world2vec paper: mechanistic world models with 10x fewer params, beating DreamerV3
burny_tech · x · 2026-11-01
An arXiv paper introduces world2vec, which learns mechanistic world models directly from pixels and actions without rewards or reconstruction. Key points:
- Latent dynamics faithfully reproduce true systems over long prediction horizons
- At least one order of magnitude fewer parameters than DreamerV3, TD-MPC2 and LeWM
- SOTA control of continuous environments from pixels
- Accurate system identification across a benchmark of 71 attractors, enabling scientific discovery
The authors argue that world models achieving strong control without capturing true dynamics limits interpretability — world2vec makes dynamics amenable to mechanistic analysis.
More from Research
- Open-source memory layer ICE loses to vector-RAG on LongMemEval: 43% vs 69.5% — Deepnar_S · 2026-11-01
- Tiny 109M model beats 72B models at image-text alignment with 42x less data — Lazylion2 · 2026-11-01
- Can you name these three ML architectures from blank diagrams? — cneuralnetwork · 2026-11-01
- TIL: Boolean circuits can be cyclic — NOR-gate latches and Rivest-style loops — thomasahle · 2026-11-01
- Platypus: an open-source continual-learning LM that infers new facts while it "sleeps" — ChrisLLM · 2026-11-01
- The Path of Averaging: Tracing the Line from Attention to TTT — SonglinYang4 · 2026-11-01