The Trinity couples verifiable world generation with agent harness evolution, gaining +26 on unseen benchmarks
YangYou1991 · x · 2026-11-04
Researchers introduce The Trinity, a capability-driven multimodal recursive self-improvement (RSI) framework that couples verifiable world generation with evolution of agent harnesses.
Key results with model weights fully frozen:
- +34 points average across 28 training benchmarks; +26 points across 18 unseen benchmarks
- Real2Sim2Real pipeline lifts real-robot task success on Unitree G1 and AlohaMini2 from 0% to 60%
The work asks whether tasks are the right unit of evolution for RSI, and how world generation and capability evolution can reinforce each other in a sustained feedback loop — pointing to shared capabilities as the answer. Project page, arXiv paper, GitHub and HF Papers are all public.
More from Embodied
- Walkie One launches as the world's first AI walkie-talkie for screenless AI access — Scobleizer · 2026-11-04
- DISCO, CoRL 2026 best-paper finalist: distilled tree search over continuous actions — davheld · 2026-11-04
- CMU CoRL paper: grasp pretraining boosts dexterous tool-use success by 33.3 points over DP3 — davheld · 2026-11-04
- University of Tokyo 3D-prints life-size musculoskeletal humanoid legs from single TPU material — IanPritchard · 2026-11-04
- Neuracore to demo robot skill deployment on ABB cobots at Dec 2 event — stepjamUK · 2026-11-04
- Training Humanoids in World Models via Differentiable Heightmap Lookups — breadli428 · 2026-11-04