Research map / Vision-language and multimodal / Multimodal LLMs
EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models
Xuanlang Dai, Yujie Zhou, Long Xing, Jiazi Bu, Xilin Wei, Yuhong Liu, Beichen Zhang, Kai Chen, Yuhang Zang
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 |
|---|---|---|---|
| add auxiliary semantic loss (ablation) | EndoCoT · Maze-32 | accuracy: 14 -> 90 | Improved |
| add auxiliary semantic loss (ablation) | EndoCoT · Maze-32 | path repetition: 67.24 -> 98.13 | Improved |
| use implicit latent tokens instead of explicit text tokens (ablation) | EndoCoT · Maze-32 | accuracy: 0 -> 90 | Improved |
| use implicit latent tokens instead of explicit text tokens (ablation) | EndoCoT · Maze-8 | accuracy: 34 -> 100 | Improved |
| jointly train MLLM and DiT instead of DiT only (ablation) | EndoCoT · Maze-32 | accuracy: 18 -> 90 | Improved |
| jointly train MLLM and DiT instead of MLLM only (ablation) | EndoCoT · Maze-32 | accuracy: 0 -> 90 | Improved |
| add second-stage terminal consolidation to single-stage training (ablation) | EndoCoT · Maze-32 | accuracy: 14 -> 90 | Improved |
| increase reasoning steps tau from 2 to 50 (ablation) | EndoCoT · Maze-32 | accuracy: 11 -> 90 | Improved |
| increase reasoning steps tau from 2 to 20 (ablation) | EndoCoT · Maze-16 | accuracy: 72 -> 100 | Improved |
| use EndoCoT instead of DiffThinker (task-specific) | Qwen-Image-Edit-2511 · Sudoku-35 | accuracy: 55 -> 95 | Improved |
| use EndoCoT instead of DiffThinker (unified training) | Qwen-Image-Edit-2511 · Visual reasoning average | accuracy: 77.1 -> 84.2 | Improved |
| use EndoCoT instead of DiffThinker (unified training) | Qwen-Image-Edit-2511 · Maze-32 unified | accuracy: 66 -> 52 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiffThinker — DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (ICML 2026) · better on 7 of 9 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Thinking Diffusion: Penalize and Guide Visual-Grounded Reasoning in Diffusion Multimodal Language Models (CVPR 2026 · similarity 0.91)
- Visual Thoughts: A Unified Perspective of Understanding Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.90)
- SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.90)
- DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (ICML 2026 · similarity 0.89)
- AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.89)
- ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning (ICLR 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.
