# open-thought/reasoning-gym

[NeurIPS 2025 Spotlight] Reasoning Environments for Reinforcement Learning with Verifiable Rewards

Repository: https://github.com/open-thought/reasoning-gym
Canonical: https://ross.abutalabs.com/products/reasoning-gym
Language: Python
License: Apache-2.0
License Family: permissive
Topics: gym, reinforcement-learning, large-language-models
Last push: 2026-04-17T19:39:15+00:00

## Health v2 (maintenance only)
Score: 66/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 77, release rhythm 65, longevity 41
- inputs: {"age_days": 587, "days_push": 138, "days_rel": 158, "gap_med": 56.5, "n_releases_24m": 5}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1494, forks 128 (observed 2026-08-28T04:04:52.816936+00:00)

## What it is
Reasoning Gym is a Python library of procedural dataset generators and algorithmically verifiable reasoning environments for training LLMs with reinforcement learning and verifiable rewards. It offers 100+ tasks across domains like algebra, logic, graph theory, and games, with adjustable complexity and a standard score_answer verification interface.

## Use cases
- generate infinite training data for RL with verifiable rewards
- train reasoning models on math and logic tasks
- evaluate LLM reasoning with algorithmic verification
- create procedurally generated puzzle environments for RLHF
- benchmark language models on reasoning tasks

## When to choose
- you need scalable, procedurally generated reasoning tasks with automatic answer verification
- you are training or evaluating LLMs with RLVR and want diverse domains
- you want adjustable task difficulty for curriculum-style RL training

## When to avoid
- you need static benchmark datasets with fixed test sets
- you are doing general-purpose supervised fine-tuning without verifiable answers
- you need non-Python environments or GPU-accelerated simulation

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, reinforcement-learning, llm-training, data-generation, benchmarking
- domain: reinforcement-learning, large-language-models, machine-learning
- platform: python, cli
- tags: gym-environments, verifiable-rewards, procedural-datasets, reasoning-tasks, rl-training-data, algorithms

## Member repositories
- open-thought/reasoning-gym (main) score 66

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:52.816936+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:33:23.697822+00:00, confidence not recorded.
  - readme: https://github.com/open-thought/reasoning-gym (fetched 2026-08-28T04:04:52.816936+00:00, sha 28e8b1912742)
  - registry_pypi: https://pypi.org/pypi/reasoning-gym/json (fetched 2026-08-29T11:38:51.380691+00:00, sha 213a5ed7f63f)
- Data as of 2026-08-30T08:39:29.467469+00:00.