The thing being improved is not the model
A preprint posted to arXiv on Monday takes aim at a specific practice: letting an agent optimise its own harness. The harness, as the authors define it, is everything around a frozen backbone model — the prompts, the control flow, the tooling, the memory and the context management. An agent’s apparent capability is largely a property of that scaffolding, which is why rewriting it has become a way of improving an agent without touching its weights.
The paper’s claim is that unconstrained rewriting overfits. The harness gets very good at the tasks it was evolved against and no better, sometimes worse, at anything else.
Three brakes
The method, called RRSI, applies regularisation in the statistical sense — deliberately limiting the search so the result generalises. It does this three ways: a budget that caps how many edits each candidate may make and shrinks over time, a critic that screens proposals before they are tried, and a pruning step that removes changes which are not earning their place.
The authors, led by Peng Xia with thirteen co-authors, report gains of up to 14.1 points on tasks drawn from the same distribution the harness was evolved on, and up to 4.7 points on out-of-distribution benchmarks. The second number is the one the paper is about. Unregularised evolution can produce the first without the second.
And it is cheaper
The resulting harness runs on about 30 per cent fewer policy tokens than the one produced by unregularised evolution. That is a straightforward consequence of pruning: an evolved harness accumulates instructions, tool definitions and context-handling steps that are never load-bearing, and every one of them is paid for on every call.
Why it lands now
Recursive self-improvement has been argued about all month in terms of capability and control. This paper treats it as an engineering problem with a measurable failure mode, and the failure mode is mundane: a system that gets better at its benchmark and no better at its job.
Caveats
This is a preprint and has not been peer reviewed. The reported gains are the authors’ own measurements on their own evaluation setup, and no independent group has reproduced them. Code and a project page are public, which at least makes reproduction possible.
What to watch
Whether the regularisation survives contact with harnesses other people built. A technique that constrains a search is only as good as the search space it is constraining, and the paper’s harnesses are the paper’s own.
