Reasoning Core: Designing Broad Procedural Data for Completion-Supervised Reasoning Training
2026-08-05 • Computation and Language
Computation and Language
AI summaryⓘ
The authors created Reasoning Core, a set of 50 computer programs that generate different types of reasoning problems in areas like math, logic, and planning. They compared Reasoning Core with other similar collections to see which helps train AI models better for reasoning tasks. Their results show that Reasoning Core generally leads to better performance, but simply having valid problems isn't enough; the difficulty and clarity of the tasks matter too. They also emphasize the importance of careful checking to catch errors in problem creation. All their tools and data are openly shared for others to use.
procedural generationfine-tuningreasoning problemssemantic scoringdifficulty controltask evaluationmachine learning datasetsmodel auditingcompletion supervisionAI benchmarking
Authors
Damien Sileo, Valentin Lacombe, Dimitri Kachler
Abstract
Procedural generators produce useful verifiable reasoning problems at scale, but have received less attention as data for completion-supervised fine-tuning. We introduce Reasoning Core, a collection of 50 generators spanning mathematics, logic, planning, state tracking, formal languages, structured data, games, causality, and code, with semantic scorers, difficulty controls, and task evaluators. Under a matched completion-supervised protocol, we compare Reasoning Core with Procedural Warmup, Reasoning Gym, and SynLogic across four base-model settings and multiple training durations. In the primary 3B comparison, Reasoning Core achieves the highest mean scores on DROP, LogiQA, and ARC-Challenge, exceeding both the baseline without procedural data and all three alternative procedural collections. Task-level analyses show that semantic validity alone does not ensure training utility, highlighting compact targets and calibrated difficulty as important design factors. We ran audits combining model-assisted review, human adjudication, and regression testing. Applied throughout Reasoning Core development and to the other collections, they reveal subtle mismatches among generation, rendering, targets, and scoring, a reminder that procedural generation alone does not guarantee correctness. The library, generated datasets, and audit material are publicly available.