Research map / Reinforcement learning / Visual and representation RL
TD-JEPA: Latent-predictive Representations for Zero-Shot Reinforcement Learning
Marco Bagatella, Matteo Pirotta, Ahmed Touati, Alessandro Lazaric, Andrea Tirinzoni
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 explicit deep state encoder to FB (ablation) | FB · OGBench proprioceptive, avg | success rate: 19.07 -> 39.04 | Improved |
| add explicit deep state encoder to Laplacian (ablation) | Laplacian · OGBench proprioceptive, avg | success rate: 9.07 -> 14.81 | Improved |
| add explicit state encoder to FB (ablation) | FB · DMC proprioceptive, avg | return: 648.9 -> 648.2 (shallow), 632.0 (deep) | No clear effect |
| add explicit state encoder to HILP (ablation) | HILP · DMC proprioceptive, avg | return: 659.4 -> 620.1 (shallow), 603.2 (deep) | Hurt |
| model multi-step policy-conditioned dynamics instead of behavior-policy dynamics (BYOL, BYOL-gamma) (ablation) | TD-JEPA · zero-shot suites, normalized performance | normalized return | Improved |
| use behavior-policy one-step or multi-step prediction on expert-like data (ablation) | BYOL baselines · OGBench | success rate | Mixed |
| use distinct state and task encoders instead of shared symmetric encoder (ablation) | TD-JEPA · zero-shot suites, per-task normalized performance | normalized return | Improved |
| symmetric variant with a single shared encoder-predictor pair (ablation) | TD-JEPA · zero-shot suites | normalized return | No clear effect |
| fine-tune whole pretrained agent versus train from scratch | TD-JEPA pixel-based DMC · DMC pixels, offline and online adaptation | sample efficiency | Improved |
| keep pre-trained state encoder frozen during adaptation instead of fine-tuning (ablation) | TD-JEPA · DMC pixels adaptation | normalized return | No clear effect |
| latent-predictive TD-JEPA versus best baseline on proprioception | TD-JEPA · OGBench proprioceptive, avg | success rate: 39.04 (FB) vs 37.98 | No clear effect |
| latent-predictive TD-JEPA versus FB on pixels | TD-JEPA · DMC pixels (RGB), avg | return: 456.2 -> 628.8 | Improved |
Later papers that compared with this method
1 later accepted paper reports TD-JEPA as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Zero-Shot Off-Policy Learning (ICML 2026 · better on 2 of 3 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
- Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.90)
- Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models (NeurIPS 2025 · similarity 0.88)
- Reinforcement Learning Under Latent Dynamics: Toward Statistical and Algorithmic Modularity (NeurIPS 2024 · similarity 0.88)
- Reasoning with Latent Diffusion in Offline Reinforcement Learning (ICLR 2024 · similarity 0.88)
- Unsupervised Zero-Shot Reinforcement Learning via Functional Reward Encodings (ICML 2024 · similarity 0.88)
- Discerning Temporal Difference Learning (AAAI 2024 · 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.
