Reproduce “CoT Vectors: Transferring and Probing the Reasoning Mechanisms of LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Li Li, Ziyi Wang, Yongliang Wu, Jianfei Cai, Xu Yang
Researchers from Southeast University and Monash University introduced "CoT Vectors," a method to efficiently transfer and probe multi-step reasoning capabilities in Large Language Models by encoding reasoning knowledge as compact additive vectors. The approach achieved superior performance over LoRA fine-tuning with significantly fewer parameters and revealed a three-stage functional organization for reasoning within LLMs.
Key insights
- Large Language Models exhibit a consistent "three-stage" functional organization for multi-step reasoning, characterized by distinct layer-wise information densities and representational complexities.
- The instability of passively "extracted" CoT Vectors in middle layers stems from the highly variable, sample-specific nature of representations during core reasoning, making them less generalizable.
- Parametrically "learnable" CoT Vectors successfully distill a robust and generalizable reasoning signal, enabling more effective and stable performance by overcoming individual layer limitations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.