· via Hacker News – Front Page (native)
Open 27B model rewrites its own weights and improves itself without training
An arXiv preprint describes an open 27B model that edits its own weights through reasoning alone, gaining 6.6 accuracy points on unseen problems with no training run.

The finding
A preprint posted to arXiv claims what its authors believe is the first demonstration of a language model improving itself by directly rewriting its own weights, with no gradients and no training run. The paper, submitted on 16 November and revised on 18 November, reached the Hacker News front page this week.
The authors — a team led by Jiajun Fan — call the technique "self-rewriting," meaning the model effectively serves as its own optimizer. The subject is an open 27-billion-parameter model. According to the paper, when given access to its own weights, its graded answers, and the ability to execute itself, the model can diagnose why it got a problem wrong and then edit the weights responsible for the failure. The authors report that the model's own edit makes it 6.6 points more accurate on problems it had never encountered.
The work builds on earlier research in which a frontier model edited a much smaller one. The difference here is that a single open model performs the diagnosis and the surgery on itself.
Iteration and self-repair
The process is not a one-shot patch. The abstract describes the model testing candidate edits on practice problems and keeping only those that help, with each accepted edit building on the previous one. Across eight consecutive sessions, the authors report, accuracy only rose.
The more dramatic demonstration is self-repair. The researchers damaged a copy of the model so severely that it could solve almost nothing. Without any retraining, and without ever reading the original weights, the model restored the broken copy to more than 80 percent of its original accuracy on image tasks it had never seen. The authors compare this to a machine mending itself rather than being rebuilt.
The capability also appears to depend on scale. Within one model family, the paper reports, smaller models either made no edits at all or actively degraded themselves; only the largest model both improved and repaired itself.
A third path between training and inference
The authors position self-rewriting between the two standard ways of making a frozen model better. Test-time techniques — prompts, harnesses, extra computation per query — can raise accuracy, but the gain disappears when the task ends and adds cost to every answer. Training, meanwhile, requires gradient updates and dedicated compute.
Self-rewriting, by the authors' account, combines advantages of both: the improvement is written into the weights, so subsequent answers cost no more than before, and the process runs only forward passes, like ordinary model serving. The paper argues that the hard boundary between training and inference reflects the limits of current tooling rather than anything fundamental about intelligence, and frames self-rewriting as a step toward recursive self-improvement — a model that keeps upgrading itself without a larger teacher, a scaffolding harness, or prompt engineering.
What remains unverified
The claims above come from a preprint abstract and have not yet been through peer review. The abstract does not specify how the weight edits are implemented, which benchmarks produced the accuracy figures, how the damaged copy was created, or which model family was tested. Readers should treat the numbers as the authors' own reporting until the full paper is examined and replicated.
Why it matters
If the results hold up, they sketch a route to model improvement that sidesteps the most expensive part of the AI pipeline: the training run. A deployed model that can diagnose and patch its own failures could adapt in place, at the cost of inference alone, and could in principle repair corrupted or degraded copies of itself.
That prospect cuts both ways. Self-improving systems that modify their own weights outside any training or evaluation loop raise obvious oversight questions, since improvements, mistakes and misuse would all be written directly into the model. The paper's own finding that smaller models break themselves while only the largest succeeds suggests the capability arrives unevenly and is not easy to control. Expect the safety discussion to follow the technical one closely.
- #language-models
- #machine-learning
- #self-improvement
- #arxiv
- #ai-safety