Research review / trendwatch

Models That Rewrite Their Weights Spark Debate About Improvement

A November 20 discussion questions what a model editing its own weights actually demonstrates. The paper's accepted changes and withdrawals provide a narrower, testable story than open-ended improvement claims.

By Sandeep Belgavi · Published 2026-11-20

The November 20, 2026 Hacker News archive records a discussion of models rewriting their own weights. This entry covers that dated debate. The underlying paper was submitted November 16 and revised November 18, rather than first published on November 20.

The dispute concerns the scope of improvement

Participants question whether targeted changes to excessive reasoning support broader self-improvement claims, and whether gradient-free changes should be distinguished from training. These are positions in a particular discussion, not a measured consensus.

The paper reports both accepted and withdrawn edits

The paper's second protocol has a 27B model withdraw changes in five of eight sessions and submit three. Session eight scores 63.8%, compared with a 57.2% baseline, across 490 held-out problems with eight answers per problem. Answers exhausting the token budget fall from 33.6% to 26.1%. These are the authors' experimental results, not a local reproduction or evidence of indefinite improvement.

A proposed gate that also records rejected patches

Original exercise, not executed: evaluate synthetic candidate parameter patches against three frozen checks: answer accuracy, completion within a token budget and an unrelated regression suite. Include a patch that improves completion while failing the regression check, and retain its rejection in the report.

The exercise would explain acceptance bookkeeping without claiming to reproduce model self-editing. Define the rejection rule before inspecting candidate results.

Sources accessed November 21, 2026. Views are my own and do not represent any employer, past or present.

I find the withdrawals as instructive as the accepted edits. I would require an external acceptance gate with frozen tests and unrelated regression checks before trusting another proposed change. The reported reduction in unfinished answers gives me a concrete hypothesis to test; it does not justify assuming that future weight edits will improve general capability.

Sources and evidence

Separate a bounded, evaluated weight change from claims of continuing improvement, and preserve rejected candidates in the evidence record.