LEPO: Latent Reasoning Policy Optimization for Large Language Models

TL;DR

LEPO introduces a novel RL-based framework for large language models that maintains stochasticity in latent reasoning to improve diversity and performance in reasoning tasks.

Contribution

LEPO is the first to apply reinforcement learning directly to continuous latent representations in large language models, enhancing reasoning diversity.

Findings

LEPO outperforms existing RL methods in reasoning tasks.

Stochastic latent reasoning improves diversity and exploration.

LEPO achieves significant performance gains in experiments.

Abstract

Recently, latent reasoning has been introduced into large language models (LLMs) to leverage rich information within a continuous space. However, without stochastic sampling, these methods inevitably collapse to deterministic inference, failing to discover diverse reasoning paths. To bridge the gap, we inject controllable stochasticity into latent reasoning via Gumbel-Softmax, restoring LLMs' exploratory capacity and enhancing their compatibility with Reinforcement Learning (RL). Building on this, we propose \textbf{\underline{L}}atent R\textbf{\underline{e}}asoning \textbf{\underline{P}}olicy \textbf{\underline{O}}ptimization~(\textbf{LEPO}), a novel framework that applies RL directly to continuous latent representations. Specifically, in rollout stage, LEPO maintains stochasticity to enable diverse trajectory sampling, while in optimization stage, LEPO constructs a unified gradient estimation for both latent representations and discrete tokens. Extensive experiments show that LEPO significantly outperforms existing RL methods for discrete and latent reasoning.