Endogenous Chain-of-Thought: Why Diffusion Models Need to Think During Generation

Diffusion models have achieved remarkable fidelity in visual synthesis, yet they remain brittle when confronted with tasks demanding structured logical reasoning. While Multimodal Large Language Models (MLLMs) have been integrated into diffusion pipelines to augment semantic understanding, they typically serve as static encoders, computing text embeddings once at the outset of generation. This architectural choice, as detailed in the recent work EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models by Dai et al., creates a fundamental mismatch between the reasoning requirements of complex tasks and the processing capabilities of current systems. The paper identifies a critical insight: reasoning is not a pre-computation that can be cached in an initial embedding, but rather a dynamic process that must evolve in tandem with the generative trajectory.

The Static Guidance Bottleneck

Current approaches to integrating MLLMs into diffusion frameworks, including Diffusion Transformers (DiTs), treat language models as lookup tables that translate instructions into fixed conditioning vectors. This paradigm suffers from two interconnected failures that the authors term limited single step reasoning and static guidance failure. First, complex spatial and logical constraints, such as those found in Sudoku puzzles or Traveling Salesperson Problem (TSP) route planning, cannot be fully encoded in a single forward pass of an MLLM. The reasoning required to solve a 35x35 Sudoku or navigate a 32x32 maze involves iterative constraint satisfaction and backtracking, processes that exceed the capacity of one shot text encoding.

Second, and more subtly, the guidance provided to the DiT remains invariant throughout the denoising process. In standard diffusion pipelines, the text embedding computed at step zero conditions every subsequent denoising step identically. This static conditioning prevents the model from progressively decomposing complex instructions into actionable sub-tasks. The authors provide empirical evidence that vanilla diffusion models commit to their final solution within the first few denoising steps, spending the remainder of the generation process merely refining visual quality rather than reasoning about logical correctness. This premature commitment explains why existing methods fail catastrophically when generalizing to novel domains, such as maze sizes or font styles not seen during training. Without the ability to revise its reasoning trajectory as the image gradually materializes, the model resorts to fragile pattern matching rather than robust problem solving.

Endogenous Reasoning Through Latent State Refinement

To address these limitations, EndoCoT introduces an endogenous mechanism, meaning the reasoning process emerges from within the model's internal dynamics rather than being imposed from outside. The framework comprises two primary architectural innovations: Iterative Thought Guidance and Terminal Thought Grounding.

The Iterative Thought Guidance module establishes a bidirectional correspondence between the MLLM's latent thought states and the DiT's denoising process. Rather than freezing the text embedding after the initial encoding, the module iteratively refines latent states within the MLLM across denoising timesteps. This creates a genuine Chain-of-Thought (CoT) process where each reasoning step informs a specific phase of image generation. The MLLM effectively maintains a working memory of logical constraints, updating its beliefs about the solution as the visual evidence becomes clearer through denoising.

Complementing this, the Terminal Thought Grounding module ensures that the reasoning trajectory remains anchored to textual supervision. By aligning the MLLM's final reasoning state with ground truth answers during training, the system prevents cumulative drift in the terminal output. This is crucial because iterative refinement without grounding can lead to hallucinated reasoning chains that diverge from valid solutions. Together, these components enable the DiT to execute instructions progressively, treating denoising as a step by step problem solving procedure rather than a single shot synthesis.

Empirical Validation and Generalization

The authors evaluate EndoCoT across four diverse reasoning benchmarks: Maze solving, TSP, Visual Spatial Planning (VSP), and Sudoku. The results demonstrate substantial improvements over state of the art baselines including DiffThinker and Qwen3-VL-8B. Across all tasks, EndoCoT achieves an average accuracy of 92.1%, representing an 8.3 percentage point improvement over the strongest baseline. On complex configurations specifically, the gains are even more pronounced: 90% accuracy on Maze-32 and 95% on Sudoku-35, outperforming prior work by 25% and 40% respectively.

Beyond raw accuracy, the generalization experiments reveal the robustness of endogenous reasoning. When tested on maze sizes or Sudoku fonts not encountered during training, EndoCoT maintains consistent performance while competitors degrade significantly. This suggests that the iterative reasoning process captures underlying structural principles rather than surface level statistical correlations. The paper also presents compelling visualizations of the latent reasoning chain, showing interpretable step by step thought processes that correspond to progressive solution construction in the image space.

Analysis and Broader Implications

The significance of EndoCoT extends beyond the specific benchmarks tested. It represents a fundamental reconceptualization of how reasoning should be structured in non autoregressive generative models. In autoregressive language models, Chain-of-Thought emerges naturally from the sequential token generation process; each token can represent an intermediate reasoning step. Diffusion models, however, operate on the entire signal simultaneously through iterative denoising. The key insight of this work is that CoT must be temporally aligned with this denoising trajectory, not prepended to it as a preamble.

This alignment opens intriguing possibilities for inference time scaling in diffusion models. Just as test time compute scaling has proven effective for large language models, EndoCoT suggests that allowing the MLLM to perform more reasoning steps, or to backtrack and revise earlier thoughts during generation, could yield further accuracy gains. The framework effectively transforms the diffusion process into a search procedure over the space of logical solutions, guided by the semantic understanding of the MLLM.

However, the approach is not without limitations. The iterative refinement of latent states introduces computational overhead compared to single pass encoding. Each denoising step now requires not just the DiT forward pass but also updates to the MLLM's internal representations. For high resolution generation or real time applications, this latency may prove prohibitive without significant optimization. Additionally, the method relies heavily on the base reasoning capabilities of the MLLM itself; if the language model lacks the capacity to solve the underlying logical problem, the iterative guidance cannot compensate for this deficit.

Conclusion and Open Questions

EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models demonstrates that treating MLLMs as static encoders fundamentally underspecifies their potential utility in generative systems. By integrating iterative reasoning directly into the denoising loop, the authors achieve substantial gains in complex visual reasoning tasks while improving generalization to novel domains. The work suggests that the next generation of multimodal systems will require tighter coupling between perception, reasoning, and generation, with each process informing the others in a continuous feedback loop.

Looking forward, several questions remain open. How does this approach scale to video generation, where temporal consistency adds additional reasoning constraints across frames? Can endogenous CoT be extended to 3D generation or other modalities where the denoising trajectory differs significantly from image synthesis? Most importantly, as we build systems that reason during generation, how do we interpret and audit these distributed reasoning chains to ensure alignment with human intent? The path toward truly intelligent generative systems requires not just better pattern matching, but genuine cognitive processes embedded within the generative act itself. EndoCoT provides a compelling template for how such integration might be achieved.