🔍 Read the full analysis: GPT-6 Astra Tried A New Kind Of AI Repair: Editing Another Model on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 2 blog post by four researchers listed with MIT CSAIL reports that GPT-6 Astra edited two layers of the open Qwen3.5-0.8B model without gradient descent. On MATH-500, the edited model scored 48.1%, up from 7.9%, while answers hitting the token limit fell from 93.0% to 33.1%. The results are early, not peer-reviewed, and the looping behavior remained in about a third of test answers.
Four researchers listed with MIT CSAIL report that OpenAI’s GPT-6 Astra edited the weights of the open Qwen3.5-0.8B model to reduce repetitive answers and improve its math benchmark scores, without using gradient descent. In their November 2 blog post, the researchers say the model’s MATH-500 score rose from 7.9% to 48.1%; the share of answers reaching the 32,768-token limit fell from 93.0% to 33.1%, meaning the loops were reduced, not eliminated.
The researchers—Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab—describe the work in a post titled “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights.” The post is accompanied by logs and an edited model on Hugging Face. The authors call the work “very early results”; it is not a peer-reviewed paper, according to the source material.
According to the published logs, Astra located a direction associated with looping in the MLP outputs of layers 15 and 19. It removed that direction from 64 output columns in each layer, changing 131,072 weights—about 0.016% of the model’s 0.8 billion weights. The researchers say the released model’s only changes from the original are in the two corresponding down-projection weight tensors.
The 48.1% MATH-500 result followed an iterative process with a budget of 50 forward passes, alongside 158 turns and 80 Python runs over 20.3 hours, according to the post. The researchers also report that the edited model scored 67.1% on GSM8K, compared with 13.8% for the base model, and 62.3% on IFEval, compared with 21.0%. The benchmarks used for final scores were not shown to Astra, the authors say.
A Targeted Fix for Open Models
The result offers an example of an AI system making a direct, localized change to another model’s weights, rather than producing an answer or modifying the model through gradient-based training. If the approach can be reproduced, it could give developers another way to diagnose and patch specific behaviors in models whose weights they can access.
But this was a targeted repair, not evidence that Astra broadly taught Qwen new mathematical reasoning skills. The source says Qwen’s model card already warns of thinking loops, and the researchers themselves caution that “Fixing loops, however, is not the same as making a model better at math.” The score gains are reported on benchmarks after a lengthy, feedback-driven search, not from an automated one-step edit.
The method also depends on access to the target model’s weights. It does not show that an outside system can make equivalent changes to a closed model. And an edited checkpoint is a modified model that needs its own testing: a patch intended to reduce one failure could affect other tasks or behaviors.
Products Worth Considering
Samsung Galaxy Tab S10 Lite – 6 GB RAM, 128 GB storage, up to 2 TB microSD, 10.9‑inch LCD, Gray
Apple MacBook Air 13-inch with M5 chip, 16GB memory, 512GB SSD, and 12MP Center Stage camera
Apple MacBook Air 15-inch with M5 chip, 16GB Memory, 512GB SSD, and Liquid Retina Display
As an affiliate, we earn on qualifying purchases.
How the Weight Search Worked
The experiment was one of eight in the researchers’ post, which describes attempts to use a language model to inspect or alter neural-network weights. Another attempt could not determine from weights alone which of two fine-tuned Gemma models had been trained to flatter users, according to the authors. The Qwen repair was the experiment that produced the largest reported benchmark change in the material supplied.
Astra did not train Qwen through gradient descent, but “without training” does not mean without testing or computation. The researchers gave it a Python environment and access to practice problems, while limiting how often it could run Qwen to observe outputs. In the 50-pass run, a pass could cover up to 512 problems. The smaller 10- and 25-pass runs produced MATH-500 scores of 12.3% and 14.1%, respectively, compared with 7.9% for the base model.
The post identifies OpenAI’s model as GPT-6 Astra, run through the Codex command-line tool at the highest reasoning setting. The researchers are not identified as OpenAI authors, and the source material reports no OpenAI statement about the experiment. The work also cites earlier research on removing refusal behavior by changing a direction in model weights, placing this experiment within a developing line of weight-editing research.
““Fixing loops, however, is not the same as making a model better at math.””
— The researchers, in “Sparks of Weight-Space Programming”
How Reliable Is the Repair?
The findings have not been independently established in a peer-reviewed paper, and the supplied account does not describe an external replication. The reported benchmark gains therefore remain the authors’ results. It is also unclear how often the same approach would work on other models, other failure modes or different evaluation sets.
The intervention did not remove looping completely: 33.1% of MATH-500 answers from the edited model still reached the 32,768-token limit. The source also reports that on Gemma 3 270M, no edit improved scores by more than about a point, and that a model without loops did not improve in the relevant test. Those findings limit claims that weight editing alone reliably boosts performance.
The available material does not establish whether the edit affects performance outside the listed evaluations, how robust the result is to different sampling settings, or whether another team can reproduce the same direction search. The authors’ logs and released checkpoint provide material for scrutiny, but do not resolve those questions by themselves.
Replication and Broader Testing
The immediate next step is independent testing of the released edited checkpoint against the original model, including checks for looping, math accuracy and unintended changes on other tasks. The researchers’ published logs and tensor hashes give others details to compare, but the supplied source does not announce a follow-up paper or a planned replication.
Further work would need to establish whether the method transfers beyond this small Qwen model and whether its gains hold under different prompts, sampling settings and evaluation conditions. Until such evidence is available, the result is best read as a specific, promising repair experiment, not a general demonstration that one AI can reliably improve another.
Key Questions
What did GPT-6 Astra change?
According to the researchers, Astra identified and removed a direction linked to repetitive outputs in the MLPs of Qwen3.5-0.8B’s layers 15 and 19. The edit changed 131,072 weights across two layer tensors.
Did the edited model stop looping?
No. On MATH-500, the share of answers reaching the 32,768-token limit fell from 93.0% to 33.1%, according to the post. The behavior was reduced, not eliminated.
Was Qwen trained by gradient descent?
The researchers say they did not use gradient descent, backward passes or fitting to a loss. They did use an iterative search with test feedback, Python runs and up to 50 forward passes in the highest-budget experiment.
Are the results peer-reviewed?
No peer-reviewed publication is identified in the source material. The findings appeared in a November 2, 2026 blog post by four researchers listed with MIT CSAIL, who described the work as very early.
Can this method be used on any AI model?
The report does not establish that. The experiment required access to Qwen’s weights, and the source describes weaker results in another model test. How well the approach transfers to other models or behaviors remains unknown.
Source: ThorstenMeyerAI.com
