pi squared: Structure-Originated Reasoning Data Improves Long-Context Reasoning Ability of Large Language Models

Name: pi squared: Structure-Originated Reasoning Data Improves Long-Context Reasoning Ability of Large Language Models
Uploaded: 2026-10-06T09:34:29.872659+00:00
Duration: 272 s
Description: The study created a pipeline called $\pi^2$ to generate high-quality reasoning data from Wikipedia tables to boost large language models' long-context reasoning. The method involves extracting tables, creating complex multi-hop questions validated by code, and generating structured reasoning traces. This curated dataset significantly improves model performance on various long-context benchmarks.

The study created a pipeline called $\pi^2$ to generate high-quality reasoning data from Wikipedia tables to boost large language models' long-context reasoning. The method involves extracting tables, creating complex multi-hop questions validated by code, and generating structured reasoning traces. This curated dataset significantly improves model performance on various long-context benchmarks.

Open the full episode