🔍 Read the full analysis: How GPT-6 Astra Changed Another Model’s Weights To Address A Loop on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 2 blog post reports that GPT-6 Astra modified two layers of the open-weight Qwen3.5-0.8B model without gradient descent. On MATH-500, the edited model scored 48.1%, up from 7.9%, while the share of answers hitting the 32,768-token limit fell from 93.0% to 33.1%. The results are early, not peer-reviewed, and the loops persisted.
Four researchers affiliated with MIT CSAIL report that OpenAI’s GPT-6 Astra changed two weight tensors in the open model Qwen3.5-0.8B, raising its MATH-500 score from 7.9% to 48.1% without gradient descent. The edit also reduced the share of answers that ran into the model’s 32,768-token limit, from 93.0% to 33.1%—but did not eliminate the looping behavior.
The work was described in a blog post published November 2 by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. The post, titled “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights,” includes experiment logs and calls the findings “very early results.” It is not a peer-reviewed paper, according to the source report, and it is not listed on arXiv.
In the reported experiment, GPT-6 Astra used a Python-based research harness to inspect and edit the model’s weights. The researchers prohibited gradient descent, backward passes and fitting weights to a loss. Astra instead used iterative testing and feedback, with the successful run involving 158 turns, 80 Python executions and 50 forward passes over as many as 512 problems each. The run took 20.3 hours, according to the published account.
The researchers say Astra traced repetitive answers to a single internal direction associated with the model’s MLPs in layers 15 and 19. It subtracted that direction from 64 output columns in each layer, changing 131,072 weights—reported as 0.016% of the model’s 0.8 billion parameters. The edited Hugging Face model reportedly differs from the original only in the two layers’ down-projection tensors.
The result depended on the amount of testing. With 10 forward passes, the MATH-500 score was 12.3%; with 25, it was 14.1%; and with 50, it reached 48.1%. The proportion of answers reaching the token limit was 87.2%, 86.4% and 33.1%, respectively. On GSM8K, the reported score rose from 13.8% to 67.1% in the 50-pass run; on IFEval, it rose from 21.0% to 62.3%.
A Targeted Repair, Not New Math Ability
The result is notable because it suggests a language model can help inspect and directly modify another model’s open weights, rather than only providing advice about code or prompts. If independently replicated, this kind of method could offer developers a way to address specific failure patterns in models they can access internally.
But the benchmark gains should not be read as proof that Astra taught Qwen new mathematical reasoning. The reported intervention targeted repetitive outputs, a weakness identified in the model’s behavior; the researchers distinguish fixing loops from making a model better at math. The improvement also required a substantial search budget, and the final model still hit its token limit on about one in three MATH-500 answers.
The method has potential risks as well as uses. The source report notes prior research describing single-direction weight edits as a way to remove refusal behavior from open models. A targeted edit can change a model’s conduct, but the effects may depend on the model and task, and the Qwen result alone does not establish how reliably such edits can be controlled.
Popular right now
Breville BES870BTR Barista Express Espresso Machine, Black Truffle
De'Longhi Magnifica Duo Fully Automatic Espresso Machine with Milk Frother, Silver
Ninja Luxe Café Mini Plus Espresso and Drip Coffee Machine with AutoFroth, Stainless Steel (ES351)
As an affiliate, we earn on qualifying purchases.
How the Qwen Experiment Was Run
Qwen3.5-0.8B is a small, open-weight model. According to the report, its model card warns that it is prone to thinking loops. In the researchers’ early practice tests, 85 of 96 answers reached the 32,768-token cap, illustrating the problem they chose to address.
Astra did not have unrestricted access to the internet or the benchmark questions used for final scoring. The researchers’ setup was CPU-only and limited Astra to at most 300 code runs. It provided practice questions drawn from GSM8K and MATH, separate from the final evaluation sets. Astra could not run Qwen directly: it had to use a limited number of forward passes to observe the model’s responses and guide its edits.
The post describes eight experiments involving different tasks, not all of them successful. The source account says Astra could not identify which of two fine-tuned Gemma models had been trained to flatter users. It also reports that edits to Gemma 3 270M changed scores by no more than about a point, and that a loop-free model did not improve under this method. Those results suggest the Qwen outcome may not generalize to other models or problems.
““Fixing loops, however, is not the same as making a model better at math.””
— The researchers, as quoted in the project post
Replication and Generalization Remain Open
The reported findings come from a blog post and its logs, not a peer-reviewed study. The available source material does not establish whether independent researchers have reproduced the edit, whether the benchmark results hold across different sampling choices, or how the modification affects other tasks and forms of mathematical reasoning.
It is also unclear how the edit would perform on different model sizes or architectures. The reported limited changes on Gemma 3 270M and the failed model-identification experiment are reasons not to assume that the technique works broadly. The source report says OpenAI did not author or acknowledge the post and that it found no statement from the company about the experiment. No independent assessment of the results is provided in the supplied material.
The reduction in looping is substantial in the reported comparison, but the behavior was not gone: 33.1% of edited-model MATH-500 answers still reached the token limit. The source material also does not provide enough information to establish whether the score gains would persist under other evaluation protocols or in real-world use.
Independent Tests Will Settle Reach
The immediate next step is independent reproduction of the reported weight edit and benchmark results. Researchers would need to check whether the published changes yield similar outcomes under the same settings, then test whether the gains persist on additional tasks and with other model families.
Further evaluation could also clarify trade-offs: whether reduced repetition affects answer quality, whether other failure modes emerge, and how much testing is needed before an edit can be judged reliable. The supplied source material does not announce a schedule for a peer-reviewed paper or follow-up experiments. Until such evidence appears, the result is best treated as a promising, early report about a targeted change to one open model.
Key Questions
What did GPT-6 Astra change?
According to the researchers’ post, Astra edited the down-projection weight tensors in layers 15 and 19 of Qwen3.5-0.8B, subtracting a direction associated with repetitive outputs.
Did the researchers train Qwen with gradient descent?
No. The reported setup prohibited gradient descent, backward passes and loss-based weight fitting. Astra used an iterative search with Python tools and forward-pass feedback instead.
Were the loops eliminated?
No. On MATH-500, the share of answers reaching the 32,768-token limit fell from 93.0% to 33.1% in the 50-pass run, so the reported edit reduced but did not remove the behavior.
How established are the results?
The findings were published in a blog post described by its authors as very early. The source report says it is not peer-reviewed and does not describe independent replication.
Does the result show that Astra can improve any model?
No. The post includes experiments with mixed outcomes, including little change on Gemma 3 270M and a task Astra could not solve. The Qwen result does not establish that the approach generalizes to other models.
Source: ThorstenMeyerAI.com
