Research map / Reinforcement learning / Visual and representation RL
CompleteP for RL: Maintaining Feature Learning When Scaling Deep Reinforcement Learning
Adam Lee, M Ganesh Kumar, Blake Bordelon, Cengiz Pehlevan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CompleteP instead of NTK parameterization | CompleteP · G1JoystickRoughTerrain (L=2) | compute to reach same reward: -31% | Improved |
| use CompleteP instead of NTK parameterization | CompleteP · G1JoystickRoughTerrain (L=2) | reward at equal compute: +33% | Improved |
| use CompleteP instead of NTK parameterization | CompleteP · AcrobotSwingupSparse (L=1) | reward at equal compute: +701% | Improved |
| use CompleteP instead of NTK parameterization | CompleteP · CartpoleSwingup (L=2) | reward at equal compute: +1% | No clear effect |
| use CompleteP instead of NTK parameterization | CompleteP · humanoid_u_maze (L=4) | reward at equal compute: +27% | Improved |
| scale width and depth with CompleteP (constant learning rate) | CompleteP · HalfCheetah, HumanoidMaze | final reward vs learning rate | Improved |
| scale depth with NTK without lowering learning rate | NTK · HalfCheetah, HumanoidMaze | final reward | Hurt |
| increase width from N=64 to N=2048 under NTK | NTK · reward vs env steps | learning speed | Hurt |
| use CompleteP instead of NTK for width scaling | CompleteP · feature kernel across seeds (CheetahRun) | CKA across seeds | Improved |
| use CompleteP instead of NTK | CompleteP · logit evolution rate vs width | MSE of logits | No clear effect |
| add orthogonal initialization or layer normalization to NTK | NTK · reward at lower compute | maximal reward | Mixed |
| use CompleteP instead of NTK with sparse rewards | CompleteP · sparse-reward tasks | efficiency | Improved |
Closest papers by meaning
- Don't be lazy: CompleteP enables compute-efficient deep transformers (NeurIPS 2025 · similarity 0.87)
- Mixtures of Experts Unlock Parameter Scaling for Deep RL (ICML 2024 · similarity 0.87)
- On the Provable Separation of Scales in Maximal Update Parameterization (ICML 2025 · similarity 0.87)
- Stable Gradients for Stable Learning at Scale in Deep Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Representation Learning Enables Scalable Multitask Deep Reinforcement Learning (NeurIPS 2026 · similarity 0.86)
- The Ladder in Chaos: Improving Policy Learning by Harnessing the Parameter Evolving Path in A Low-dimensional Space (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
