Researchers presented the SIFT framework for accelerating self-supervised learning of AI agents

Researchers presented SIFT (Recursive Self Improvement via Fast Tree-search) — a framework for recursive self-improvement of coding agents. The system allows models to modify their own code while minimizing the computational resources usually required to evaluate the effectiveness of changes.

The main bottleneck of existing solutions is the need to run benchmarks multiple times to test each iteration of the code. In SIFT, this process is replaced by using LLM as a judge that conducts pairwise comparisons of improvement candidates. The results of the comparisons are aggregated using the Bradley-Terry model, forming strength estimates that guide the search in the decision tree.

Expensive test runs in the new architecture are only performed for the most promising nodes, selected based on the preliminary analysis of the judge. This approach significantly reduces CPU-hour consumption and API call costs compared to traditional search methods while maintaining high code generation accuracy.

During testing on the Polyglot benchmark, the framework demonstrated superiority over existing tree search-based self-evolution systems. The study confirms that using intermediate signals from the judge model allows agents to more effectively explore the space of possible modifications without overloading the computational infrastructure.

This development changes the approach to creating autonomous systems capable of self-optimization. The transition from full test enumeration to candidate ranking through LLM-as-a-judge makes the process of iterative code improvement accessible for implementation under tight budget constraints.