Motion direction forms a ring in a video model’s activation space
mathemagic1an · x · 2026-07-29
A follow-up to the video-world-models interpretability work shows that motion direction forms a ring-shaped manifold in the model’s activation space.
- Averaging block-5 activations by motion direction and projecting them with PCA produces a circular structure.
- Nearby directions in the physical world remain neighbors in activation space.
- The key point is that the model never saw a controller or explicit actions—only pixels—yet a steerable motion representation still emerges.
More from Research
- An AI digest scans 92 journals every week and turns them into one RSS feed — Afinetheorem · 2026-07-29
- A weekly PDB-synced leaderboard tracks open cofolding models — rishabh16_ · 2026-07-29
- Agentic AI Summit sets robotics and world models session with Sergey Levine and Jim Fan — dawnsongtweets · 2026-07-29
- New guidelines say GenAI still needs human oversight in systematic literature reviews — DrDatta_AIIMS · 2026-07-29
- Wharton labs open-sources AIBO for statistically valid AI behavior experiments — emollick · 2026-07-29
- A Unity RPG prompt pushes multi-agent coding loops to absurd AAA perfection — chongdashu · 2026-07-29