Research map / Large language models / Reasoning and chain of thought
Diffusion of Thought: Chain-of-Thought Reasoning in Diffusion Language Models
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.
Proposes Diffusion-of-Thought (DoT), which fine-tunes diffusion language models (Plaid, SEDD) to perform chain-of-thought reasoning by diffusing thought latents over timesteps, with scheduled and coupled sampling for self-correction.
Problem
Whether diffusion language models can leverage chain-of-thought-style reasoning, given that autoregressive CoT is slow and error-prone and diffusion LMs have not been tested for complex reasoning.
Approach
Reasoning steps are represented as latent thoughts that are progressively denoised in parallel over diffusion timesteps, with classifier-free-style conditioning on the problem (partial noising), a multi-pass thought-by-thought variant, training-time scheduled and coupled sampling for self-correction, and a conditional ODE solver to speed up continuous diffusion inference.
Key results
- DoT reaches 100% accuracy on 4x4 and 5x5 multiplication and boolean logic, with up to 27x speed-up over GPT-2 CoT.
- DoT-SEDD-medium outperforms GPT-2-medium with CoT on GSM8K by around 10% (53.5 vs 43.9 accuracy).
- DoT from scratch achieves only about 4.6% accuracy on GSM8K, showing pre-trained language understanding is vital.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DoT instead of GPT2-small CoT | 4×4 Mult. · 4×4 Mult. | throughput: 2.3 -> 62.5 | Helped |
| Use DoT instead of GPT2-small CoT | 5×5 Mult. · 5×5 Mult. | throughput: 1.5 -> 61.8 | Helped |
| Fine-tune a pre-trained diffusion model instead of training DoT from scratch | DoT on GSM8K-Aug · GSM8K-Aug | accuracy: 4.6 -> 53.5 | Helped |
| Use DoT with SEDD-medium instead of GPT2-medium CoT | GSM8K-Aug · GSM8K-Aug | accuracy: 43.9 -> 53.5 | Helped |
| Use multi-pass rather than single-pass DoT | DoT with Plaid · GSM8K-Aug | accuracy: 32.6 -> 37.7 | Helped |
| Use multi-pass rather than single-pass DoT | DoT with SEDD-medium · GSM8K-Aug | accuracy: 53.5 -> 53.3 | No clear effect |
| Add self-consistency decoding | DoT and multi-pass DoT · GSM8K | accuracy | Helped |
| Add self-consistency decoding | DoT with SEDD · GSM8K | accuracy | Helped |
| Use gradient-based conditioning instead of classifier-free conditioning (ablation) | Plaid DoT fine-tuned on GSM8K · GSM8K | accuracy | Hurt |
| Remove scheduled sampling (ablation) | DoT · GSM8K | accuracy | Hurt |
| Remove coupled sampling (ablation) | Multi-pass DoT · GSM8K | accuracy | Hurt |
| Increase DoT generation timesteps | DoT-SEDD-medium · GSM8K | accuracy | Helped |
Surprising findings
- Training DoT from scratch on GSM8K reaches only about 5% accuracy, far below fine-tuned GPT-2.
- Gradient-based classifier guidance fails to recover exact conditioning tokens, motivating partial-noising conditioning instead.
Limitations
- Requires additional training to achieve accurate reasoning rather than working out of the box.
- Pre-trained diffusion LMs are much smaller than modern autoregressive LLMs; DoT lags far behind LLMs hundreds of times larger.
- Evaluation limited to math and logic tasks.
Experimental setup
Datasets: GSM8K-Aug BIG-bench multiplication boolean logic dataset OpenWebText
Benchmarks: GSM8K 4x4 multiplication 5x5 multiplication boolean logic
Metrics: accuracy throughput
Baselines: GPT-2 CoT GPT-2 answer-only Implicit CoT ChatGPT few-shot CoT Plaid DoT SEDD DoT
Builds on: Plaid SEDD DiffuSeq Implicit CoT self-consistency
Closest papers by meaning
- On the Reasoning Abilities of Masked Diffusion Language Models (ICLR 2026 · similarity 0.88)
- Reinforcing the Diffusion Chain of Lateral Thought with Diffusion Language Models (NeurIPS 2025 · similarity 0.88)
- LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning (ICLR 2026 · similarity 0.87)
- Beyond Autoregression: Discrete Diffusion for Complex Reasoning and Planning (ICLR 2025 · similarity 0.87)
- A Formal Comparison Between Chain of Thought and Latent Thought (ICML 2026 · similarity 0.87)
- d2: Improving Reasoning in Diffusion Language Models via Trajectory Likelihood Estimation (ICML 2026 · similarity 0.87)
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.