AI
New training method helps AI models reason in hidden thought spaces while avoiding feature collapse
Researchers propose Latent Thoughts Tuning, a framework that enables large language models to perform reasoning in continuous hidden states rather than relying on explicit written chains of thought.
Large language models have traditionally improved their reasoning abilities by explicitly writing out step-by-step Chain-of-Thought processes. However, this approach confines the model's internal reasoning to a fixed vocabulary. A new method offers an alternative: enabling models to reason in continuous latent space—essentially hidden thought vectors that live outside the discrete word space.
The challenge with latent reasoning#
Reasoning in latent space has theoretical appeal, but prior attempts have struggled with two major problems. Models experience feature collapse, where the hidden representations lose diversity and become too similar, and they suffer from instability when reusing these states across multiple reasoning steps. Some approaches have tried to mitigate this by relying on assistant models for guidance, but this creates alignment issues and additional complexity.
How Latent Thoughts Tuning works#
Researchers introduce Latent Thoughts Tuning (LT-Tuning), a post-training framework that takes a different approach. Rather than building latent thoughts from raw hidden states alone, the method incorporates a Context-Prediction-Fusion mechanism. This process simultaneously draws on two sources: contextual information from the model's hidden layers, and semantic guidance derived from the vocabulary embedding space. The fusion of these two signals provides richer, more stable representations.
The framework also implements a progressive three-stage curriculum learning pipeline, which gradually teaches the model to switch fluidly between latent reasoning and explicit thought modes depending on what the task requires.
Results and impact#
Experimental results show that LT-Tuning outperforms existing latent reasoning baselines. The method successfully mitigates feature collapse while maintaining robust reasoning accuracy, suggesting a viable path toward more efficient, flexible reasoning in large models without abandoning the structure needed for stability.
