$Ï^2$: Structure-Originated Reasoning Data Improves Long-Context Reasoning Ability of Large Language Models
This addresses the challenge of enhancing reasoning abilities for AI systems handling complex, multi-hop queries, though it is incremental as it builds on existing fine-tuning methods.
The paper tackles the problem of improving long-context reasoning in large language models by curating high-quality reasoning data from structured sources like Wikipedia tables, resulting in average accuracy gains of up to +4.4% on benchmarks.
We study a pipeline that curates reasoning data from initial structured data for improving long-context reasoning in large language models (LLMs). Our approach, $Ï^2$, constructs high-quality reasoning data through rigorous QA curation: 1) extracting and expanding tables from Wikipedia, 2) from the collected tables and relevant context, generating realistic and multi-hop analytical reasoning questions whose answers are automatically determined and verified through dual-path code execution, and 3) back-translating step-by-step structured reasoning traces as solutions of QA pairs given realistic web-search context. Supervised fine-tuning with \textsc{\small{gpt-oss-20b}} and \textsc{\small{Qwen3-4B-Instruct-2507}} on $Ï^2$ yields consistent improvements across four long-context reasoning benchmarks and our alike $Ï^2$-Bench, with average absolute accuracy gains of +4.3% and +2.7% respectively. Notably, our dataset facilitates self-distillation, where \textsc{\small{gpt-oss-20b}} even improves its average performance by +4.4% with its own reasoning traces, demonstrating $Ï^2$'s usefulness. Our code, data, and models are open-source at https://github.com/vt-pi-squared/pi-squared.