Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Zero-Shot Adaptation of Behavioral Foundation Models to Unseen Dynamics
Maksim Bobrin, Ilya Zisman, Alexander Nikulin, Vladislav Kurenkov, Dmitry V. Dylov
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 transformer belief encoder to FB (Belief-FB) | Vanilla-FB · Randomized-FourRooms (test) | zero-shot return: 0.15 -> 0.4 | Improved |
| add transformer belief encoder to FB (Belief-FB) | Vanilla-FB · Randomized-PointMass (test) | zero-shot return: 0.1 -> 0.45 | Improved |
| add transformer belief encoder to FB (Belief-FB) | Vanilla-FB · Ant-Wind (test) | zero-shot return: 250 -> 550 | Improved |
| add vMF task-vector sampling around context (Rotation-FB) (ablation) | Belief-FB · Randomized-FourRooms (test) | zero-shot return: 0.4 -> 0.61 | Improved |
| add vMF task-vector sampling around context (Rotation-FB) (ablation) | Belief-FB · Randomized-PointMass (test) | zero-shot return: 0.45 -> 0.55 | Improved |
| add vMF task-vector sampling around context (Rotation-FB) (ablation) | Belief-FB · Ant-Wind (test) | zero-shot return: 550 -> 640 | Improved |
| add vMF task-vector sampling around context (Rotation-FB) (ablation) | Belief-FB · OGBench Scene (test) | zero-shot return: 0.45 -> 0.55 | Improved |
| use oracle environment ID instead of inferred belief | FB · Randomized-FourRooms | zero-shot return: 0.90 -> 0.10 (train -> test) | Mixed |
| increase number of training layouts beyond 25 (ablation) | Belief-FB · Randomized-FourRooms | zero-shot return: <3% | No clear effect |
| increase transformer context length beyond a single episode (100 steps) (ablation) | Belief-FB · Randomized-FourRooms and PointMass | zero-shot return | No clear effect |
| use context length shorter than one episode (50 vs 100 steps) (ablation) | Belief-FB · Randomized-FourRooms and PointMass | zero-shot return | Hurt |
| condition backward network B on context (ablation) | Belief-FB · Randomized-FourRooms | zero-shot return | Hurt |
Closest papers by meaning
- Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.88)
- Towards Robust Zero-Shot Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Distributional Successor Features Enable Zero-Shot Policy Optimization (NeurIPS 2024 · similarity 0.87)
- Self-Improving Embodied Foundation Models (NeurIPS 2025 · similarity 0.86)
- Behavioral Foundation Models for Quality Diversity (NeurIPS 2026 · similarity 0.86)
- Behavior-agnostic Task Inference for Robust Offline In-context Reinforcement Learning (ICML 2025 · 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.
