Research map / Reinforcement learning / Visual and representation RL
Robust Subtask Learning for Compositional Generalization
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 |
|---|---|---|---|
| Train subtask policies with worst-case adversarial subtask selection | ROSAC · Rooms and F1/10th | success probability and number of subtasks completed | Helped |
| Train subtask policies with worst-case adversarial subtask selection | AROSAC · Rooms and F1/10th | success probability and number of subtasks completed | Helped |
| Use asynchronous parallel training of subtask policies | ROSAC · Rooms and F1/10th | success probability and number of subtasks completed | Hurt |
| Train subtasks independently to complete each subtask, using a manually designed initial-state distribution | Naive · Rooms and F1/10th | long-sequence task performance | Hurt |
| Use dataset aggregation for the initial-state distribution instead of a manually designed distribution | Naive · Rooms and F1/10th | long-sequence task performance | No clear effect |
| Treat the problem as a generic Markov game rather than exploiting its compositional structure | MADDPG · Rooms and F1/10th | robust option performance | Hurt |
| Use a non-compositional PPO-based protagonist and antagonist | Paired · Rooms and F1/10th | robust option performance | Hurt |
| Select training subtasks randomly instead of using epsilon-greedy selection (ablation) | ROSAC · Rooms | sample steps to learn good subtask policies | Hurt |
| Evaluate with a random adversary instead of an MCTS adversary | ROSAC and AROSAC · Rooms and F1/10th | success probability | No clear effect |
Closest papers by meaning
- Hierarchical Reinforcement Learning for Zero-shot Generalization with Subtask Dependencies (NeurIPS 2018 · similarity 0.90)
- How does RL Post-training Induce Skill Composition? A Case Study on Countdown (NeurIPS 2026 · similarity 0.87)
- Compositional Reasoning in Language Models under Reinforcement Learning Post-Training (NeurIPS 2026 · similarity 0.87)
- One Subgoal at a Time: Zero-Shot Generalization to Arbitrary Linear Temporal Logic Requirements in Multi-Task Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Dynamic Optimizations of LLM Ensembles with Two-Stage Reinforcement Learning Agents (ICML 2026 · similarity 0.87)
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · 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.