KAGE-Bench preprint published

KAGE-Bench preprint is now available on arXiv! We propose an environment (and a benchmark built on it) for studying visual generalization in RL (i.e., training on one image distribution and validating on another). We wrote the benchmark in JAX because alternatives are too slow for rapid hypothesis testing. We also added the ability to customize the environment along 93 easily configurable parameters, as other benchmarks often confound multiple shifts in observation distributions (for example, changing not only illumination but also textures, etc.).

Link to visualizations (there are many more in the paper): https://avanturist322.github.io/KAGEBench/ Paper: https://arxiv.org/abs/2601.14232 Code: https://github.com/CognitiveAISystems/kage-bench Hugging Face: https://huggingface.co/papers/2601.14232