---
title: "Reasoning Core: Designing Broad Procedural Data for Completion-Supervised Reasoning Training"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.05148v1"
permalink: "https://feed.astralor.com/items/itm_dfcceb4e3e4ab0bec8d916a7"
published: "2026-08-05T17:59:44.000Z"
score: 77
channel: "ai"
tags: ["arXiv","Reasoning Core","Completion-Supervised Training","Reasoning Models","Procedural Data Generation"]
---

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

## AI commentary

> Synthetic reasoning data is a major lever for improving model reasoning, yet most teams focus only on generation scale, overlooking how target structure and difficulty calibration shape training outcomes. The paper's true value is not just another dataset, but a reusable audit methodology, reminding us that procedural generation does not equate to correctness. For those working on reasoning data engineering or preference fine-tuning, this open-source toolkit and comparative findings are worth a close read.

## At a glance

This paper introduces Reasoning Core, a collection of 50 procedural generators spanning mathematics, logic, planning, state tracking, formal languages, structured data, games, causality, and code, equipped with semantic scorers, difficulty controls, and task evaluators. Under a matched completion-supervised protocol, Reasoning Core achieves the highest mean scores on DROP, LogiQA, and ARC-Challenge in a primary 3B comparison, outperforming baselines and 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. Audits combining model-assisted review, human adjudication, and regression testing reveal subtle mismatches among generation, rendering, targets, and scoring, reminding that procedural generation alone does not guarantee correctness. The library, generated datasets, and audit material are publicly available.

## Key points

- Reasoning Core comprises 50 procedural generators covering nine reasoning domains, with semantic scorers and difficulty controls.
- At 3B scale, Reasoning Core achieves higher mean scores on DROP, LogiQA, and ARC-Challenge than Procedural Warmup, Reasoning Gym, and SynLogic under completion supervision.
- Semantic validity alone does not guarantee training utility; compact targets and calibrated difficulty are critical design factors.
- Model-assisted audits reveal mismatches among generation, rendering, targets, and scoring, showing procedural generation requires rigorous validation.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_dfcceb4e3e4ab0bec8d916a7