Read original
arXivDamien SileoPapers83

Reasoning Core: Designing Broad Procedural Data for Completion-Supervised Reasoning Training

Reasoning Core is a collection of 50 procedural generators spanning mathematics, logic, planning, state tracking, formal languages, structured data, games, causality, and code. It includes semantic scorers, difficulty controls, and task evaluators. Under a matched completion-supervised protocol, the paper reports that its primary 3B setting achieved the best mean performance across DROP, LogiQA, and ARC-Challenge versus no procedural-data baseline, Procedural Warmup, Reasoning Gym, and SynLogic. Audits also uncovered subtle inconsistencies between generation, rendering, targets, and scoring.

Why it's worth reading

Procedural reasoning data is becoming a practical training resource, and this work provides timely comparative evidence that compact targets, calibrated difficulty, and scorer consistency matter as much as task breadth.

Tags

Reasoning Coreprocedural datareasoning trainingcompletion supervisiondata quality3B modelsbenchmarkingopen source