🔍 Read the full analysis: AI Repairs AI: GPT-6 Astra’s Weight Edit In Perspective on ThorstenMeyerAI.com
Get the little things that make your day 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
A November 2 blog post by four researchers affiliated with MIT CSAIL reports that GPT-6 Astra edited two layers of the open-weight model Qwen3.5-0.8B, raising its MATH-500 score from 7.9% to 48.1%. The approach used iterative testing without gradient descent, but it took 20.3 hours and 50 forward passes, and 33.1% of edited-model answers still hit the token limit. The result is an early, non-peer-reviewed report, not evidence that the loop was eliminated or that the model gained broad mathematical ability.
Four researchers affiliated with MIT CSAIL report that OpenAI’s GPT-6 Astra edited 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 November 2 blog post describes an iterative weight-editing experiment, not a peer-reviewed study; the edit reduced but did not eliminate the model’s tendency to run into its token limit.
The researchers—Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab—published “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights” with logs of the experiment. They report that GPT-6 Astra identified a direction associated with repetitive outputs and subtracted it from output columns in the MLP components of layers 15 and 19. The changed weights amount to a reported 0.016% of the model’s 0.8 billion parameters. The edited model and weight files are available on Hugging Face, according to the source account.
On MATH-500, the base model scored 7.9%, while 10-, 25- and 50-pass editing runs scored 12.3%, 14.1% and 48.1%, respectively. The share of answers reaching the 32,768-token limit fell from 93.0% for the base model to 33.1% after the 50-pass run. The researchers also report that the 50-pass version scored 67.1% on GSM8K, compared with 13.8% for the base model, and 62.3% on IFEval, compared with 21.0%.
The process was not a single automatic fix or a training run in the conventional gradient-descent sense. According to the published logs, the winning run involved 158 turns, 80 Python runs and 50 forward passes, and took 20.3 hours. The model had access to practice problems, but the researchers say the final benchmark questions were withheld. Smaller budgets produced much smaller MATH-500 gains: 12.3% after 10 passes and 14.1% after 25.
What Weight Editing Could Change
The result matters because it suggests a capable model may be able to help modify an open-weight model directly, rather than only provide instructions for a human to retrain it. If similar methods prove reliable, developers could investigate targeted edits to address particular failures in models they can inspect and modify. That could be relevant when full retraining is costly or unnecessary for a narrow defect.
But the benchmark gain should be read as a repair of a specific failure mode, not proof of a general advance in mathematical reasoning. The source says Qwen’s model card already warned that the model was prone to thinking loops, and the researchers explicitly distinguish fixing loops from making a model better at math. Even after editing, roughly one-third of MATH-500 answers hit the maximum token length. The reported improvements also depend on repeated model evaluations and a substantial time budget.
There is a broader governance implication for users of models with accessible weights: a small change can alter how a model behaves, and a modified copy is not the same artifact as its original release. The experiment’s authors cite earlier work on removing refusal behavior through weight edits, underscoring that targeted edits can have effects beyond benchmark performance. Whether this particular procedure is dependable across models and tasks remains unestablished.
Products Worth Considering
Garmin Forerunner 55 GPS Running Smartwatch, Black
Garmin Forerunner 165 GPS Running Smartwatch, 43 mm, Black
Garmin Forerunner 570 GPS Running Smartwatch, 47mm, Aluminum, Black
As an affiliate, we earn on qualifying purchases.
How the Experiment Was Run
The post covers eight experiments involving GPT-6 Astra and model weights, including attempts to identify training information and repair a failing driving policy. The source reports that one identification task 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 the eighth experiment, rather than the sole subject of the post.
For the Qwen test, the researchers provided Astra a Python environment to inspect and edit the 24-layer model’s weights. Their log says the setup was CPU-only, without internet access, and allowed up to 300 code runs. Astra could not run the model itself; it had to use limited forward passes to observe outputs. The researchers supplied practice material from GSM8K and MATH, while reserving the final benchmark sets from the editing process.
The post is described as very early results, is not listed as peer-reviewed, and, according to the source account, was not on arXiv at publication. The authors are listed with MIT CSAIL, with Takehi and Teoh marked as equal contributors. The source also says OpenAI was not an author and had made no statement about the experiment. GPT-6 Astra’s role, as presented, was to carry out the researchers’ constrained editing task; the report does not establish that OpenAI sponsored or endorsed the work.
““Fixing loops, however, is not the same as making a model better at math.””
— The researchers, in “Sparks of Weight-Space Programming”
Limits of the Reported Results
The reported scores come from a single blog-post account, and independent replication or peer review is not established in the source material. It is not clear whether the score changes will hold across different sampling settings, repeated benchmark runs, or other versions of Qwen. Nor does the report show whether the edit preserves performance on a broader set of tasks or creates new failure modes.
The looping was reduced, not removed: 33.1% of MATH-500 answers from the 50-pass version still reached the token limit. The source does not provide enough information to settle how much of the score gain comes from improved mathematical accuracy versus the reduced frequency of unproductive, overlong answers. The reported benchmark results should therefore be understood within the test setup and measures described by the authors.
It also remains unclear whether another team could reproduce the edit with the same time and compute budget, or whether the method would work on larger or differently structured models. The post’s claims are about this experiment; they do not show that AI systems can generally improve other models without substantial evaluation or expert oversight.
Replication and Broader Testing
The immediate next step is independent testing of the published model and logs, including checks of the edited weights, the stated benchmark settings and the token-limit results. Researchers would also need to test the edit on additional math problems and unrelated tasks to see whether it is a narrow repair or produces wider behavioral changes.
The source material does not identify a scheduled peer-reviewed paper, a follow-up release or an OpenAI statement. For now, the key milestone is whether other researchers can reproduce the result and establish its limits. Until then, the reported improvement is a promising but preliminary demonstration of weight editing through iterative evaluation, not a settled new training method.
Key Questions
What did GPT-6 Astra change?
According to the researchers’ blog post, Astra edited weight values in the MLP components of layers 15 and 19 of Qwen3.5-0.8B, targeting a direction they associated with repetitive outputs.
Did the Qwen model stop looping?
No. The reported share of MATH-500 answers reaching the 32,768-token limit fell from 93.0% to 33.1% after the 50-pass edit, so the behavior persisted in a substantial share of answers.
Was this done without training?
It was done without gradient descent, according to the post. But the work still involved iterative tests, Python runs and forward passes, and the reported winning run took 20.3 hours.
Has the result been independently verified?
The supplied source describes a blog post published November 2, 2026, and says it was not peer-reviewed and was not on arXiv at that time. Independent replication is not established in the available material.
Does this show AI can generally improve other AI models?
No. It reports one preliminary experiment on Qwen3.5-0.8B, alongside other limited demonstrations. The source does not establish that the approach works reliably across models or tasks.
Source: ThorstenMeyerAI.com
