---
title: "推理核心：为完成监督推理训练设计广泛程序化数据"
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","完成监督训练","推理模型","程序化数据生成"]
---

# 推理核心：为完成监督推理训练设计广泛程序化数据

## AI 点评

> 合成推理数据是当前提升模型推理能力的重要杠杆，但绝大多数团队只关注生成规模，忽视了目标结构和难度校准对训练效果的决定性影响。这篇论文最有价值的地方不是又多了一套数据集，而是给出了可复用的审计方法论，提醒我们程序化生成不等于正确。对从事推理数据工程或偏好微调研究的人，这份开源工具和对比结论值得细读。

## 速览

本文提出Reasoning Core，一个包含50个程序化生成器的数据集集合，覆盖数学、逻辑、规划、状态追踪、形式语言、结构化数据、游戏、因果关系和代码等领域，并配备语义评分器、难度控制和任务评估器。在匹配的完成监督协议下，与Procedural Warmup、Reasoning Gym和SynLogic对比，Reasoning Core在3B模型上于DROP、LogiQA和ARC-Challenge取得最高平均分。任务分析表明，语义有效性不足以保证训练效用，紧凑目标和校准难度是关键设计因素。审计揭示了生成、渲染、目标和评分之间的细微不匹配，提示程序化生成并不能保证正确性。库、数据集和审计材料已公开。

## 关键点

- Reasoning Core包含50个程序化数据生成器，覆盖9大推理领域，并支持语义评分与难度控制。
- 在3B模型规模下，完成监督训练中Reasoning Core在DROP、LogiQA和ARC-Challenge上平均分优于Procedural Warmup、Reasoning Gym和SynLogic。
- 语义有效不等于训练有用，紧凑目标与校准难度对训练效果影响显著。
- 模型辅助审计揭示了生成、渲染、目标和评分之间的不一致，说明程序化生成仍需严格质量验证。

## 全文

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.

---

引用前请核对原文;本文由星际流动整理:https://feed.astralor.com/items/itm_dfcceb4e3e4ab0bec8d916a7