Research map / Reinforcement learning / Visual and representation RL
ELLA: Exploration through Learned Language Abstraction
Suvir Mirchandani, Siddharth Karamcheti, Dorsa Sadigh
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 relevant low-level GoTo rewards | PPO on PutNext-Room · PutNext-Room | sample efficiency | Improved |
| add relevant low-level GoTo rewards | PPO on PutNext-Maze · PutNext-Maze | sample efficiency | Improved |
| add low-level Pick rewards for collecting relevant keys | PPO on Unlock-Maze · Unlock-Maze | sample efficiency | Improved |
| add relevant low-level rewards | PPO on Combo-Maze · Combo-Maze | sample efficiency | No clear effect |
| add low-level Open and GoTo rewards | PPO on Open&Pick-Maze · Open&Pick-Maze | sample efficiency | Improved |
| add relevant low-level rewards | PPO on Sequence-Maze · Sequence-Maze | sample efficiency | Improved |
| reduce termination-classifier training data from 15K to 2K (ablation) | ELLA on Open&Pick-Maze · Open&Pick-Maze | average reward: 15K; 2K | No clear effect |
| reduce termination-classifier training data to 500 (ablation) | ELLA on Open&Pick-Maze · Open&Pick-Maze | average reward: 500 | Hurt |
| reduce termination-classifier training data to 100 (ablation) | ELLA on Open&Pick-Maze · Open&Pick-Maze | termination-classifier accuracy: 66% | Hurt |
| reduce termination-classifier training data to 100 (ablation) | ELLA on Open&Pick-Maze · Open&Pick-Maze | learning-curve quality: 100 | Hurt |
| replace online relevance classifier with an oracle after initially rewarding all low-level tasks (ablation) | ELLA · Four evaluated tasks | average reward: 2.5 million frames | Improved |
| replace online relevance classifier with an oracle after initially rewarding all low-level tasks (ablation) | ELLA · Sequence-Maze | average reward: 2.5 million frames | No clear effect |
Closest papers by meaning
- Subwords as Skills: Tokenization for Sparse-Reward Reinforcement Learning (NeurIPS 2024 · similarity 0.88)
- Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data (NeurIPS 2025 · similarity 0.88)
- ARIA: Training Language Agents with Intention-driven Reward Aggregation (NeurIPS 2025 · similarity 0.88)
- Dense Reward for Free in Reinforcement Learning from Human Feedback (ICML 2024 · similarity 0.88)
- Motif: Intrinsic Motivation from Artificial Intelligence Feedback (ICLR 2024 · similarity 0.88)
- Generalizing Verifiable Instruction Following (NeurIPS 2025 · similarity 0.88)
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.
