New method uses geometric probes to improve LLM math reasoning by 62%

Researchers have introduced Spectral Orthogonal Exploration (SOE), a novel framework designed to combat 'Reasoning Collapse' in large language models during complex mathematical tasks. SOE operates under a 'Student Guides Teacher' paradigm, where a weaker model doesn't imitate but instead probes the teacher model orthogonally to its dominant reasoning subspace. This intervention encourages more diverse reasoning trajectories, leading to significant improvements in accuracy and sampling efficiency on mathematical benchmarks. Preliminary results also indicate SOE's effectiveness in logic and code generation tasks. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Mitigates reasoning collapse in LLMs, potentially improving performance on complex tasks like math and code generation.

RANK_REASON Academic paper detailing a new method for improving LLM reasoning.

Read on arXiv cs.AI →

• Spectral Orthogonal Exploration
• SOE
• Large Language Models
• Dayu Wang
• arXiv

• paper
• safety