The Trinity evolves multimodal-agent harnesses from shared capabilities
Original title: The Trinity:Capability-Driven Recursive Self-Improvement for Multimodal Agents
The authors report mean gains of 34 points on 28 training benchmarks and 26 points on 18 unseen benchmarks.
The Trinity is a recursive self-improvement framework that couples world generation with harness evolution through Creator, Explorer, and Evolver roles. Evolver reorganizes tasks into a hierarchical capability tree from verified execution trajectories, and Creator builds simplified simulated worlds for shared requirements while evaluation stays on the original tasks. Explorer grows reusable capability modules into a harness tree, and a Capability-Aware Harness Router picks module configurations from public task information. The paper says evolved harnesses raise mean performance by 34 points across 28 training benchmarks and 26 points across 18 unseen benchmarks, including two real-world embodied tasks.
Unconfirmed: The 34- and 26-point mean gains are claims from the paper, not independently verified.
- Roles are Creator, Explorer, and Evolver.
- Capability-Driven Real2Sim2Real abstracts shared requirements into verifiable simulations.
- Harness updates accumulate inside a fixed topology.
- Gains include two real-world embodied tasks among the unseen set.
From the original article
1 min read at huggingface.co
Multimodal intelligence can be viewed as human-guided co-evolution between worlds and agents through interaction. However, scaling this process faces two bottlenecks: constructing heterogeneous, verifiable environments is costly, and manual Harness iteration cannot keep pace with growing task diversity and complexity.
By Weidong Tang, Ruiheng Wang, Jingxuan Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.