Research map / Reinforcement learning / Visual and representation RL
PaCo: Parameter-Compositional Multi-task Reinforcement Learning
Lingfeng Sun, Haichao Zhang, Wei Xu, Masayoshi Tomizuka
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 |
|---|---|---|---|
| add w-reset on top of loss maskout (ablation) | PaCo · MT10-rand | success rate (%): 76.0 -> 85.4 | Improved |
| add loss maskout on top of vanilla (no stabilization) (ablation) | PaCo · MT10-rand | success rate (%): 71.6 -> 76.0 | Improved |
| add both stabilization schemes to vanilla PaCo (ablation) | PaCo · MT10-rand | success rate (%): 71.6 -> 85.4 | Improved |
| raise parameter set size K from 3 to 5 (ablation) | PaCo · MT10-rand | success rate (%): 74.4 -> 85.4 | Improved |
| raise parameter set size K from 5 to 8 (ablation) | PaCo · MT10-rand | success rate (%): 85.4 -> 81.0 | Hurt |
| set K=10 with one-hot initialization of w (ablation) | PaCo · MT10-rand | success rate (%): 85.4 -> 88.9 | Improved |
| apply compositional structure to actor only instead of output layer only (ablation) | PaCo · MT10-rand | success rate (%): 64.0 -> 73.3 | Improved |
| apply compositional structure to actor and critic instead of actor only (ablation) | PaCo · MT10-rand | success rate (%): 73.3 -> 85.4 | Improved |
| PaCo versus CARE with task metadata | PaCo · MT10-rand | success rate (%): 76.0 -> 85.4 | Improved |
| PaCo versus CARE | PaCo · MT10-fixed | success rate (%): 86.6 -> 93.3 | Improved |
| PaCo with K=20 versus CARE | PaCo · MT50-rand | success rate (%): 50.8 -> 57.3 | Improved |
| PaCo versus Multi-Task SAC | PaCo · MT10-rand | success rate (%): 62.9 -> 85.4 | Improved |
Later papers that compared with this method
6 later accepted papers report PaCo as a baseline in their results tables; 6 of them report a better number on most of the results they share with it.
- MTRL-CG: Multi-Task Reinforcement Learning Method with Spectral Clustering-Based Task Grouping (AAAI 2026 · better on 4 of 4 shared results)
- HyMTRL: A Hybrid Multi-Task Reinforcement Learning Framework via Phased Policy Evolution (ICML 2026 · better on 2 of 2 shared results)
- TOPPO: Rethinking PPO for Multi-Task Reinforcement Learning with Critic Balancing (NeurIPS 2026 · better on 1 of 1 shared result)
- Scalable Multi-Objective and Meta Reinforcement Learning via Gradient Estimation (AAAI 2026 · better on 1 of 1 shared result)
- Hard Tasks First: Multi-Task Reinforcement Learning Through Task Scheduling (ICML 2024 · better on 5 of 5 shared results)
- Multi-Task Reinforcement Learning with Mixture of Orthogonal Experts (ICLR 2024 · better on 2 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- QMP: Q-switch Mixture of Policies for Multi-Task Behavior Sharing (ICLR 2025 · similarity 0.88)
- Multi-Task Reinforcement Learning with Mixture of Orthogonal Experts (ICLR 2024 · similarity 0.87)
- TOPPO: Rethinking PPO for Multi-Task Reinforcement Learning with Critic Balancing (NeurIPS 2026 · similarity 0.87)
- HarmonyDream: Task Harmonization Inside World Models (ICML 2024 · similarity 0.86)
- HarmoDT: Harmony Multi-Task Decision Transformer for Offline Reinforcement Learning (ICML 2024 · similarity 0.86)
- Meta-Reinforcement Learning with Universal Policy Adaptation: Provable Near-Optimality under All-task Optimum Comparator (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.
