Compositional energy models: learn reusable evaluation factors, compose them for new tasks
du_yilun · x · 2026-09-09
Using the example of placing a mug on a crowded table, the author argues that instead of directly predicting the placement, one can learn reusable local evaluation factors — support, clearance, reachability — each judging a local relationship rather than the whole scene; inference composes these factors into an evaluator and searches for a placement satisfying all constraints together. In their compositional-energy work, factors learned on small SAT problems compose to solve larger ones. The principle extends to planning and reasoning: by combining familiar skills, reasoning operations and expertise, a system can construct both a solution and the reasoning process used to find it.
Related event: Du Yilun Proposes Reasoning-Driven Robot Generalization(3 posts)→
More from Research
- MissionReport benchmark: only 3 of 540 AI-generated government reports pass all criteria — Exp_Mark · 2026-10-27
- Finch 1: a geospatial foundation model trained on imagery from 50+ sensors — Scobleizer · 2026-10-27
- Skyfall AI's SKY-001 world model turns sketches into CAD builds in 3.9s, vs 199s for GPT-6 Astra — rohanpaul_ai · 2026-10-27
- Physicists Let AI Pick Its Own Problem—It Wrote a Publishable Paper in 2 Hours — burny_tech · 2026-10-27
- Researcher flags unresolved data transparency issues in compositionality evals — najoungkim · 2026-10-27
- Separate Forms from Estimators: Why Steer by Difference of Means When You Can Use SGD? — johnhewtt · 2026-10-27