MIT team uses OpenAI's GPT-6 Astra to directly edit Qwen's open model weights
Researchers bypassed retraining, letting an AI system rewrite another model's weights to fix repetitive reasoning and lift math scores.
What to know
- Researchers edited an open model's weights directly using another AI system, rather than retraining, to fix a specific performance weakness.
- The gain is attributed to correcting repetitive reasoning patterns in Qwen, though the editing process required about 50 trial-and-error attempts.
- A parallel attempt to apply the same technique to Google's Gemma model to remove sycophancy failed, suggesting the method does not generalize easily.
- Online discussion is split between those who see early signs of recursive self-improvement and those who dispute that LLMs can be said to 'reason' or that weights encode 'repetition' at all.
The dispute Commenters disagree on whether the result meaningfully resembles recursive self-improvement or whether describing the process in terms of LLM 'reasoning' misrepresents what weight editing actually does.
This looks like an early, incomplete form of recursive self-improvement and is a meaningful step beyond AI-generated code claims.
-
“Interesting to finally see something that more credibly resembles recursive self improvement even though this was just low hanging fruit and they didn't technically complete the cycle.”
Slashdot commenter · Slashdot ↗
The framing of LLMs 'reasoning' or weights controlling 'repetition' is technically wrong or misleading.
-
“LLM's don't "reason" much less repetitively. And weight values don't control "repeating" or recursion. Worse, it would seem that without detailed information on the structure of the model it would be impossible to do any aspect of this at…”
Slashdot commenter · Slashdot ↗
The method is still crude and human-dependent, not autonomous: it needed a human trainer to initiate and succeeded through trial and error rather than clean generalization.
-
“It is telling that this had to be initiated by a human trainer, but it does not seem to require further prompting once in an instance”
Slashdot commenter · Slashdot ↗
AI and human labor should be developed together rather than AI being pushed to replace workers.
-
“I am serious when I say that humans and AI are not mutually exclusive, and the benefits are great if managed well”
Slashdot commenter · Slashdot ↗
“The improvement came primarily from correcting repetitive reasoning—a distinction that shaped both the technical findings and the debate over their implications.”
VentureBeat article, reporting summary · Slashdot ↗ · Nov 1
MIT researchers Study authorsGPT-6 Astra OpenAI model used as the editing toolQwen Alibaba's open-weight model, the edit targetGemma Google's open-weight model, a second edit target
How it unfolded 2 developments, newest first · click a bar or a number to jump articlesposts
-
2
Reddit and Slashdot pick up the study, spurring debate
The story was reposted to r/ and discussed on Slashdot, where commenters debated whether the technique constitutes a form of recursive self-improvement and disputed the framing of LLMs 'reasoning' or 'repeating.'
-
1
MIT researchers report GPT-6 Astra editing Qwen's weights directly
A study describes using GPT-6 Astra to manipulate Qwen's open weights without retraining, boosting its math performance by addressing repetitive reasoning patterns; the paper also notes Astra wrote and ran a Python program to perform the edit after roughly 50 trial-and-error attempts, and a similar edit attempted on Gemma to remove sycophancy was unsuccessful.
“Like gene editing for neural networks…”
— study researchers -
2 outlets first by Slashdot, 5d ago · also VentureBeat · read ↗
-