3arXiv cs.AI (Artificial Intelligence)·
FootsiesGym: Open-source fighting game benchmark for two-player zero-sum imperfect-information RL
FootsiesGym is a new open-source reinforcement learning environment built on the minimalist 2D fighting game Footsies, designed to isolate cyclic, non-transitive strategic interactions in two-player zero-sum imperfect-information settings. The authors provide a vectorized simulator for high-throughput training on standard hardware and benchmark several RL algorithms. The environment targets a gap in the benchmark landscape for multi-agent game-theoretic RL research.
Related guides (1)
Related events (8)
OpenAI Gym Beta Release
OpenAI released the public beta of OpenAI Gym, a toolkit for developing and comparing reinforcement learning algorithms. The toolkit includes a suite of environments ranging from simulated robots to Atari games, along with a site for comparing and reproducing results. This represented a significant early infrastructure contribution to the RL research community.
Safety Gym: OpenAI Releases RL Safety Constraint Benchmark Suite
OpenAI released Safety Gym, a suite of environments and tools designed to measure progress in training reinforcement learning agents that respect safety constraints during training. The toolkit targets the challenge of constrained RL, where agents must optimize objectives without violating specified safety boundaries. This represents an early formal effort by OpenAI to provide standardized benchmarking infrastructure for safe RL research.
EvoPolicyGym benchmark evaluates autonomous policy evolution by LLM agents in RL environments
A new arXiv preprint introduces EvoPolicyGym, a benchmark for evaluating how LLM-based agents iteratively improve executable policies in compact interactive RL environments under a fixed interaction budget. The benchmark provides trajectory-level diagnostics beyond aggregate scores, distinguishing how agents allocate budget and convert feedback into parametric tuning. GPT-5.5 achieves the strongest aggregate rank score and top-two performance across all 16 environments. The work targets a gap in agent evaluation where iterative policy refinement is conflated with open-ended software engineering progress.
OpenAI Releases Full Version of Gym Retro with 1,000+ Games
OpenAI has released the full version of Gym Retro, a reinforcement learning research platform supporting over 1,000 games across multiple emulators. This expands the previously public release of ~70 Atari and ~30 Sega games. OpenAI is also releasing the tooling used to add new games to the platform, enabling community expansion.
MobileGym: Verifiable Parallel Simulation Platform for Mobile GUI Agent Training
MobileGym is a browser-hosted simulation environment for mobile GUI agent research that enables deterministic outcome verification via structured JSON state and scalable online RL through hundreds of parallel instances (~400 MB/instance, ~3s cold start). The accompanying MobileGym-Bench provides 416 parameterized task templates across 28 apps with deterministic judges. A sim-to-real case study using GRPO on Qwen3-VL-4B-Instruct achieves +12.8 percentage points on the 256-task test set, with real-device execution retaining 95.1% of simulation-side training gains.
OpenAI Releases CoinRun Environment for Measuring RL Generalization
OpenAI released CoinRun, a procedurally generated platformer training environment designed to measure reinforcement learning agents' ability to generalize to novel situations. The environment is positioned as simpler than Sonic the Hedgehog benchmarks but still challenging enough to expose generalization failures in state-of-the-art RL algorithms. It addresses a longstanding puzzle in RL research around overfitting to training environments versus true generalization.
RevengeBench: Benchmark for Reconstructing Agent Decision Programs from Behavioral Observations
RevengeBench is a new benchmark of 75 LLM-generated, Elo-calibrated policies across five game environments that tests whether a learner can reconstruct a hidden agent's decision program as executable code from behavioral traces alone. The benchmark draws from CodeClash tournament trajectories and allows the learner to design controlled behavioral probes (custom opponent policies) to elicit informative behavior before submitting an executable hypothesis. Evaluated across twelve frontier LLMs, recovery quality ranges from 34 to 72% of initial action-distance closed, with reconstructed policies providing measurable competitive advantage especially for weaker models. The work frames policy reconstruction as a tractable inverse problem in code-space, with implications for opponent modeling and policy interpretability.
Putting RL back in RLHF: RLOO Implementation on Hugging Face
Hugging Face published a blog post introducing RLOO (REINFORCE Leave-One-Out), a reinforcement learning algorithm aimed at making the RL component of RLHF more practical and effective. The post discusses implementation details and motivations for revisiting pure RL-based fine-tuning approaches within the TRL library. This represents a technical contribution to the alignment and RLHF tooling ecosystem, offering an alternative to PPO-based RLHF pipelines.