PRISM-MCTS：元认知反思驱动的蒙特卡洛树搜索推理框架

PRISM-MCTS is a reasoning framework addressing traditional MCTS inefficiencies via a process reward model (PRM) and dynamic shared memory. It optimizes reasoning trajectories efficiently, halving required trajectories on the GPQA benchmark, aligning with the shift to test-time computation and meta-cognitive AI.

OpenAI o1's 'slow thinking' marks a shift from pre-training to test-time computation. Traditional MCTS in reasoning treats trajectories as isolated, lacking info sharing—leading to redundant computation and low efficiency.

Inspired by human meta-cognition, PRISM-MCTS includes:

1. Process Reward Model (PRM): Evaluates intermediate steps to identify promising paths early.
2. Dynamic Shared Memory: Stores validated heuristics (effective strategies) and fallacies (error-prone patterns) for cross-trajectory sharing.
3. Branch Pruning & Reinforcement: Uses memory to cut error-prone branches and strengthen successful strategies.

A few-shot training strategy enables PRM to achieve high-fidelity evaluation with minimal labeled data, making it feasible for resource-limited environments.

PRISM-MCTS reduced required trajectories by half on GPQA while outperforming MCTS-RAG and Search-o1, proving 'smart reasoning' is more efficient than exhaustive search.

PRISM-MCTS shows meta-cognitive abilities (learning from experience) are key to advanced AI. It offers a path to high-quality reasoning under limited compute—'fewer but smarter' searches are practical for real-world applications.