard CoT | AtomThink-LlamaV · MathVista | accuracy (%): 50.4 -> 57.1 | Improved |
| add PRM-guided step-wise beam search to SCoT | AtomThink-LlamaV · MathVista | accuracy (%): 57.1 -> 58.4 | Improved |
| add PRM-guided search to SCoT | AtomThink-LLaVA · WeMath | accuracy (%): 34.5 -> 44.2 | Improved |
| add PRM-guided search to SCoT | AtomThink-LlamaV · ScienceQA | accuracy (%): 85.9 -> 85.8 | Hurt |
| use beam search instead of greedy step-wise search (ablation) | AtomThink-EMOVA · MathVerse (300-sample subset) | accuracy (%): 38.3 -> 45.3 | Improved |
| use MCTS instead of beam search (ablation) | AtomThink-EMOVA · MathVerse (300-sample subset) | accuracy (%): 45.3 -> 47.6 | Improved |
| use Best-of-N Avg aggregation instead of SCoT alone (ablation) | AtomThink-EMOVA · MathVista-M vs MathVista-G | accuracy (%): 54.2 -> 58.7 on MathVista-M; 46.7 -> 40.5 on MathVista-G | Mixed |
| increase beam-search candidates from 0 to 3 (ablation) | AtomThink-LlamaV · MathVision-mini | accuracy (%): 13.9 -> 23.3 | Improved |
| scale AMATH-SFT data from 0k to 124k (ablation) | AtomThink-LLaVA · MathVision-mini | accuracy (%): 9.28 -> 12.45 | Improved |
| add 100k LLaVA-Instruct data instead of none (ablation) | AtomThink-LLaVA · MathVision-mini | accuracy (%): 8.81 -> 12.45 | Improved |
| scale LLaVA-Instruct from 100k to 400k (ablation) | AtomThink-LLaVA · MathVision-mini | accuracy (%): 12.45 -> 12.50 | No clear effect |
Closest papers by meaning
- AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.92)
- ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.91)
- DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (ICML 2026 · similarity 0.90)
- CMMCoT: Enhancing Complex Multi-Image Comprehension via Multi-Modal Chain-of-Thought and Memory Augmentation (AAAI 2026 · similarity 0.90)
- Thinking Diffusion: Penalize and Guide Visual-Grounded Reasoning in Diffusion Multimodal Language Models (CVPR 2026 · similarity 0.89)
- SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.89)
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.
