🔍 Read the full analysis: GPT-6 Astra Edited Another Model. Was It A Repair Or An Upgrade? 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 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 without gradient descent. Its MATH-500 score rose from 7.9% to 48.1%, while answers reaching the token limit fell from 93.0% to 33.1%; the results are early, not peer-reviewed, and depend on substantial iterative testing.
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 says the edit targeted repetitive reasoning loops without gradient descent; it is an early, non-peer-reviewed report, and the loops remained in a substantial share of answers.
The experiment addressed a specific failure: Qwen3.5-0.8B often continued generating repetitive reasoning until it reached the 32,768-token limit. Before the edit, 93.0% of its MATH-500 answers hit that limit. After the reported 50-pass run, the share fell to 33.1%, while the benchmark score rose to 48.1%. The result is a reported improvement, not evidence that looping was eliminated.
According to the researchers’ published logs, Astra identified an internal direction associated with looping and removed it from output weights in layers 15 and 19. The change affected 131,072 weights, or 0.016% of the model’s 0.8 billion weights. The authors say the edited model’s two relevant weight tensors differ from the original; the edited model and associated hashes are available on Hugging Face.
The phrase “without training” needs qualification. The authors prohibited gradient descent and fitting weights to a loss, but Astra carried out an iterative search using test feedback. The winning run involved 158 turns, 80 Python runs and 50 forward passes, with each pass able to assess up to 512 problems. The work took 20.3 hours, according to the post. At 10 and 25 passes, the MATH-500 scores were 12.3% and 14.1%, respectively.
A Targeted Fix, Not New Math Skill
The result matters because it suggests a language model may help inspect and edit the weights of another model, including a model that can be run outside a closed service. For organizations using open-weight systems, a targeted patch could offer another way to address a specific failure without retraining the model through gradient descent.
But the evidence points to a repair of a known weakness, not a general leap in mathematical ability. The source article says Qwen’s model card already warns of thinking loops, and the researchers caution that fixing loops is not the same as making a model better at math. Because the intervention changes weights, the patched model is also a distinct version that would need its own evaluation for accuracy and other behaviors.
The report also indicates that the method may rely heavily on the search budget: smaller runs produced much smaller score gains. The researchers report that a model without loops did not improve in the same way, while tests on Gemma 3 270M produced changes of about a point or less. Those limits make the Qwen result promising as a specific case, but not proof of a reliable, general-purpose editing method.
Popular right now
Breville BES870BTR Barista Express Espresso Machine, Black Truffle
Ninja Foodi Air Fryer, 8 Qt DualZone, 6-in-1, Grey, DZ201
Vitamix Propel Series 510 Professional-Grade Blender, 3 Programs, Black
As an affiliate, we earn on qualifying purchases.
How the Weight Edit Was Tested
The experiment is one of eight described in the blog post, Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights, published November 2 by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. The source identifies all four with MIT CSAIL and marks Takehi and Teoh as equal contributors. The post is not a peer-reviewed paper, and the authors describe the work as “very early results.”
According to the logs, Astra ran through the Codex command-line tool at its highest reasoning effort. The researchers’ own harness gave it a Python environment to inspect and edit the model’s weights, with no internet access and a limit of 300 code runs. Astra could use practice problems from GSM8K and MATH, but the final benchmark questions were withheld. Gradient descent and backward passes were not allowed.
The post reports gains beyond MATH-500: on GSM8K, the score rose from 13.8% to 67.1% after 50 passes; on IFEval, it rose from 21.0% to 62.3%. These are results reported by the authors, not independent replications. The article also says one of the post’s eight experiments failed: Astra could not tell from the weights alone which of two fine-tuned Gemma models had been trained to flatter users.
“Fixing loops, however, is not the same as making a model better at math.”
— The researchers, as quoted in their blog post
Replication and Generality Remain Open
The reported numbers have not been independently verified in the source material, and the blog post has not undergone peer review. It is not yet clear whether other researchers can reproduce the same edit and benchmark gains using the published setup, or whether comparable results would hold across model sizes, tasks and evaluation conditions.
The source says the edit reduced the rate of answers hitting the token limit, but 33.1% still reached it after the 50-pass run. It also does not establish whether the gains persist under different sampling settings or whether the weight change affects performance on tasks beyond those reported. OpenAI is not listed as an author, and the source reports no statement from the company about Astra’s role in the experiment.
Independent Checks Will Test the Claim
The immediate next step is independent testing of the published edited model against the original, using the same benchmarks and settings and checking for regressions on other tasks. The researchers have shared logs and model weights, giving other teams material to examine, but the source provides no announced date for a peer-reviewed paper or a follow-up study.
Until those checks are available, the result is best treated as a documented early experiment: Astra reportedly made a narrow change that substantially improved selected benchmark scores while reducing, not eliminating, looping. Whether this approach proves useful beyond this model and failure mode remains open.
Key Questions
What did GPT-6 Astra change?
According to the researchers’ report, Astra removed an internal direction associated with repetitive output from the model’s weights in two layers. The edit affected 0.016% of Qwen3.5-0.8B’s weights.
Did the edited model score 48.1% on every math test?
No. The 48.1% score was reported for the MATH-500 benchmark after the 50-pass run. The authors reported a separate 67.1% result on GSM8K for that run.
Were the model weights changed without any training or testing?
No gradient descent or loss-based weight fitting was used, according to the post. Astra still ran an iterative search with substantial testing feedback: the winning run used 50 forward passes, 80 Python runs and 158 turns.
Did the edit stop the looping completely?
No. The share of MATH-500 answers that reached the 32,768-token limit fell from 93.0% to 33.1%, so the reported problem was reduced but remained.
Has the result been peer-reviewed?
The source describes it as a blog post published November 2, 2026, and says it is not peer-reviewed. Independent replication is not reported in the material provided.
Source: ThorstenMeyerAI.com
