Latent Thought Flow: Efficient Latent Reasoning in Large Language Models

2026-06-15

AI Summary

This work addresses the high computational cost of explicit chain-of-thought (CoT) reasoning in large language models and the inability of existing implicit methods to properly model the probability distribution over multiple reasoning trajectories. To this end, the authors propose the Latent Thought Flow framework, which formulates reasoning as variable-length continuous trajectories and trains a sampler via continuous GFlowNets to approximate the posterior distribution defined jointly by answer quality and computational cost. By introducing an entropy-weighted subtrajectory balance objective and a reference prior regularizer, the method achieves, for the first time, a principled probabilistic modeling of multi-path reasoning in implicit inference. Experiments demonstrate that the approach significantly outperforms both explicit CoT and current implicit baselines under both fine-tuning and transfer settings, yielding an average accuracy improvement of 9.5% while reducing reasoning length by 27.2%.

Abstract

Large Language Models (LLMs) increasingly rely on intermediate reasoning, yet explicit Chain-of-Thought (CoT) suffers from a linguistic space bottleneck: each thought must be decoded into tokens, causing high inference overhead. Latent reasoning moves deliberation into continuous space, but existing methods mostly learn deterministic or reward-maximizing paths, lacking a principled way to allocate probability across trajectories with different correctness and costs. We propose Latent Thought Flow (LTF), which models reasoning as variable-length continuous trajectories and trains a sampler to match a reward-induced posterior over answer quality and computation cost. We instantiate this with a continuous GFlowNet using stochastic latent transitions. To handle sparse answer supervision, we introduce an Entropy-Weighted Subtrajectory Balance objective for intermediate rewards and a reference-prior regularizer to anchor exploration. Experiments under finetuning and transfer learning settings show that LTF outperforms explicit CoT and latent reasoning baselines, improving accuracy by 9.5% while reducing reasoning length by 27.2% on average compared with strong latent reasoning baselines.

Problem

• Latent Reasoning
• Chain-of-Thought
• Large Language Models
• Reasoning Trajectories
• Inference Overhead

Innovation

• Latent Thought Flow
• continuous reasoning
• GFlowNet
• subtrajectory balance
• reward-induced posterior