When better SI test scores mean worse results on new tasks

Researchers tested a way to improve the framework around an SI agent without retraining its model. Their method limits changes that help familiar tests but fail to carry over to new tasks.

Researchers tested a way to improve the framework around an SI agent without retraining its model. Their method limits changes that help familiar tests but fail to carry over to new tasks.
An SI agent can get better at its practice tasks and worse at unfamiliar ones—even when its underlying model stays unchanged. Researchers from Google Cloud AI Research and several universities tested a method intended to limit that problem. According to The Decoder, their method improved results on unfamiliar tasks while using fewer runtime tokens than a version without those controls.
An SI agent uses a language model inside a “harness”: the prompts, tools, memory, workflows and rules that guide its work. The harness can determine whether the agent reads a file before editing it, recovers from a mistake or presents its result properly. Improving that framework is different from training the model itself.
People have traditionally improved harnesses by reviewing failures and making changes by hand. Automated methods instead ask a language model to rewrite the harness repeatedly using feedback from test tasks. The researchers call this recursive self-improvement: the system changes the framework that controls its own behavior.
Repeated feedback from a limited set of tasks can lead to overfitting—changes that improve familiar-test scores without helping on new tasks. The researchers identify several routes to this problem: rules that suit only one benchmark, proposals that score well by chance, and extra complexity that raises scores without improving broader capability. A benchmark is a set of tasks used to measure performance.
RRSI stands for Regularized Recursive Self-Improvement of Agent Harnesses. Here, regularization means placing controls on the improvement process to limit overfitting. The harness remains fully editable, but each proposal has a cap on the number of independent changes it can contain. That cap shrinks over time: early rounds allow larger rewrites, while later rounds permit only small changes whose effects are easier to trace.
A critic reviews each proposal and rejects changes that build particular task names, solutions or other benchmark-specific tricks into the harness. A separate acceptance rule permits higher computing costs only when they bring a measurable performance gain. Components that no longer help are removed.
The system also records earlier attempts to avoid repeating failed ideas. When progress stalls, it tries changes in parts of the harness it has not yet explored.
The researchers tested eight benchmarks covering coding, office tasks performed by agents and engineering design. They compared RRSI with an unmodified baseline harness and four recent optimization methods. The underlying language model, Claude Opus 4.8, stayed frozen: its learned internal parameters, called weights, did not change.
| Measure | Reported result | Scope or comparison |
|---|---|---|
| Score gain on familiar tasks | Up to 14.1 points | Tasks used to optimize the harness |
| Score gain on unfamiliar tasks | Up to 4.7 points | Five benchmarks not used during optimization |
| Runtime token use | About 30% lower | Compared with the unregularized variant, not the unmodified baseline |
All the optimization methods did well on familiar tasks, but two performed below the unmodified baseline on unfamiliar ones. RRSI had the smallest gain on familiar tasks and improved on all five unseen benchmarks, with its largest gain on JobBench. The reported trade-off was less improvement on the tasks used for optimization, but better results beyond them.
Tokens are pieces of text a model processes or generates. RRSI used the fewest tokens and steps among the optimized harnesses, according to the researchers, but the unmodified baseline remained leaner. The lower token use therefore does not mean RRSI was cheaper to run than the original harness.
The authors limit their findings to harnesses built around frozen models; they did not study systems whose model weights also change. Whether independent testing would reproduce the gains, or whether the approach would work alongside model training, remains open.

The award will recognize climate learning on diplomas or transcripts. A California bill summary describes voluntary participation and deadlines for setting the award criteria.
2 min

Research reviews examine pain mechanisms and sex chromosomes, while outside experts raise questions about menopause perspectives and the path into care.

IEA global averages favor electric cars and heat pumps, but upfront costs, financing and other barriers complicate the switch.
3 min
4 min