Can the performance improvement of AI models only be achieved by blindly stacking parameters and consuming computing power? Google has offered a completely new solution!
Recently, Google officially opened source its latest breakthrough in the field of recursive self-improvement (RSI) – the Harness² framework. This solution breaks traditional path dependencies; without modifying the weights of the large model base, simply by optimizing the peripheral running scaffold, the performance of the large model can improve dramatically. It is worth noting that 78% of the core R&D team for this framework consists of Chinese people.
Break through the two major weaknesses of industries: task-level and improvement-level double recursion
Past RSI (Recursive Self-Improvement) solutions often face two major challenges: first, they are prone to “going off track” when there is no standard answer; second, they easily over-fit a single task. To address this, Harness² innovatively introduces a double-layer recursive architecture.
Task-level recursion (proposal-detection-synthesis mechanism): In an open environment without a standard answer, the system uses multi-step verification to effectively apply changes and safely integrate them into the running scaffold, while supporting parallel expansion to greatly expand the search space.
Improved level recursion (experience library accumulation and reuse): Deposit valuable experiences from the debugging process into a continuous improvement experience library. When facing new tasks later, the Agent can directly retrieve and reuse historical experiences, enabling continuous and efficient iteration of improvement capabilities.
Actual performance is impressive: scores soar across all categories, and cross-domain migration capabilities are at full capacity
In multiple authoritative benchmark tests, Harness²'s actual performance is quite impressive:
In the WorkBuddy-Bench testing, the GLM-5 score directly increased by 11.5 points after integrating this framework, even surpassing the native Gemini3.5Flash.
In tests across multiple fields such as JobBench and legal benchmark LAB, the model performance also saw a significant improvement; by leveraging cross-field migration experience databases, performance can be further enhanced.
Enable the new path for Runtime Cognitive Engineering
The introduction of this achievement marks a shift in the industry away from simply relying on the scaling law of base model parameters, and officially opens up a new path for runtime cognitive engineering where agents can self-perceive and self-improve during operation. With the full open source release of Harness², developers will be able to make their AI models continuously evolve and become smarter as they are used, at a very low cost!