Reasoning Core: A Scalable RL Environment for LLM Symbolic Reasoning (github.com)

Reasoning Core is a new, open-source text-based reinforcement-learning environment designed to train and evaluate large language models on symbolic reasoning. The toolkit bundles an expressive suite of task families — full first-order logic, formal mathematics via TPTP, novel formal planning domains, arithmetic and syntax challenges — plus verifiers and scoring utilities. It provides a compact Python API (pip install reasoning_core) to generate examples, score answers, run multi-threaded data generation (JSON output consumable by Hugging Face Datasets), and directly evaluate models through an LLM client (examples show running gpt-4.1-mini via OpenRouter). Its significance is practical and methodological: Reasoning Core scales symbolic reasoning benchmarks for RL-style training and supports reproducible dataset creation and automated verification, which are exactly the capabilities needed to advance RL fine-tuning and verifier-guided training of LLMs on rigorous reasoning tasks. Technically, it is interoperable with reasoning_gym (tasks can be registered and combined into weighted composite datasets), supports rollouts and dataset export, and includes ready-to-use task generation and scoring functions (e.g., get_task('arithmetics'), score_answer). By focusing on formal, verifiable tasks and a lean interface, the project aims to make large-scale, RL-driven research on symbolic reasoning more tractable and benchmarkable.