Dissecting Multilingual Reasoning: Step and Token Level Attribution in CoT
Published:Nov 19, 2025 21:23
•1 min read
•ArXiv
Analysis
This research dives into the critical area of explainability in multilingual Chain-of-Thought (CoT) reasoning, exploring attribution at both step and token levels. Understanding these granular attributions is vital for improving model transparency and debugging complex multilingual models.
Key Takeaways
- •Investigates the attribution of reasoning steps in multilingual CoT.
- •Analyzes the contribution of individual tokens to the reasoning process.
- •Aims to improve explainability and debuggability of multilingual models.
Reference
“The research focuses on step and token level attribution.”