Research map / Reinforcement learning / Visual and representation RL
Semantic Exploration from Language Abstractions and Pretrained Representations
Allison Tam, Neil Charles Rabinowitz, Andrew Kyle Lampinen, Nicholas Andrew Roy, Stephanie C.Y. Chan, DJ Strouse, Jane X Wang, Andrea Banino, Felix Hill
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 |
|---|---|---|---|
| use CLIP/ALM text embeddings of oracle captions as NGU state instead of controllable state | NGU · City coverage | coverage (bins): 83 -> 241 | Improved |
| use pretrained CLIP/ALM image embeddings (LSE-NGU) instead of controllable state | NGU · City coverage | coverage (bins): 83 -> 162 | Improved |
| replace language-shaped image embeddings with ImageNet embeddings (ablation) | LSE-NGU · City coverage | coverage (bins): 162 -> 111 | Hurt |
| use language oracle captions with random features (Lang-RND) instead of visual random features (ablation) | RND · Playroom lift/put | sample efficiency: 33% faster on lift | Improved |
| semantic caption targets (LD) instead of shuffled statistically matched targets (S-LD) (ablation) | language distillation · Playroom lift/put/find | sample efficiency | Improved |
| pretrained language-shaped representations in NGU instead of Vis-NGU | NGU · Playroom lift/put/find | sample efficiency: 50-70% on lift/put, 18-38% on find | Improved |
| ALM network distillation targets instead of random target (ND vs Vis-RND) | RND · Playroom find | sample efficiency: 41% faster | Improved |
| ImageNet embeddings instead of language-shaped embeddings in LSE-NGU (ablation) | LSE-NGU · Playroom find | sample efficiency | Hurt |
| ImageNet embeddings instead of Vis-NGU controllable state (ablation) | NGU · Playroom lift/put | sample efficiency | Improved |
| freeze a pretrained inverse-dynamics model for controllable states in Vis-NGU (ablation) | Vis-NGU · Playroom | learning speed | No clear effect |
Closest papers by meaning
- Improving Intrinsic Exploration with Language Abstractions (NeurIPS 2022 · similarity 0.92)
- Representation-Based Exploration for Language Models: From Test-Time to Post-Training (ICLR 2026 · similarity 0.90)
- SENSEI: Semantic Exploration Guided by Foundation Models to Learn Versatile World Models (ICML 2025 · similarity 0.89)
- Guiding Pretraining in Reinforcement Learning with Large Language Models (ICML 2023 · similarity 0.89)
- Do LLMs Build World Representations? Probing Through the Lens of State Abstraction (NeurIPS 2024 · similarity 0.88)
- Bridging Environments and Language with Rendering Functions and Vision-Language Models (ICML 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.
