Modeling multiple future thoughts improves continuous reasoning accuracy
Next Thoughts Are Distributions: Generative Autoregressive Reasoning in the Latent Space
Artificial Intelligence
Summary
Solving problems often involves many possible next steps, but computers usually pick only one next thought at a time. The authors present a method called Autoregressive Thought Flow that keeps track of many possible next thoughts as distributions in a compact space. This approach allows the system to explore different ways to solve problems continuously and improves performance on math reasoning tasks. It also benefits from learning methods that reward better thinking and can try multiple reasoning paths at test time for broader solutions.
What this means in practice
- •For mathematics software developers: Improve accuracy and diversity of mathematical reasoning algorithms by modeling multiple possible next reasoning steps in latent space.
- •For ai research engineers: Design AI systems that explore multiple plausible future thoughts to enhance reasoning and decision-making processes continuously.
Authors
Yang Li, Yi Wang, Shiyuan Huang, Yang Liu, Hao Wang, Chengzhi Mao
Abstract
Reasoning problems often admit multiple valid ways to proceed. Continuous reasoning promises to move computation beyond language tokens into a more compact latent space, but representing several plausible ways to think next remains difficult. We introduce Autoregressive Thought Flow (ATF), which models the next continuous thought as a multimodal distribution. A causal autoregressive model performs the reasoning computation, while a lightweight diffusion head generates a plausible next thought from the resulting condition. The sampled thought is fed back into the model, allowing continuous reasoning to unfold for a variable number of steps while preserving the pretrained backbone. Across mathematical reasoning tasks, ATF improves accuracy with compact latent traces and benefits from reinforcement learning and additional test-time thinking. Multi-sample evaluation shows broader solution coverage, indicating that its multimodal predictions capture useful diversity among reasoning paths. Our results suggest that continuous reasoning is more effective when multiple possible next thoughts remain available rather than being collapsed into a single prediction.