REASONING GYM: Reasoning Environments for Reinforcement Learning with Verifiable Rewards

2025/05/30 by Zafir Stojanovski, Stojanovski, Zafir, Oliver Stanley +12 · 10 voices · 34 citations
Computer Science · #Reinforcement Learning in Robotics #cs.AI #cs.CL #cs.LG

paper · pdf · doi:10.48550/arxiv.2505.24760

openalex publication_date 2025/05/30 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

We introduce Reasoning Gym (RG), a library of reasoning environments for reinforcement learning with verifiable rewards. It provides over 100 data generators and verifiers spanning multiple domains including algebra, arithmetic, computation, cognition, geometry, graph theory, logic, and various common games. Its key innovation is the ability to generate virtually infinite training data with adjustable complexity, unlike most previous reasoning datasets, which are typically fixed. This procedural generation approach allows for continuous evaluation across varying difficulty levels. Our experimental results demonstrate the efficacy of RG in both evaluating and reinforcement learning of reasoning models.

Citations

• Absolute Zero: Reinforced Self-play Reasoning with Zero Data
• EnigmaEval: A Benchmark of Long Multimodal Reasoning Challenges
• How to Merge Your Multimodal Models Over Time?
• Are We Done with MMLU?
• Learning to summarize from human feedback

Cited by

• Trace: A Taxonomy-Guided Environment for Multidomain Visual Reasoning
• MemoHarness: Agent Harnesses That Learn from Experience
• On-Policy Delta Distillation
• Multi-Task GRPO: Reliable LLM Reasoning Across Tasks
• A Comedy of Estimators: On KL Regularization in RL Training of LLMs
• AutoBaxBuilder: Bootstrapping Code Security Benchmarking
• How Does RL Post-training Induce Skill Composition? A Case Study on Countdown
• SPHINX: A Synthetic Environment for Visual Perception and Reasoning
• Differential Smoothing Mitigates Sharpening and Improves LLM Reasoning
• Scaling Agentic Reinforcement Learning for Tool-Integrated Reasoning in VLMs
• RLVE: Scaling Up Reinforcement Learning for Language Models with Adaptive Verifiable Environments
• Reasoning Curriculum: Bootstrapping Broad LLM Reasoning from Math
• Output Supervision Can Obfuscate the Chain of Thought
• Beyond Pass@k: Breadth-Depth Metrics for Reasoning Boundaries
• Pragyaan: Designing and Curating High-Quality Cultural Post-Training Datasets for Indian Languages
• Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models
• Think Right: Learning to Mitigate Under-Over Thinking via Adaptive, Attentive Compression
• h1: Bootstrapping LLMs to Reason over Longer Horizons via Reinforcement Learning
• BuilderBench -- A benchmark for generalist agents
• What Can You Do When You Have Zero Rewards During RL?
• Studying the Korean Word-Chain Game with RLVR: Mitigating Reward Conflicts via Curriculum Learning
• GEM: A Gym for Agentic LLMs
• BroRL: Scaling Reinforcement Learning via Broadened Exploration
• Beyond Log Likelihood: Probability-Based Objectives for Supervised Fine-Tuning across the Model Capability Continuum
• Recursive Self-Aggregation Unlocks Deep Thinking in Large Language Models
• Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning
• Learning to Reason with Mixture of Tokens
• RL Grokking Recipe: How Does RL Unlock and Transfer New Algorithms in LLMs?
• Reasoning Core: A Scalable RL Environment for LLM Symbolic Reasoning
• Sharing is Caring: Efficient LM Post-Training with Collective RL Experience Sharing
• Loong: Synthesize Long Chain-of-Thoughts at Scale through Verifiers
• OptimalThinkingBench: Evaluating Over and Underthinking in LLMs
• Enhancing Vision-Language Model Training with Reinforcement Learning in Synthetic Worlds for Real-World Success
• Seemingly Simple Planning Problems are Computationally Challenging: The Countdown Game

Discussions

• ReasoningGym: Reasoning Environments for RL with Verifiable Rewards [hn, 105 points, 28 comments]
• ReasoningGym: Reasoning Environments for RL with Verifiable Rewards https://arxiv.org/abs/2505.24760 (https://news.ycombinator.com/item?id=44157077) [bsky, 0 points, 0 comments]
• ReasoningGym: Reasoning Environments for RL with Verifiable Rewards [bsky, 0 points, 0 comments]
• ReasoningGym: Reasoning Environments for RL with Verifiable Rewards #HackerNews https://arxiv.org/abs/2505.24760 [bsky, 0 points, 0 comments]
• ReasoningGym: Reasoning Environments for RL with Verifiable Rewards https://arxiv.org/abs/2505.24760 [bsky, 0 points, 0 comments]
• ReasoningGym: Reasoning Environments for RL with Verifiable Rewards https://arxiv.org/abs/2505.24760 (https://news.ycombinator.com/item?id=44157077) [bsky, 0 points, 0 comments]
• ReasoningGym: Reasoning Environments for RL with Verifiable Rewards https://arxiv.org/abs/2505.24760 [bsky, 0 points, 0 comments]
• ⚡ Hackernews Top story: ReasoningGym: Reasoning Environments for RL with Verifiable Rewards [bsky, 0 points, 0 comments]
• https://arxiv.org/abs/2505.24760 この論文では、検証可能な報酬を用いた強化学習のための推論環境であるREASONING GYMを紹介しています。 REASONING GYMは、エージェントが推論能力を向上させるための多様な環境を提供します。 これにより、強化学習エージェントの性能向上に貢献することが期待されます。 [bsky, 0 points, 0 comments]
• ReasoningGym: Reasoning Environments for RL with Verifiable Rewards View Article | Join the HN Conversation Summary of HN discussion 🧵👇 #hacker-news [bsky, 0 points, 1 comments]

Related

• Reinforcement Learning is all You Need
• Can One Domain Help Others? A Data-Centric Study on Multi-Domain Reasoning via Reinforcement Learning
• Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
• A Reinforcement Learning Environment for Mathematical Reasoning via Program Synthesis
• PUZZLES: A Benchmark for Neural Algorithmic Reasoning