Spectral Orthogonal Exploration Improves LLM Reasoning
A new study on arXiv has introduced a framework called Spectral Orthogonal Exploration (SOE) that seeks to tackle an issue known as 'Reasoning Collapse' in Large Language Models (LLMs) when they deal with complicated math problems. Researchers found that ineffective reasoning often gets stuck in a low-rank bias area within the model's hidden-state structure, limiting the ability to find correct answers. SOE utilizes a 'Student Guides Teacher' approach, where a weaker auxiliary agent acts as an orthogonal probe, injecting diverse reasoning signals into the main teacher's subspace. This strategy pushes the teacher to explore a wider array of reasoning options, improving exploration compared to traditional random sampling. Tests on math datasets confirm its success.
Key facts
- Paper titled 'Student Guides Teacher: Weak-to-Strong Inference via Spectral Orthogonal Exploration'
- arXiv identifier: 2601.06160v2
- Addresses 'Reasoning Collapse' in LLMs on mathematical reasoning tasks
- Failed reasoning traces associated with low-rank bias manifold in hidden-state geometry
- Proposes Spectral Orthogonal Exploration (SOE) framework
- Uses 'Student Guides Teacher' paradigm with weak auxiliary agent as orthogonal probe
- Injects heterogeneous reasoning signals into teacher's orthogonal complement of dominant subspace
- Experiments show improved exploration over standard sampling
Entities
Institutions
- arXiv