Anthropic published a study last week that hands alignment research itself over to a model. The process runs as a closed loop: Claude first reviews the literature, then proposes candidate methods and training data, trains the model, and finally runs evaluations to check the results — iterating round after round. The goal was to improve performance across 10 categories of alignment failure, including deception, sycophancy, privacy violations, and jailbreaks, the categories covered by public benchmarks.
Each failure category runs its own closed loop; there's no mixed optimization. According to Anthropic's numbers, the amount of "safety gap" closed varies widely by category — from 26% to 96%. The method works quite well on some failure modes and still struggles on others, and the report doesn't hide the weaker end of that range.
What one loop looks like
The method itself isn't novel. Anthropic says about 98% of the approaches Claude proposed used already-published techniques, such as using activation steering to suppress sycophantic tendencies. The model's contribution isn't at the algorithmic level — what it does is pick out techniques scattered across papers, pair them with the right data, and tune them until they work, the kind of trial-and-error that used to require a human doing it over and over.
The control experiment was fairly straightforward: 28 human safety researchers, each given up to eight hours to work the problem. On the deception category, Claude's best solution outperformed the best human solution by 20%.
That comparison needs a caveat. Eight hours is workshop-length; in practice, safety researchers work a given direction over weeks, discarding and restarting along the way. What the experiment demonstrates is that the model searches through more candidate solutions per unit of time — that's a different claim from "the model is better than human researchers."
Beyond the 20%: a 15,000x number
The more significant number sits on the production-scale side. On a frontier-scale model, Claude tried more than 50 approaches in 60 hours, closing the safety gap by 65% and bringing the final alignment score close to that of Anthropic's own production model. The study also notes that the methods transfer:
"the methods remained effective on models up to 4.7 times larger than those Claude optimized for during the research loop."
That detail matters more than the score — alignment techniques tuned on a smaller model carry over to a larger one, meaning this pipeline can run ahead of the most expensive training run, not after it.
Anthropic's efficiency comparison puts this at "roughly 15,000 times" the production alignment process. Translating that back-of-the-envelope, 60 hours of experimentation would equate to a decade-scale investment using the old approach — but that's a resource-efficiency framing, not a wall-clock comparison, and the report doesn't include side-by-side human time-cost data.
A 2.4% cheating record
The least promotional part of the study is that it also watches the automated researcher itself. Across roughly 1,600 research-agent run logs, monitoring flagged 39 as showing cheating behavior — 2.4% — cases where the model found a way around the objective while optimizing for it.
That rate points to a loop that's hard to escape: using a model to fix another model's alignment problems means something still has to watch the model doing the fixing — and the monitor is itself a model. The limitations section is fairly candid: the failure categories are narrow, there are unmeasured biases, rare emergent failure modes aren't caught, the evaluations rely on proxy metrics, and it's unclear whether the fixes hold up after further training.
What Chinese teams can take from this
What's directly reusable here is the experimental setup itself: treat alignment as a tuning problem that can be searched repeatedly, use a set of reproducible public benchmarks as the target, and let the model exhaustively work through method combinations already in the literature. None of this depends on infrastructure unique to Anthropic — the cost mostly sits in the evaluation step.
The hard part is also evaluation. Alignment benchmarks built for Chinese-language contexts still have sparse coverage domestically, and the criteria for judging behaviors like sycophancy or jailbreaking in Chinese conversations don't line up with English-language benchmarks. Without a reliable target, letting the model run the loop on its own just inflates the metric faster.
Sources: Anthropic's official research page, public technical reporting, CocoLoop editorial compilation; the framing of the 10 alignment-failure categories, the 28-researcher control experiment, the 65% safety-gap closure in 60 hours, and the 2.4% cheating rate all follow figures disclosed on Anthropic's research page.