Compact "Gene" Encoding Beats Documentation for Reusable AI Problem-Solving Experience

A new technical report examines a fundamental question: how should AI systems encode reusable problem-solving experience so it can guide behavior at test time and serve as the basis for ongoing improvement?

Researchers tested this across 4,590 controlled experiments spanning 45 scientific code-solving tasks, comparing different ways of representing and packaging what a system has learned.

Documentation vs. Compactness#

The findings challenge the intuitive approach of capturing experience as detailed documentation-style skill packages. These verbose representations proved unreliable as control signals: the useful information they contained was scattered and sparse, and expanding them into fuller instructional packages often failed to help—sometimes making overall performance worse.

The breakthrough came from a more economical approach. A compact representation scheme called "Gene" consistently outperformed documentation-heavy alternatives. Gene-based systems remained robust when subjected to structural disruptions and beat comparable Skill fragments with the same computational budget. Notably, adding back documentation material to Gene representations typically degraded rather than improved results.

Beyond One-Shot Learning#

The study also examined how experience accumulates over time. Gene proved superior as a substrate for iterative learning, particularly when failure data was attached. When systems evolved through repeated cycles, encoding warnings and failure histories in the compact Gene format outpaced storing the same information as raw Skill fragments or freeform text. The structure of the encoding itself—not just its content—shaped how effectively accumulated experience could guide future behavior.

Real-World Performance Gains#

On the CritPt benchmark, systems evolved using Gene representations improved substantially: base models jumped from 9.1% to 18.57% success, while stronger baselines advanced from 17.7% to 27.14%. These results suggest that the bottleneck in experience reuse is not the quantity of experience but rather its format: the most powerful representation is one engineered for both compact storage and easy evolution.