CoT reasoning generalizes via shared predictors in LLMs
Original title: Understanding the Out-of-Distribution Generalization of Chain-of-Thought Reasoning in LLMs
CoT generalizes OOD by composing reusable ID transitions with risk bounded by O(k) coefficient.
Chain-of-Thought improves LLMs but OOD generalization mechanisms were underexplored. CoT acts as compositional OOD generalization using shared single-step predictors on atomic ID data. Theoretical analysis shows OOD task risk controlled by sum of ID sub-task risks scaled by O(k), with estimation term of order O(sqrt(T/n)). Experiments on Atom-Task and GSM8K support short-to-long and compositional generalization, with explicit state tracking crucial.
- Generalizes by applying shared single-step predictor
- OOD risk controlled by ID risks scaled by O(k)
- Estimation term O(sqrt(T/n))
- Experiments support short-to-long generalization
- Code at https://github.com/maxliang114514/CoT-OOD
From the original article
1 min read at arxiv.orgChain-of-Thought (CoT) reasoning substantially improves Large Language Models (LLMs), yet its out-of-distribution (OOD) generalization mechanism remains theoretically underexplored. We study CoT as compositional OOD generalization: the target task lacks complete reasoning trajectories or direct input-output pairs, while training provides only in-distribution (ID) atomic step data.
By Nuo-Jing Liang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.