Google has recently open-sourced its Recursive Self-Improvement (RSI) framework, Harness², a tool that substantially boosts model performance. It achieves this by refining the surrounding operational infrastructure—often referred to as "scaffolding"—without altering the core weights of large models. Harness² features a two-tiered recursive architecture that operates at both the task and improvement levels. This innovative design effectively tackles two key challenges that plagued earlier RSI solutions: the tendency to drift off-target in the absence of standard answers and the propensity to overfit to individual tasks.

At the task level, Harness² utilizes a recursive mechanism that validates beneficial modifications and incorporates them into the scaffolding through a propose-probe-synthesize process. This is particularly useful in scenarios where standard answers are unavailable, and it also supports parallel expansion to widen the search space. Meanwhile, at the improvement level, the framework accumulates debugging experiences into a continuously evolving library. This library can then be accessed and reused for subsequent new tasks, fostering a cycle of ongoing improvement.

Practical tests have demonstrated the effectiveness of Harness². After integrating the framework, the GLM-5 model saw a significant boost in its score on the office benchmark WorkBuddy-Bench, jumping by 11.5 points and outperforming the native Gemini 3.5 Flash. Impressive performance gains were also observed in tests such as JobBench and the legal benchmark LAB. Furthermore, the cross-domain transfer experience library within Harness² has the potential to further enhance model performance across various domains.

Harness² represents a departure from the traditional approach of relying solely on scaling base model parameters. Instead, it explores a new avenue in Runtime cognitive engineering, empowering agents to self-assess and self-correct during operation. An interesting note is that 78% of the core team members behind Harness² are of Chinese descent, highlighting the global collaboration and diversity in cutting-edge AI research.