English Español Français Deutsch
← Back to the list
Reasoning and decomposition Methodologists Directly upstream of ai100's methodology

Zhiting Hu

Treating a language model as one component in a larger planning system — instead of asking it to do reasoning end-to-end in its own head.

The dominant frame for LLM reasoning is to keep everything inside the language model: prompt it to produce reasoning steps, sample from the resulting distribution, hope the right answer falls out. RAP (Reasoning as Planning, 2023, with Hu as senior author) makes a different bet — use the LLM as a world model that predicts what happens after a candidate action, and combine it with classical Monte Carlo Tree Search to choose which action to actually take. The architectural consequence is that reasoning becomes auditable in a way pure prompting isn't: you can inspect the search tree, see which states the model evaluated as promising, and trace why the final answer was chosen.

Worth following when
you want to know how LLM reasoning behavior changes when the model is constrained to operate inside an explicit search structure rather than freelancing through prompted text.
Topics
language models as world models for planning; integration of classical search (MCTS) with LLM components; auditable reasoning traces in agentic systems.
Key works
RAP — Reasoning with Language Model is Planning (2023, senior author); broader publications on combining symbolic and neural methods for structured generation.