Research map / Reinforcement learning / Visual and representation RL
Modular Lifelong Reinforcement Learning via Neural Composition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Reuse and compose shared neural modules across sequential tasks | Comp.+Struct. and Comp.+Search · 64 discrete 2-D tasks | average return during learning | Helped |
| Reuse and compose shared neural modules across sequential tasks | Compositional methods · 48 robot manipulation tasks | average success during learning | Helped |
| Use offline batch RL over replayed task experience to incorporate knowledge into shared modules | Lifelong learners with a multi-stage incorporation phase · 2-D tasks | performance on earlier tasks after later training | Helped |
| Use offline batch RL over replayed task experience to incorporate knowledge into shared modules | Lifelong learners with a multi-stage incorporation phase · robot manipulation tasks | performance on earlier tasks after later training | Mixed |
| Use a monolithic network with a progress phase and a shared-parameter compression phase instead of compositional modules | P&C · 64 discrete 2-D tasks | retention of earlier-task performance | Hurt |
| Use monolithic lifelong baselines rather than compositional modules | Online EWC and CLEAR · 64 discrete 2-D tasks | average return | Hurt |
| Train modules on later tasks and recombine them for unseen task combinations | Modular multi-task model · unseen discrete 2-D task combinations | zero-shot performance | Helped |
| Use a random task sequence without requiring the initial tasks to have disjoint components (ablation) | Comp.+Search-NC · 64 discrete 2-D tasks | forward and backward transfer | No clear effect |
| Modify online PPO training for continuous actions to enable successful task training (ablation) | Compositional methods · robot manipulation tasks | zero-shot transfer | Hurt |
| Use batch RL for offline knowledge incorporation rather than standard offline training (ablation) | P&C · 2-D tasks | performance | Helped |
Closest papers by meaning
- Self-Composing Policies for Scalable Continual Reinforcement Learning (ICML 2024 · similarity 0.90)
- Separating the 'what' and 'how' of compositional computation to enable reuse and continual learning (NeurIPS 2025 · similarity 0.89)
- Compositional Reasoning in Language Models under Reinforcement Learning Post-Training (NeurIPS 2026 · similarity 0.88)
- How does RL Post-training Induce Skill Composition? A Case Study on Countdown (NeurIPS 2026 · similarity 0.88)
- From f(x) and g(x) to f(g(x)): LLMs Learn New Skills in RL by Composing Old Ones (ICLR 2026 · similarity 0.87)
- Autonomous Option Invention for Continual Hierarchical Reinforcement Learning and Planning (AAAI 2025 · similarity 0.87)
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.