01 — Method
How reasoning steers diffusion
Task-Specific Accuracy — EndoCoT vs. DiffThinker
Maze-8
100%
Maze-32
90% vs 65%
Sudoku-45
100%
Sudoku-35
95% vs 55%
VSP-Super-32
85%
Average Accuracy
Task-specific · all benchmarks
92.1%
+8.3pp over DiffThinker
Generalization to Novel Domains
DiffThinker
Std. Size ✓
Novel Size ✗
Novel Font ✗
EndoCoT
Std. Size ✓
Novel Size ✓
Novel Font ✓
Architecture Overview
Figure 1: EndoCoT architecture. The CoT module generates reasoning tokens that dynamically condition ...
Key Innovations
Genuine Endogenous CoT
The first diffusion framework to enable genuine chain-of-thought reasoning through iterative latent state refinement, bypassing standard single-pass solutions.
Demystifying Diffusion Reasoning
Localizes the source of reasoning via layer-wise sensitivity and attention entropy analysis, identifying the key bottlenecks limiting prior methods.
Empirical Gains on Reasoning and Editing Tasks
Achieves 25-40% improvements on complex visual reasoning benchmarks, enabling controllable inference-time scaling and clearer image editing trajectories.
Training Pipeline
1
Multimodal Encoding
Prompt and image are encoded into latent embeddings
3
Supervised CoT Alignment
MSE loss supervises the correctness of intermediate CoT representations
4
Image Generation Training
Train MMDiT with denoising / reconstruction loss
Inference Pipeline
1
Iterative Reasoning
Run VL reasoning for N steps in the latent space
2
Final Answer / Image Generation
Decode final result from the last latent state
