🔍 Read the full analysis: GPT-6 Astra Edited Another AI: A Repair Story, Not An IQ Story on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 2, 2026 blog post by four researchers listed with MIT CSAIL reports 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 work used iterative searches and test feedback rather than gradient descent, and it reduced but did not eliminate answers that ran into the token limit. The result is an early, non-peer-reviewed report, not evidence that Astra broadly improved another AI’s reasoning ability.
Four researchers listed 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%. The November 2 blog post describes an iterative search that changed model weights without gradient descent; it also reports that repetitive answers fell sharply but continued in 33.1% of test answers.
The experiment targeted a specific failure: Qwen3.5-0.8B often repeated its reasoning until reaching a 32,768-token limit. According to the researchers’ post and logs, 93.0% of the base model’s MATH-500 answers reached that limit. Astra identified a direction associated with the loops and subtracted it from output columns in the model’s layers 15 and 19, changing 131,072 weights, or about 0.016% of the model’s 0.8 billion weights.
The reported MATH-500 score improved as the search budget increased: runs with 10 and 25 forward passes scored 12.3% and 14.1%, while the 50-pass run scored 48.1%. The share of answers hitting the token limit was 87.2%, 86.4% and 33.1%, respectively. On GSM8K, the 50-pass version scored 67.1%, compared with 13.8% for the base model; on IFEval, it scored 62.3%, compared with 21.0%.
The authors say the highest-budget run took 20.3 hours and involved 158 turns, 80 Python runs and 50 forward passes. Each pass could evaluate up to 512 problems. The researchers say Astra did not see the final benchmark questions, and that gradient descent, backward passes and fitting weights to a loss were prohibited. The edited model and logs were made available publicly, according to the source post.
A Repair, Not a New Math Ability
The result matters because it suggests a capable model may help diagnose and patch a specific weakness in an open-weight system, rather than only generate text or code. For teams that can inspect and modify model weights, an edit targeting a repeated failure could offer another way to adapt a model without the standard gradient-descent training process.
But the headline score should not be read as proof that Astra taught Qwen broad mathematical reasoning. The experiment addressed a known looping problem that interfered with answering questions, and the authors themselves caution that fixing loops is not the same as making a model better at math. The remaining 33.1% token-limit rate and the large difference between smaller and larger search budgets underline how dependent the result was on the particular repair and amount of evaluation.
This approach also has limits for deployment. An edited set of weights is a modified model that needs its own testing for accuracy, safety and reliability. The post cites earlier work showing that single-direction weight edits can remove refusal behavior from open chat models, a reminder that a targeted change can have consequences beyond the intended task. The current report does not establish how often those side effects occur or how the edit behaves across broader use.
Products Worth Considering
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Hazelnut Taupe
Amazon Kindle Scribe (64GB) with Premium Pen, Metallic Jade
Bose S1 Pro+ All-in-One PA Speaker
As an affiliate, we earn on qualifying purchases.
How the Weight Search Worked
The report, titled “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights,” describes eight experiments, including attempts to infer information about models and repair a failing driving policy. The Qwen math experiment was the one that drew widespread attention. The authors call the work “very early results” and say its capabilities remain limited. It is a blog post, not a peer-reviewed paper, and the source says it was not posted on arXiv.
In the Qwen test, Astra was given a Python environment to inspect and alter the model’s weights, but it could not run Qwen directly. It relied on paid-for forward passes to observe outputs. The researchers also supplied practice questions from GSM8K and MATH, while keeping the final benchmark questions hidden. The base Qwen model card warns that the model is prone to thinking loops, according to the source report.
The article identifies Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab as the authors, all listed with MIT CSAIL; Takehi and Teoh are marked as equal contributors. OpenAI is not listed as an author, and the source says it found no OpenAI statement about the experiment. GPT-6 Astra’s role in the study is therefore reported through the researchers’ account and logs, not an OpenAI announcement.
“Fixing loops, however, is not the same as making a model better at math.”
— The researchers, in their blog post
Questions About Reliability and Scope
The findings have not been independently validated in a peer-reviewed paper, and the reported numbers come from the authors’ blog post and published logs. It remains unclear whether other researchers can reproduce the same edit and results, or whether the gain holds across different prompts, sampling settings and broader math tasks. The source reports the benchmark figures but does not provide enough detail here to establish the model’s performance across real-world deployments.
The mechanism also needs further scrutiny. The researchers associate the looping with a direction linked to layers 15 and 19, but the report does not establish that this direction is the sole cause of repetition or explain all effects of removing it. The edited model still hit the token cap in about one-third of MATH-500 answers. It is also not yet clear what unintended changes the edit may cause to other capabilities or safety behavior.
Finally, “without training” is a limited description. The method avoided gradient descent, but it still used 50 forward passes, code runs and iterative test feedback over more than 20 hours. The post’s results on a second small model, Gemma 3 270M, reportedly showed no edit improving its score by more than about a point, so the Qwen result may not generalize to other models.
Replication and Broader Model Tests
The immediate next step is independent replication of the Qwen edit, including verification of the released weights, evaluation setup and benchmark results. Researchers would also need to test whether the reported improvement persists on additional math datasets and whether the reduction in loops comes with regressions elsewhere.
The authors’ published work covers multiple experiments, but the supplied source does not announce a timetable for a peer-reviewed paper or further releases. Until more evidence is available, the result is best treated as a promising but preliminary repair demonstration. Open-weight model developers considering similar edits would need to assess each modified model separately before relying on it.
Key Questions
What did GPT-6 Astra change in Qwen3.5-0.8B?
According to the researchers, Astra identified a direction associated with repetitive outputs and subtracted it from output columns in layers 15 and 19. The edit changed 131,072 weights, about 0.016% of the model’s total weights.
Did the edit eliminate Qwen’s looping?
No. The reported share of MATH-500 answers that reached the 32,768-token limit fell from 93.0% to 33.1% in the highest-budget run, so the problem was reduced, not eliminated.
What does “without training” mean in this report?
The researchers say they did not use gradient descent, backward passes or loss-based weight fitting. Astra still used iterative searches, Python runs and 50 forward passes in the strongest reported run.
Has the result been peer reviewed?
No peer-reviewed publication is identified in the source material. The result appeared in a blog post on November 2, 2026, which the authors describe as very early work.
Did OpenAI endorse or co-author the experiment?
The source identifies four researchers listed with MIT CSAIL as the authors and says OpenAI is not listed as a co-author. It also reports that no statement from OpenAI about the experiment was found.
Source: ThorstenMeyerAI.com
