🔍 Read the full analysis: Can Editing Weights Repair An AI Loop? Astra’s Case on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Four researchers affiliated with MIT CSAIL report that GPT-6 Astra edited two layers of the open-weight Qwen3.5-0.8B model, raising its MATH-500 score from 7.9% to 48.1%. The result comes from an early, non-peer-reviewed blog post and a substantial iterative search; looping fell but remained, with 33.1% of edited-model answers still reaching the token cap.
Four researchers affiliated with MIT CSAIL report that OpenAI’s GPT-6 Astra modified two layers of the open-weight model Qwen3.5-0.8B, raising its reported MATH-500 score from 7.9% to 48.1% without gradient descent. The change reduced, but did not eliminate, repetitive answers: 33.1% of edited-model responses still reached the benchmark’s 32,768-token limit, according to the researchers’ November 2 blog post.
The post, titled “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights,” says Astra identified an internal direction associated with looping and removed it from output weights in layers 15 and 19 of Qwen3.5-0.8B. The authors report changing 131,072 weights, or 0.016% of the model’s 0.8 billion weights. A published version on Hugging Face and the researchers’ logs are cited in the source material as records of the edit.
The reported MATH-500 comparison used the same sampling settings for the base and edited model. The base scored 7.9%, while 10 and 25 forward-pass runs reached 12.3% and 14.1%; the 50-pass run reached 48.1%. The share of responses hitting the token cap fell from 93.0% to 33.1% in the 50-pass result. On GSM8K, the reported score rose from 13.8% to 67.1% at 50 passes; on IFEval, it rose from 21.0% to 62.3%.
The term “without training” needs qualification. The researchers say they prohibited gradient descent, backward passes and fitting weights to a loss, but Astra still conducted an iterative search using model feedback. The 50-pass run took 20.3 hours and included 158 turns and 80 Python runs, according to the published account. Astra did not see the final benchmark questions, the authors say; it used practice problems from GSM8K and MATH instead.
A Repair, Not a Math Breakthrough
The result matters because it suggests a language model may be able to help diagnose and alter an open model’s behavior directly in its weights, rather than relying on conventional gradient-based fine-tuning. For teams running models they can inspect and modify, a targeted repair could offer another way to address a specific failure, such as repetitive generation. The reported change was small relative to the model, but it required a search process and access to the model’s weights.
The benchmark gains should not be read as proof that Astra gave Qwen a general new mathematical ability. The researchers characterize the work as repairing a weakness, and the source says Qwen’s model card already warns of thinking loops. The authors put the distinction plainly: “Fixing loops, however, is not the same as making a model better at math.” Because many base responses consumed their token budget without producing a useful answer, reducing loops can raise measured accuracy by allowing more answers to finish.
The method also carries a broader implication for model control. The source notes that single-direction weight edits have been studied as a way to remove refusal behavior from open chat models. That does not establish that the same edit or risks apply in every case, but it underlines that small weight changes can affect behavior in ways that developers may want to inspect before deploying a modified model.
Products Worth Considering
Dell XPS 13 9350 Copilot+ Laptop, 13.4-inch 120Hz Display, Intel Core Ultra 7 256V, 16GB RAM, 512GB SSD, Ubuntu Linux 24.04 LTS
Apple MacBook Pro 14 2021 with M1 Max, 64GB RAM, 2TB SSD, Silver
Apple MacBook Air 13-inch with M5 chip, 16GB memory, 1TB SSD, and 12MP Center Stage camera (Midnight)
As an affiliate, we earn on qualifying purchases.
How the Weight Search Worked
The experiment was one of eight described in the blog post, which also reports attempts to identify what a small network was trained to do and to repair a failing car-driving policy. The post says one experiment failed: Astra could not determine from weights alone which of two fine-tuned Gemma models had been trained to flatter users. The Qwen math test was therefore one result in a broader, mixed set of demonstrations, not evidence that weight inspection works reliably across tasks.
For the Qwen experiment, Astra ran as gpt-6-astra through the Codex command-line tool at the highest reasoning effort, according to the logs described in the source. Built-in tools were disabled; the researchers’ harness provided a Python kernel to work with the model’s weights. The setup was CPU-only and had no internet access, with a limit of 300 code runs. The authors also supplied practice sets containing 7,469 GSM8K problems and 7,275 MATH problems.
Astra could not run Qwen freely: each forward pass was a limited opportunity to observe its outputs. A pass could cover up to 512 problems or read up to 1,024 texts. The researchers compared budgets of 10, 25 and 50 passes. The modest results at lower budgets—12.3% and 14.1% on MATH-500—show that the large reported jump did not appear with little feedback. The source identifies the authors as Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, with Takehi and Teoh listed as equal contributors.
““Fixing loops, however, is not the same as making a model better at math.””
— The authors of “Sparks of Weight-Space Programming”
What the Results Do Not Establish
The source describes the findings as a blog post published November 2, 2026, not a peer-reviewed paper or an arXiv publication. The reported scores and logs are attributable to the researchers; independent replication is not established in the supplied material. The account also says OpenAI is not listed as an author and that no statement from OpenAI about the experiment was found.
It remains unclear how reliably the method would work on other models, tasks or benchmark settings, and whether the improvement would hold under independent evaluation. The post includes experiments that did not succeed, while the Qwen result required the largest tested feedback budget. The available information does not establish the method’s cost or reliability in production, nor whether the modified model would retain other behaviors important to users.
Looping was substantially reduced in the reported run, not removed. Even after the edit, roughly one in three MATH-500 answers reached the token limit. The results also do not show that a model without looping would necessarily gain accuracy: the source says that a model without loops did not improve in the reported comparison, and that edits to Gemma 3 270M moved its score by no more than about a point.
Replication and Broader Testing
The next useful test is whether independent researchers can reproduce the weight edit and benchmark changes using the published logs and model files. The source material identifies no scheduled peer review or follow-up release, so a formal next milestone is not confirmed. Further work would need to test whether the approach generalizes beyond this small Qwen model and whether improvements persist across tasks and evaluation setups.
For now, readers can treat the reported numbers as an early result with a defined scope: one model, one loop-related weakness and a time-intensive search that avoided gradient descent but used repeated feedback. The key unresolved issue is whether this can become a dependable method for modifying open models, rather than a striking demonstration limited to a particular experiment.
Key Questions
What did GPT-6 Astra change in Qwen3.5-0.8B?
According to the researchers, Astra removed an internal direction associated with looping from output weights in layers 15 and 19. They report changing 131,072 weights, a small share of the model’s total weights.
Did the edit eliminate the model’s loops?
No. The reported share of MATH-500 answers reaching the 32,768-token limit fell from 93.0% to 33.1% in the 50-pass run. The researchers’ figures show a reduction, not a complete fix.
Was this done without any training or testing?
It was done without gradient descent, backward passes or fitting weights to a loss, according to the authors. But Astra used iterative test feedback and a substantial search process, including 50 forward passes in the strongest reported run.
Has the result been peer-reviewed?
The supplied source describes it as a blog post published November 2, 2026, and says it was not on arXiv and had not been peer-reviewed at the time of writing. Independent replication is not confirmed in the source material.
Source: ThorstenMeyerAI.com
