Reasoning Core: A Scalable Procedural Data Generation Suite for Symbolic Pre-training and Post-Training

2026-03-02 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

Reasoning Core is a new, scalable software suite designed to procedurally generate verifiable symbolic reasoning data for language model pre-training and post-training. It covers five formal domains: PDDL planning, first-order logic with equality, context-free grammar parsing/generation, causal reasoning with Bayesian networks, and systems of equations. Each task includes an external solver for rigorous verification and offers continuous difficulty control, facilitating curriculum design. The suite can optionally provide solver-derived reasoning traces for supervised training and offers verifiable reward functions for reinforcement learning. Experiments demonstrate that integrating Reasoning Core data into pre-training enhances downstream reasoning capabilities while maintaining or slightly improving language modeling quality. The code and data are publicly available under the MIT license.

Key takeaway

For research scientists developing or fine-tuning large language models, incorporating Reasoning Core's procedurally generated, verifiable symbolic data into pre-training or post-training can significantly improve reasoning capabilities. You should explore its continuous difficulty control and solver-derived traces to design effective curricula and enhance model performance on complex logical and planning tasks.

Key insights

Reasoning Core generates verifiable symbolic data to enhance language model reasoning and can be used for pre-training and post-training.

Principles

• Verifiable symbolic data improves language model reasoning.
• Continuous difficulty control aids curriculum learning.
• Solver-derived traces enable supervised reasoning training.

Method

Reasoning Core procedurally generates symbolic data across five formal domains, using external solvers for verification and offering difficulty control and reasoning traces for training.

In practice

• Use for pre-training language models on symbolic tasks.
• Integrate into post-training for targeted reasoning improvement.
• Apply for reinforcement learning with verifiable rewards.

Topics

• Symbolic Reasoning
• Procedural Data Generation
• Language Model Pre-training
• Formal Logic
• PDDL Planning

Best for: Research Scientist, AI Researcher, AI Scientist, NLP Engineer

Related on AIssential

• Constructing Evaluation Datasets for Procedural Reasoning: Balancing Naturalness, Grounding, and Multi-Hop Coverage — Explicit representation-aware validation is crucial for ensuring grounding in procedural reasoning evaluation datasets.
• Unlocking the Working Memory of Large Language Models for Latent Reasoning — Large language models can be trained to use working memory for efficient, latent reasoning, decoupling internal computation from external communication.
• Framing, Judging, Steering: An Assessable Competency Model for Teach-ing Students to Reason With Generative AI — Productive generative AI use requires distinct skills: Framing, Judging, and Steering, which can be independently assessed.
• Multilingual Reasoning Gym: Multilingual Scaling of Procedural Reasoning Environments — The Multilingual Reasoning Gym offers procedurally generated, verifiable reasoning problems across 14 languages for AI model training.
• Controllable and Verifiable Process Data Synthesis for Process Reward Models — Synthesized, verifiable process data with controlled errors significantly improves PRM performance in reasoning tasks.
• How Large Languages Models Are Really Made — Language model evolution is driven by increasingly specific training signals, moving from raw statistics to explicit reasoning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.