Paper claims language models can rewrite their own weights through reasoning, with no gradient training
Summarized from the Hacker News discussion because the article couldn't be read
- A commenter quotes the paper saying a language model rewrites its own weights by its own reasoning and improves without gradients or training.
- Per the quoted text, the model tests its edits on practice problems, submits only edits that work, and its accuracy rises over eight consecutive sessions.
- A commenter reports that the only modification the model made was shortening its reasoning traces to avoid hitting the token limit.
- Commenters report that Appendix E5 groups errors, computes the aggregate state difference, and pushes weights along that direction, approximating gradient descent.
- A commenter notes the paper never mentions alignment, and that the lead author's publication page has no related work.
Hacker News opinions
Honestly the alignment point is the one that gets me. The author never mentions it once, and a commenter checked the lead author's publication page and found nothing on it there either.
Wait, so the whole 'learning' is just the model grading its own homework? I want to see the test set before I believe the accuracy going up over eight sessions.
The test material is just interaction with the environment though. That's how learning works. The model isn't cheating, it's doing the same thing we do.
I'd push back a bit. The only edit the commenter says it made was shortening its reasoning traces so it stopped hitting the token limit. That's a pretty obvious fix, not deep insight.
Same thing with the weight-space programming post from rikiyat. The author there found the infinite loop bug and fixed it, and a commenter says that bug was findable without the LLM at all.
That's a different paper though, different authors and a different approach. Astra in that one can read the weights and write new ones from a spec, so don't lump them together.
I've been messing with tiny models using Astra and it really does seem to see stuff in the weights I can't. Being able to hotfix a trained model and git bisect it is kind of wild.
Appendix E5 is the part worth reading. It groups errors, takes the average state difference, and pushes the weights along that direction. So it's basically reinventing gradient descent by hand, which is clever but not superhuman.