Frozen model, scaled exploration: RAX beats prior SOTA by 37.4 points
du_yilun · x · 2026-10-17
RAX scales the agent's budget of online exploration episodes, not model size: execute a strategy, observe what happened, revise perception and control code, execute again—each round building on earlier programs, measurements, and discoveries. The underlying model stays frozen; learning happens in the agent's growing workspace of code, notes, and verified skills.
Across five benchmarks, success climbs from 24.9% at 4 episodes to 62.3% at 16 and 78.6% at 128, versus a 41.2% average for previous SOTA models—a +37.4-point lead. More exploration yields better strategies turned into fully autonomous programs that transfer to unseen scenes.
More from Embodied
- REK for iOS beta runs full physics and robot AI simulation on a phone — cixliv · 2026-10-17
- A laptop expansion card with a liquid CO₂ thruster that survives a 10 m drop — gnukeith · 2026-10-17
- Harness Pro: a $299 desk screen to watch and direct AI agents on every machine — dee_hw · 2026-10-17
- X-Reset eases RL exploration for dexterous hands by resetting to human hand-object states — weichiuma · 2026-10-17
- Princeton's LabOS, Featured in Nature: An AI Co-Scientist for Hypothesis and Experiment Design — MengdiWang10 · 2026-10-17
- Founder's China trip: robotics has no single hub, and robot firms work less hard — DominiqueCAPaul · 2026-10-17