Curiosity-Driven Exploration Meets JEPA Architecture
Author: @shxf0072 | Date: 2026-05-03
Observation that curiosity-driven exploration in RL papers resembles JEPA-like architectures trained end-to-end with RL. The prediction error that drives curiosity is structurally similar to JEPA’s predictive embedding objective.
Key Claims
- Curiosity-driven RL exploration uses prediction error as intrinsic reward
- This structure resembles JEPA (predicting embeddings rather than raw observations)
- End-to-end RL training of these predictive architectures creates emergent exploration
- Connects two seemingly separate research threads
Takeaways
- JEPA and curiosity-driven RL may be convergent ideas from different communities
- Prediction in latent space (vs pixel space) is key to both approaches
- Unifying these frameworks could improve both representation learning and exploration