🔍 Read the full analysis: How GPT-6 Astra Edited Another AI Model’s Weights To Fix A Loop on ThorstenMeyerAI.com
Get movie nights 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, 2026 blog post by four researchers listed with MIT CSAIL reports that GPT-6 Astra edited two layers of Qwen3.5-0.8B to reduce repetitive reasoning loops. The model’s reported MATH-500 score rose from 7.9% to 48.1%, but 33.1% of answers still reached the token limit, and the work is early, non-peer-reviewed research.
Four researchers listed with MIT CSAIL report that OpenAI’s GPT-6 Astra edited the weights of the open model Qwen3.5-0.8B to reduce repetitive reasoning loops, raising its reported MATH-500 score from 7.9% to 48.1%. The November 2 blog post describes an iterative, tool-assisted edit rather than conventional training; it is not peer-reviewed, and the edited model still hit its token limit in 33.1% of benchmark answers.
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 experiment logs. The authors call the findings “very early results”. The post is not on arXiv and has not undergone peer review, according to the source report.
In the math experiment, Astra examined the model’s behavior and edited two layers, numbered 15 and 19. The researchers say it identified a direction associated with looping and subtracted it from 64 output columns in each layer, changing 131,072 weights—reported as 0.016% of the model’s 0.8 billion weights. The edited Hugging Face model reportedly differs from the original only in the two layers’ down-projection weight tensors.
The reported MATH-500 result came after the largest tested budget: 50 forward passes. At 10 and 25 passes, scores were 12.3% and 14.1%, respectively. The share of answers reaching the 32,768-token limit fell from 93.0% for the base model to 87.2%, 86.4% and 33.1% across the 10-, 25- and 50-pass runs. The loop was reduced, not eliminated.
A Different Route to Model Repair
The experiment points to a possible way to alter an open model’s behavior by editing weights directly, rather than updating them through gradient descent. If the approach proves repeatable, developers may be able to address specific failures in models they can inspect and modify. That distinction matters: the method depends on access to the model’s weights, so it does not describe a general way to patch closed systems.
The result also carries a practical warning. The score gain required a substantial search budget and did not remove the underlying failure in every response. According to the source report, the full run used 158 turns, 80 Python runs and 50 forward passes over 20.3 hours. Those costs, alongside remaining loops, make this an early repair technique rather than evidence that a small model has acquired broadly reliable mathematical ability.
Weight edits can have consequences beyond the intended task. The researchers’ post references earlier work showing that a single-direction edit can remove refusal behavior from open chat models. That makes verification important: a targeted change may alter other behaviors, and any edited weights should be tested as a distinct model before deployment.
Popular right now
Ninja Luxe Café Mini Plus Espresso and Drip Coffee Machine with AutoFroth, Stainless Steel (ES351)
Dyson Car+Boat Cordless Handheld Vacuum
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Hazelnut Taupe
As an affiliate, we earn on qualifying purchases.
How the Qwen Loop Was Tested
The base Qwen3.5-0.8B model showed a severe repetition problem in the researchers’ test: 85 of its first 96 practice answers ran to the 32,768-token limit. The model card itself warns of a tendency toward thinking loops, according to the source article. The researchers framed the task as repairing that weakness, not giving the model a new mathematical capability.
Astra operated through the Codex command-line tool with a Python environment provided by the researchers. The setup was CPU-only, offline, and capped at 300 code runs. Astra could inspect and modify the weights but could not run the model without spending a forward pass. Each pass allowed up to 512 problem answers or 1,024 texts to be read. Gradient descent, backward passes and fitting weights to a loss were prohibited.
The researchers gave Astra practice material from GSM8K and MATH, while keeping the final benchmark questions unseen. In addition to MATH-500, the source reports that the 50-pass edit raised GSM8K accuracy from 13.8% to 67.1% and IFEval from 21.0% to 62.3%. These are results from the reported experiment, not independent validation. The post also describes other tests, including one in which Astra could not identify 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
Limits of the Reported Results
The findings have not been peer-reviewed, and the source material does not report independent replication. It is also not clear from the available account how consistently the edit would work on other prompts, model versions or evaluation setups, or whether the change affects capabilities beyond those measured.
The benchmark figures are reported by the researchers, and the highest score depended on the largest tested budget. The edited model still reached the token limit in 33.1% of MATH-500 answers. The source article also reports that a model without loops did not improve in the same way, but the supplied material does not provide enough detail to establish how broadly that comparison applies.
OpenAI is not listed as an author, and the blog post does not thank the company. The source report says it found no statement from OpenAI about the experiment. The model and method’s longer-term reliability, costs and possible side effects remain unsettled.
Replication and Broader Testing
The next useful step is independent testing of the published edited model against the original, using the same sampling settings and benchmarks and reporting variation across repeated runs. Researchers will also need to check whether the reduction in loops holds on new tasks and whether the weight changes affect unrelated behaviors.
The source material does not announce a scheduled follow-up or a peer-reviewed publication. For now, readers can examine the researchers’ blog, published logs and Hugging Face model files; whether other teams reproduce the gains or the authors expand the study remains unclear.
Key Questions
What did GPT-6 Astra change?
According to the researchers, Astra removed a direction associated with looping from output columns in layers 15 and 19 of Qwen3.5-0.8B. The edit changed 131,072 weights, reported as 0.016% of the model.
Did the method use no training at all?
The researchers prohibited gradient descent, backward passes and fitting weights to a loss. But Astra still used an iterative search, Python runs and forward-pass feedback. “Without training” here means without gradient descent, not without computation or effort.
Did the edit eliminate the model’s loops?
No. The share of MATH-500 answers reaching the 32,768-token limit fell from 93.0% to 33.1% in the 50-pass run. The reported results show a reduction, not elimination.
Has the result been independently verified?
The source report describes a blog post with logs and an edited model available on Hugging Face, but says the work is not peer-reviewed. The supplied material does not establish independent replication.
Source: ThorstenMeyerAI.com
