LTA-thinker: Latent Thought-Augmented Training Framework for Large Language Models on Complex Reasoning

2025-09-16

AI Summary

To address the efficiency degradation and performance bottlenecks in large language models (LLMs) caused by “overthinking” during complex reasoning, this paper proposes the Latent Thinking Enhancement (LTE) training framework. LTE introduces a learnable-prior-driven latent thinking generation architecture that explicitly models thought vectors in a continuous latent space. It employs a distribution-guided multi-objective optimization strategy, jointly minimizing: (i) supervised fine-tuning loss, (ii) KL-divergence-based semantic alignment loss, and (iii) contrastive learning–guided reasoning focus loss—thereby enhancing both the diversity of the thought distribution and information utilization efficiency. Evaluated on multiple reasoning benchmarks, LTE achieves state-of-the-art performance, significantly improving both reasoning quality and computational efficiency. Moreover, it demonstrates superior generalization and scalability, effectively mitigating the overthinking problem in LLMs.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Complex Reasoning in Large Language Models can be dynamically optimized using Test-Time Scaling (TTS) to mitigate Overthinking. Methods such as Coconut, SoftCoT and its variant are effective in continuous latent space inference, the core bottleneck still lies in the efficient generation and utilization of high-quality Latent Thought. Drawing from the theory of SoftCoT++ that a larger variance in the generated Latent Thought distribution more closely approximates the golden truth distribution, we propose a Latent Thought-Augmented Training Framework--LTA-Thinker, which improves distributional variance and enhances reasoning performance from two perspectives. First, LTA-Thinker constructs a Latent Thought generation architecture based on a learnable prior. This architecture aims to increase the variance distribution of generated Latent Thought Vectors in order to simplify the overall structure and raise the performance ceiling. Second, LTA-Thinker introduces a distribution-based directional optimization paradigm that jointly constrains both distribution locality and distribution scale. This mechanism improves information efficiency and computational cost through a multi-objective co-training strategy, which combines standard Supervised Fine-Tuning (SFT) loss with two novel losses: Semantic Alignment Loss, which utilizes KL divergence to ensure that the Latent Thought is highly relevant to the semantics of the question; Reasoning Focus Loss, which utilizes a contrastive learning mechanism to guide the model to focus on the most critical reasoning steps. Experiments show that LTA-thinker achieves state-of-the-art (SOTA) performance among various baselines and demonstrates a higher performance ceiling and better scaling effects.

Problem

• Optimizing complex reasoning in LLMs with latent thoughts
• Increasing variance in latent thought distribution for accuracy
• Improving efficiency and performance via multi-objective training

Innovation

• Learnable prior architecture boosts latent thought variance
• Distribution optimization with locality and scale constraints
• Multi-objective co-training with novel alignment and focus losses