🔍 Read the full analysis: GPT-6 Astra’s Model-Editing Move: Repairing A Loop, Not Raising IQ on ThorstenMeyerAI.com
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. The edited model scored 48.1% on MATH-500, up from 7.9%, while answers hitting the 32,768-token limit fell from 93.0% to 33.1%. The result is an early, non-peer-reviewed report, and the researchers describe the change as loop repair rather than a general increase in mathematical ability.
Four researchers listed with MIT CSAIL report that OpenAI’s GPT-6 Astra edited two layers of the open-weight Qwen3.5-0.8B model, raising its reported MATH-500 score from 7.9% to 48.1% without gradient descent. The change also reduced, but did not eliminate, answers that ran into the model’s 32,768-token limit: that share fell from 93.0% to 33.1%.
The result appears in “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights,” a blog post published November 2 by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. Takehi and Teoh are identified as equal contributors. The authors call the work “very early results”; the post is not peer-reviewed and was not found on arXiv in the source account.
According to the researchers’ published logs, Astra used a Python-based harness to inspect and edit the model’s weights. It identified a direction associated with repetitive outputs in the MLP components of layers 15 and 19, then subtracted that direction from 64 output columns in each layer. The edit affected 131,072 weights, or 0.016% of the model’s 0.8 billion weights. The edited model hosted on Hugging Face differs from the original in the two corresponding down-projection tensors, the source reports.
The strongest reported result came after 50 forward passes. The run used 80 Python executions and lasted 20.3 hours, according to the logs. Smaller budgets produced more modest MATH-500 scores: 12.3% after 10 passes and 14.1% after 25, compared with 7.9% for the base model. The researchers also report improvements on GSM8K and IFEval after the 50-pass run, while emphasizing that their main example is a targeted repair of a known failure mode.
A Targeted Fix for Open Models
The report offers a case study in using a capable model as an editor of another model’s open weights, rather than retraining it with gradient descent. If the method can be reproduced and applied reliably, developers may have another way to diagnose and alter specific behaviors in models they can inspect directly. The edit here was small relative to the full network, but required substantial testing and an iterative search.
For users, the central practical point is narrower than the headline score suggests. Qwen3.5-0.8B’s model card already warns that it is prone to thinking loops, according to the source account. The reported score increase accompanies a sharp drop in looping, so it does not establish that Astra independently gave the model a broad new mathematical capability. The authors themselves caution that “Fixing loops … is not the same as making a model better at math.”
There are also operational implications for organizations using open-weight systems: an edited checkpoint is a changed model, not merely a prompt adjustment. Teams would need to evaluate the patched version for regressions, safety effects and performance across their own tasks. The reported figures come from particular benchmarks and settings; they are not a guarantee of results in other deployments.
Popular right now
Ninja Flip Toaster Oven & Air Fryer, Large Capacity, Stainless, SP151
De'Longhi Magnifica Duo Fully Automatic Espresso Machine with Milk Frother, Silver
Dreame H12 Pro Wet and Dry Vacuum Mop
As an affiliate, we earn on qualifying purchases.
How the Weight Search Worked
Astra did not freely run the model or access the internet in the experiment. The published setup describes a CPU-only harness with a Python kernel for working with the 24-layer model’s weights. The team supplied practice material from GSM8K and MATH, but withheld the benchmarks used for the final reported scores. A forward pass let Astra observe responses to up to 512 problems or read up to 1,024 texts; the researchers tested budgets of 10, 25 and 50 such passes.
The phrase “without training” needs qualification. The researchers say they barred gradient descent, backward passes and fitting weights to a loss. Astra still made repeated edits and used feedback from test runs to guide its search. The winning run, as reported, involved 158 turns, 80 Python runs and 50 forward passes over 20.3 hours. The source also says that a model without the looping problem did not improve in the same way, and that tests on Gemma 3 270M produced changes of no more than about a point.
The blog covers eight experiments, not only the Qwen math result. One test, asking Astra to identify which of two fine-tuned Gemma models had been trained to flatter users, failed: the researchers say it could not distinguish them from the weights alone. The authors also connect their method to earlier work on removing behaviors through changes to a model’s internal directions, including a 2024 paper by Arditi and colleagues on refusal behavior.
„“Fixing loops, however, is not the same as making a model better at math.”“
— The researchers, in “Sparks of Weight-Space Programming”
Reproducibility and Remaining Loops
The results are reported in a blog post, not a peer-reviewed paper, and the supplied source says the work was not posted on arXiv. Independent replication and external evaluation are not described in the material provided. It is also unclear how consistently the edit would help on other model versions, different prompts, or math tasks beyond the reported benchmarks.
The repair was incomplete: 33.1% of edited-model MATH-500 answers still reached the 32,768-token limit, based on the reported run. The figures also do not establish whether the change affected other capabilities or introduced failures outside the tested tasks. OpenAI is not listed as an author, and the source reports no OpenAI statement about the experiment; the nature of any organizational involvement beyond use of GPT-6 Astra is not established.
Replication and Broader Testing
The immediate next step is independent testing of the published edited checkpoint against the original model, using the same evaluation settings and additional tasks. Further work would need to establish whether the improvement holds across repeated runs, whether remaining loops can be reduced without harming other behavior, and how much expert oversight the search requires.
The researchers’ post describes several demonstrations and calls the work early, but the source material gives no announced date for a paper, a follow-up release or a broader evaluation. Until more evidence is available, the result is best read as a specific reported model repair—not proof that AI systems can routinely improve other models without trade-offs.
Key Questions
What did GPT-6 Astra change in Qwen3.5-0.8B?
According to the researchers’ logs, Astra removed a direction linked to repetitive outputs from output columns in the MLP components of layers 15 and 19. The edit changed 131,072 weights, the source says.
Did the edit eliminate the model’s loops?
No. The reported share of MATH-500 answers hitting the 32,768-token limit fell from 93.0% to 33.1%, so a substantial share still reached the limit.
Was the model retrained?
The researchers say they did not use gradient descent, backward passes or loss-based weight fitting. Astra nevertheless used iterative edits and test feedback, so “without training” does not mean without computation or experimentation.
How strong is the evidence so far?
The figures are from a November 2, 2026 blog post with published logs and an edited model checkpoint, as described in the source. The authors call the results very early; the report is not peer-reviewed, and independent replication is not established in the material provided.
Did Astra make Qwen generally better at math?
The reported MATH-500 score rose from 7.9% to 48.1%, but the researchers attribute the change to reducing a looping failure and caution that fixing loops is not the same as broadly improving mathematical ability. The evidence covers specific benchmarks and settings.
Source: ThorstenMeyerAI.com
