Noisy
← this week  ·  Subscribe →

The world is being quietly rearranged by people who write very long documents.

arXiv · April 8, 2026
The title they went with $\pi^2$: Structure-Originated Reasoning Data Improves Long-Context Reasoning Ability of Large Language Models Noisy translates that to

Researchers built a dataset pipeline that lets AI models reason better over long documents

What happened
A team created a method to generate high-quality reasoning questions from structured data (like Wikipedia tables) and use that data to fine-tune large language models. Models trained on this dataset performed 2.7% to 4.3% better on reasoning tasks that require understanding long contexts.
Why it matters
This is a straightforward engineering contribution: better training data produces better model performance on a measurable task. The work shows that reasoning ability in large language models can be systematically improved by using structured data as a foundation for generating questions, then verifying answers through code execution. The open-source release means other teams can apply this approach to their own models and datasets.
The signal
Track whether models fine-tuned on this dataset show improvements on reasoning tasks outside the benchmarks tested here, or whether the gains flatten when applied to genuinely new problem domains.
If you insist
Read the original →