Base Models Know How to Reason, Thinking Models Learn When
📰 ArXiv cs.AI
Discover how thinking language models learn to reason during training using sparse autoencoders and model diffing, and apply these techniques to improve your own models
Action Steps
- Train a sparse autoencoder on sentence-level activations of reasoning traces to discover a model's reasoning behaviors
- Apply constructive model diffing to reconstruct the base-to-fine-tuned difference and identify key components
- Use the discovered reasoning taxonomies to fine-tune your own language models and improve their reasoning capabilities
- Analyze the differences between base and fine-tuned models to understand what the thinking models learn during training
- Implement the method in your own NLP pipeline to gain insights into your models' reasoning behaviors
Who Needs to Know This
NLP researchers and engineers can benefit from this technique to analyze and improve their language models, while data scientists can apply the method to other domains
Key Insight
💡 Thinking language models learn to reason during training by discovering and refining their reasoning behaviors, which can be analyzed and improved using sparse autoencoders and model diffing
Share This
🤖 New method discovers reasoning behaviors in language models using sparse autoencoders and model diffing! 📊
Key Takeaways
Discover how thinking language models learn to reason during training using sparse autoencoders and model diffing, and apply these techniques to improve your own models
Full Article
Title: Base Models Know How to Reason, Thinking Models Learn When
Abstract:
arXiv:2510.07364v4 Announce Type: replace Abstract: What do thinking language models learn during training that their base models lack? We first present an unsupervised method that discovers a model's reasoning behaviors by training small Sparse Autoencoders on sentence-level activations of reasoning traces, yielding interpretable reasoning taxonomies. Building on this, we introduce constructive model diffing, which aims to reconstruct the base-to-fine-tuned difference from interpretable compone
Abstract:
arXiv:2510.07364v4 Announce Type: replace Abstract: What do thinking language models learn during training that their base models lack? We first present an unsupervised method that discovers a model's reasoning behaviors by training small Sparse Autoencoders on sentence-level activations of reasoning traces, yielding interpretable reasoning taxonomies. Building on this, we introduce constructive model diffing, which aims to reconstruct the base-to-fine-tuned difference from interpretable compone
DeepCamp AI