News
No need to change model weights
3 min read
Source: Telegram AI频道
No need to change model weights! Google's open source ace framework Harness²: allows AI to self-repair during operation, and running scores soar across the board. Can the performance improvement of AI models only rely on blindly stacking parameters and burning computing power? Google has given a new solution! Recently, Google officially open sourced its latest breakthrough in the field of recursive self-improvement (RSI) - the Harness² framework. This solution breaks the traditional path dependence. There is no need to modify the base weight of the large model. Only by optimizing the peripheral running scaffold (Scaffold), the performance of the large model can achieve a dramatic increase. It is worth mentioning that 78% of the core R&D team of this framework are Chinese. Breaking the two major industry Achilles’ heel: task-level and improvement-level double-layer recursion. Past RSI (recursive self-improvement) solutions often faced two major pain points: first, it is easy to “go astray” when there is no standard answer; second, it is extremely easy to overfit a single task. To this end, Harness² innovatively introduces a two-layer recursive architecture: Task-level recursion (proposal-detection-synthesis mechanism): In open scenarios that lack standard answers, the system verifies effective changes through multi-step comparisons and safely incorporates them into the running scaffolding. It also supports parallel expansion to significantly broaden the search space. Improvement-level recursion (precipitation and reuse of experience library): Precipitate valuable experience during the debugging process into the continuous improvement experience library. When faced with new tasks later, the Agent can directly retrieve and reuse historical experience to achieve continuous and efficient iteration of improved capabilities. The actual test performance is amazing: running scores soar across the board, and cross-domain migration capabilities are fully achieved. In a number of authoritative benchmark tests, Harness²’s actual test performance is outstanding: In the office benchmark WorkBuddy-Bench test, the GLM-5 score after accessing the framework soared by 11.5 points, even surpassing the native Gemini3.5Flash. In multi-domain tests such as JobBench and legal benchmark LAB, model performance has also been significantly improved; by migrating experience libraries across domains, the performance can be further amplified. Opening a new path for runtime cognitive engineering. The launch of this achievement marks that the industry is jumping out of the traditional red ocean of relying solely on base model parameter scaling (Scaling Law), and officially exploring a new path for runtime cognitive engineering that allows Agents to self-perceive and self-repair during operation. With the full open source of Harness², developers will be able to make the AI models in their hands continuously evolve in actual operation at a very low cost, and become smarter as they are used! via AI News (author: AI Base)