Paper review · October 29, 2026
The Trinity: Capability-driven recursive self-improvement for multimodal agents
The Trinity:Capability-Driven Recursive Self-Improvement for Multimodal Agents
The method reorganizes tasks around shared capabilities rather than surface task types, automatically generates simulation worlds to practice those capabilities, and evolves the Harness. This gave average gains of 34 points on training benchmarks and 26 points on unseen benchmarks.
Problem
Improving multimodal agents requires constant work on diverse verifiable environments and on the Harness (the control code and modules that wrap the model), which is costly for people to do. Prior work on Harness self-improvement optimizes per task on text tasks such as coding and math, so it generalizes poorly to heterogeneous tasks.
Method
The system has three roles: Creator, Explorer and Evolver. The Evolver extracts the capabilities each task requires from verified execution Trajectories and restructures the task tree into a hierarchical capability tree. The Creator turns the shared requirements of tasks under the same node into simplified, verifiable simulation worlds (Capability-Driven Real2Sim2Real). The Explorer follows the capability tree, selectively inheriting and locally improving six protocol modules such as Perception, Memory and Planning, to grow a Harness tree. The backbone model stays frozen. At inference time, a Capability-Aware Harness Router picks a suitable Harness branch based on public task information and past experience.
Results
- The authors reported that the evolved Harness improved by 34 points on average across the 28 benchmarks used for training.
- On 18 benchmarks not used for training, it improved by 26 points on average, including 2 real-world embodied tasks.
Limitations
Experimental details (baselines, per-benchmark numbers, cost per iteration) are cut off in the provided text, so the source of the gains cannot be examined closely. Because the model weights are frozen and only the Harness evolves, improvement stays within the backbone's capabilities. Defining capability categories and building the trees involve several LLM agents and reviewers, making the pipeline complex.
Why it matters
In line with the trend that agent performance depends on Harness design as much as on the model itself, the paper presents a way to automatically improve and route Harnesses using shared capabilities, not task types, as the unit. It points to a way for organizations deploying agents across many kinds of work to reduce the burden of per-task tuning.
Key terms
- Harness
- A bundle of control code and modules that wraps a frozen model and coordinates perception, memory, planning, tool selection and the execution loop.
- Recursive self-improvement (RSI)
- A process in which an agent system repeatedly revises its own components based on the results of its own runs.
- Real2Sim2Real
- An approach that moves the core requirements of real tasks into simple simulations for training, then evaluates transfer performance back on the original real tasks.
Original
Authors: Weidong Tang, Ruiheng Wang, Jingxuan Zhang, Hexin Wang, Haoyuan Chen, Zhengfeiyang Zhang, Yixing Ma, Zichen Jiang, et al. · Posted 2026-10-26 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.