We’re proud to share that the research paper “1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities”, co-authored by our Chief Scientist Tomasz Trzciński, has been awarded Best Paper at NeurIPS 2025 – one of the most prestigious AI and machine learning conferences in the world.
This paper challenges the conventional assumption that the information provided by reinforcement learning (RL) is insufficient to effectively guide the numerous parameters of deep neural networks, hence suggesting that large AI systems be predominantly trained through self-supervision, with RL reserved solely for fine-tuning. Specifically, it:
- Demonstrates that extreme depth (up to 1000+ layers) unlocks new RL capabilities;
- Introduces CRL, learning from full trajectories rather than sparse rewards;
- Achieves up to 2× gains overall and 50× improvements on complex humanoid tasks;
- Depth outperforms width, with deeper, smaller models beating much wider, shallower ones;
- Possible applications: robotics, locomotion, navigation, and high-dimensional control systems.
More is explored in this article.
This recognition highlights the strength of our scientific work and the continued contributions our team makes to the global research community.
This work is the result of collaboration between researchers: Kevin Wang, Ishaan Javali, Michał Bortkiewicz, Tomasz Trzciński, and Benjamin Eysenbach.
Link to the project website: https://wang-kevin3290.github.io/scaling-crl/
Link to the paper in arXiv base: https://arxiv.org/abs/2503.14858
