Google and Partners Reveal AI Self-Improvement Methods Boosting Code Performance
Researchers from Google, Northwestern University, and Penn State University have published findings on advanced methods for AI agent harnesses to self-improve. The studies introduce techniques like Self-Improvement Memory (SIM) and gated modular edits, which allow AI systems to rewrite their own operational code while retaining context and validating changes. These approaches significantly outperform baseline models, with SIM boosting workplace task scores to 84.75 and gated edits reducing token consumption by nearly ten times compared to unchecked rewrites. The research demonstrates that preserving the rationale behind code modifications is critical for maintaining performance gains across different model architectures.
Editorial responsibility
- No named human review is recorded for this page.
- Reports are grouped by semantic similarity and deterministic rules. Language models may assist titles, summaries, translation and cross-source analysis; the page reads the event directly, while its address stays stable when the title changes.
- Summary covers the current reports
Cross-source coverage
Reporting timeline
Google Paper: Gated Modular Edits Outperform Full Rewrites in Agent Harness Efficiency
A new Google research paper reveals that self-improving agent harnesses achieve optimal results when utilizing gated modular edits rather than unchecked full rewrites. The study demonstrates that while a full rewrite of the agent harness matched accuracy levels, it consumed nearly ten times more tokens per task compared to the gated approach. The gated method involves an optimizer model improving the harness by editing one code module at a time, with each change required to pass validation on held-out tasks and adhere to a token cap. In testing on 40 SWE-bench Lite tasks using Gemini 3.8 Flash, the gated harness solved 40.0% of tasks using 46.7k tokens per task, whereas the unchecked full rewrite solved 42.5% but required 451.6k tokens. Additionally, relying solely on prompt rules resulted in a significant drop in accuracy to 2.5%. The findings suggest that enforcing changes through code-based gates and validation sets is superior to allowing agents to tune their own harnesses without constraints.
Google AI Rewrites Agent Harness, Improving Workplace Task Scores to 84.75
A new Google preprint details an experiment in recursive self-improvement where an AI system was tasked with rewriting its own agent harness. The system logged every modification, the rationale behind it, and the resulting performance outcomes. In a comparative test, the AI-modified harness achieved a score of 84.75 on workplace tasks, significantly outperforming the baseline version which scored 73.42. Additionally, in coding evaluations, the improved system performed better in nine out of ten runs. The study also involved swapping the Opus 5.5 model for Gemini 3.8 Flash to continue the iterative improvement process. This research suggests that future AI models may no longer require starting from scratch, as the ability to iteratively refine their own code allows capabilities to compound over time.
Read sourceGoogle and Penn State Study: AI Self-Improvement Memory Boosts Code Rewrite Performance
A new research paper from Google and Penn State University introduces a method called Self-Improvement Memory (SIM) to enhance how AI agents rewrite their own operational code. The study addresses the challenge where an AI model, after rewriting the harness code for another AI, often scores lower on benchmarks due to a lack of context regarding previous edits. SIM solves this by logging every code change with its targeted problem, reasoning, and measured results. When transferring these optimized harnesses to new models or workspaces, the optimizer reads the log to decide which edits to retain or revise. In a specific test case moving a harness from Claude Opus 5.5 to Gemini 3.8 Flash, SIM improved performance scores from 68.41 to 84.75, significantly outperforming the baseline score of 73.42 achieved without the log. Across ten repeated coding runs, the SIM-enhanced approach finished ahead nine times, averaging a 2.36-point advantage. This finding suggests that preserving the rationale behind AI-driven code modifications is critical for maintaining and carrying over performance gains across different model architectures.