disco-torch brings DeepMind Disco103 to PyTorch
disco-torch is a new open-source PyTorch port of DeepMind's Disco103 meta-learned reinforcement learning update rule from the 2025 Nature paper, packaged with a pip install, Colab notebook, pretrained weights, and a higher-level DiscoTrainer API. The repo claims numerical parity with the JAX reference and a 99% catch rate on the reference Catch benchmark, making a research-heavy result much easier to experiment with.
The Claude Code angle is the hook, but the durable story is reproducibility: a state-of-the-art RL update rule just got turned into something ordinary PyTorch users can actually run. That matters more than the Reddit post itself because ports like this shorten the gap between reading a paper and testing whether it survives outside the original lab stack.
- –Packaging Disco103 as a pip-installable PyTorch library lowers the barrier for RL researchers who do not want to work inside JAX-first research code
- –The included Colab notebook, pretrained weights, and DiscoTrainer wrapper make this feel closer to a usable research toolkit than a one-off code dump
- –If the validation numbers hold up beyond the Catch benchmark, this could become a convenient baseline for testing learned update rules against hand-designed PPO- and GRPO-style training
- –The repo is still very early, so the real signal will be independent reproduction and whether the port works cleanly in larger custom agent pipelines
DISCOVERED
80d ago
2026-03-09
PUBLISHED
80d ago
2026-03-08
RELEVANCE
AUTHOR
Far-Respect-4827