The Trinity reorganized tasks into a capability tree so multimodal agents could generate their own training worlds and evolve reusable harnesses, adding 34 points across 28 training benchmarks
AspireAI community · arXiv (Shanghai Jiao Tong / Shanghai AI Lab)
The Trinity, an arXiv paper submitted October 26, 2026, frames multimodal intelligence as human-guided co-evolution between worlds and agents, and names two bottlenecks: building heterogeneous verifiable environments is costly, and manual harness iteration cannot keep pace with growing task diversity. Its answer is that tasks with different modalities and surface forms can share underlying capability requirements, so the Evolver reorganizes a flat task tree into a hierarchical capability tree based on verified execution trajectories. Creator then abstracts the common requirements of heterogeneous tasks at a node into simplified, verifiable simulated worlds through a Capability-Driven Real2Sim2Real procedure, while evaluation on the original tasks measures transfer, and Explorer progressively evolves reusable capability modules through selective inheritance to form a Harness tree whose validated updates accumulate across iterations. A Capability-Aware Harness Router then picks compatible module configurations from that tree using public task information. The authors report mean performance improving by 34 points across 28 training benchmarks and 26 points across 18 unseen benchmarks, including two real-world embodied tasks where SOON went from 0 to 80% in simulation and 0 to 60% on the robot, and CaP-X from 10 to 60% and 0 to 60%. The limitation is that the gains come with substantial self-play compute, and the headline numbers are averages across a benchmark set the authors themselves assembled.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.