Lightweight Latent Reasoning for Narrative Tasks

2025-12-01

AI Summary

To address the excessive reasoning chain length and high computational cost of large language models (LLMs) in narrative tasks, this paper proposes LiteReason: a lightweight, plug-and-able reasoning projection module that maps discrete reasoning steps into continuous latent variables and dynamically skips redundant steps; it further incorporates reinforcement learning to optimize reasoning paths and enables adaptive switching between continuous token generation and standard sampling. Its core innovation lies in the first-ever dynamic coupling of discrete and continuous reasoning with explicit path control. Evaluated on plot-hole detection and chapter generation, LiteReason reduces average reasoning length by 77%–92%, matches the performance of non-implicit reinforcement learning baselines, and significantly outperforms existing implicit reasoning methods—while substantially improving training efficiency and scalability.

Technology Category

• Natural Language Processing: Generation
• Knowledge Representation and Reasoning: Reasoning with Beliefs
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Large language models (LLMs) tackle complex tasks by generating long chains of thought or"reasoning traces"that act as latent variables in the generation of an output given a query. A model's ability to generate such traces can be optimized with reinforcement learning (RL) to improve their utility in predicting an answer. This optimization comes at a high computational cost, especially for narrative-related tasks that involve retrieving and processing many tokens. To this end, we propose LiteReason, a latent reasoning method that can be interleaved with standard token sampling and easily combined with RL techniques. LiteReason employs a lightweight Reasoning Projector module, trained to produce continuous latent tokens that help the model'skip'reasoning steps. During RL, the policy model decides when to activate the projector, switching between latent and discrete reasoning as needed. Experimental results on plot hole detection and book chapter generation show that our method outperforms latent reasoning baselines and comes close to matching non-latent RL training, while reducing final reasoning length by 77-92%. Overall, LiteReason guides RL training to a more efficient part of the performance-computation tradeoff curve.

Problem

• Optimizes latent reasoning for narrative tasks efficiently
• Reduces computational cost of reinforcement learning in LLMs
• Shortens reasoning length while maintaining performance in generation

Innovation

• Lightweight latent reasoning with continuous tokens
• Interleaved latent and discrete reasoning via RL
• Reduces reasoning length by 77-92% efficiently