In-Place and Bidirectional Chain-of-Thought: Beyond Sequential Reasoning

Most CoT work assumes left-to-right, sequential reasoning, but as Jin et al. (2025) show, diffusion LLMs (dLLMs) enable in-place and bidirectional reasoning. This opens the door to “in-place” CoT prompting, where reasoning steps can be flexibly revised, reordered, or early-exited based on confidence estimates. This idea would systematically develop and evaluate such prompting strategies, investigating how they affect reasoning quality, interpretability, and efficiency. The project would also explore hybrid models that combine in-place editing with dynamic step control (see also Xu et al. 2023 on “re-reading”). The goal: create more human-like, editable, and efficient reasoning traces that better match how people solve problems.

References:

1. Thinking Inside the Mask: In-Place Prompting in Diffusion LLMs. Xiangqi Jin, Yuxuan Wang, Yifeng Gao, Zichen Wen, Biqing Qi, Dongrui Liu, Linfeng Zhang (2025). arXiv.org.
2. Re-Reading Improves Reasoning in Large Language Models. Xiaohan Xu, Chongyang Tao, Tao Shen, Can Xu, Hongbo Xu, Guodong Long, Jian-Guang Lou (2023). Conference on Empirical Methods in Natural Language Processing.