Sampling via Decision-Flow: Training-Free Extraction of Improved Latent Reasoning Paths in Large Language Models

2026-09-10

AI Summary

本文提出决策流采样法，无需训练即可提取大语言模型中的潜在推理路径，解决了是否需要强化学习来激活模型潜在能力的问题。

Abstract

A central question in LLM reasoning is whether reinforcement learning (RL) instills genuinely new capabilities or merely reshapes how existing knowledge is expressed during inference. Building on the distribution-sharpening hypothesis, which holds that RL reallocates probability mass toward high-reward trajectories already latent in base models, we ask: can we unlock those latent paths without costly RL fine-tuning? We present Decision-Flow Sampling (DF-Sample), a training-free, data-free inference-time framework that constructs a hierarchical reasoning tree, scores terminal nodes for quality, and back-propagates utilities to inform each intermediate branching decision. Unlike conventional sampling strategies that make purely local step-wise choices, DF-Sample performs explicit global trajectory evaluation before committing to a path, recovering high-quality but low-probability reasoning chains that standard decoding overlooks. On GPQA, DF-Sample achieves 45.6% accuracy, surpassing power sampling (38.9%) and GRPO (39.9%), showing that a training-free method can outperform a trained one. Across three models and four benchmarks, DF-Sample consistently outperforms baselines, indicating substantial latent reasoning potential in pretrained base models.

Problem

• large language models
• reinforcement learning
• latent reasoning paths
• training-free extraction

Innovation

• Decision-Flow Sampling
• hierarchical reasoning tree
• global trajectory evaluation
• training-free