Research map / Reinforcement learning / Visual and representation RL
LISA: Learning Interpretable Skill Abstractions from Language
Divyansh Garg, Skanda Vaidyanath, Kuno Kim, Jiaming Song, Stefano Ermon
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 |
|---|---|---|---|
| hierarchical LISA instead of flat Lang DT, 1k demos | Lang DT · BabyAI GoToSeq | success rate (%): 49.3 -> 59.4 | Improved |
| LISA instead of flat Lang DT at 100k demos | Lang DT · BabyAI SynthSeq | success rate (%): 64.2 -> 61.2 | Hurt |
| LISA instead of flat Lang DT at 100k demos | Lang DT · BabyAI BossLevel | success rate (%): 72.0 -> 69.8 | Hurt |
| LISA instead of Lang DT on LOReL with states | Lang DT · LOReL Sawyer states | success rate (%): 33.3 -> 66.7 | Improved |
| LISA instead of Lang DT on LOReL with images | Lang DT · LOReL Sawyer images | success rate (%): 15 -> 40 | Improved |
| LISA instead of flat baseline on unseen compositions | Flat · LOReL Sawyer composition | success rate (%): 13.33 -> 20.89 | Improved |
| LISA instead of LOReL planner on unseen compositions | LOReL Planner · LOReL Sawyer composition | success rate (%): 18.18 -> 20.89 | Improved |
| learned VQ skills instead of k-means skill clusters (ablation) | LISA · BabyAI BossLevel 1k | success rate (%): 20.2 -> 49.1 | Improved |
| skill horizon 1 instead of 5 (ablation) | LISA · BabyAI BossLevel 1k | success rate (%): 32 -> 52 | Improved |
| skill horizon 10 or 50 instead of 5 (ablation) | LISA · BabyAI BossLevel 1k | success rate (%): 52 -> 47 | Hurt |
| increase number of options from 50 to 100 (ablation) | LISA · BabyAI BossLevel 1k | success rate (%): 47 -> 43 | Hurt |
| vary number of options 10 to 50 (ablation) | LISA · BabyAI BossLevel 1k | success rate (%): 47 -> 47 | No clear effect |
Closest papers by meaning
- SkillDiffuser: Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution (CVPR 2024 · similarity 0.89)
- Learning with Language-Guided State Abstractions (ICLR 2024 · similarity 0.89)
- STAR: Learning Diverse Robot Skill Abstractions through Rotation-Augmented Vector Quantization (ICML 2025 · similarity 0.89)
- Goal-Oriented Skill Abstraction for Offline Multi-Task Reinforcement Learning (ICML 2025 · similarity 0.88)
- LLM-based Skill Diffusion for Zero-shot Policy Adaptation (NeurIPS 2024 · similarity 0.87)
- QueST: Self-Supervised Skill Abstractions for Learning Continuous Control (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.
