🔍 Read the full analysis: A Loop Fix, Not An IQ Jump: GPT-6 Astra Edited Another Model on ThorstenMeyerAI.com
Get hardware and tech essentials 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 OpenAI’s GPT-6 Astra edited two weight tensors in Qwen3.5-0.8B without gradient descent. On MATH-500, the edited model’s score rose from 7.9% to 48.1%, while the share of answers reaching the 32,768-token limit fell from 93.0% to 33.1%. The results are early, not peer-reviewed, and the method required extensive iterative testing; the reported change appears to reduce a looping failure rather than establish a general increase in mathematical ability.
Four researchers affiliated with MIT CSAIL report that OpenAI’s GPT-6 Astra edited weights in the open model Qwen3.5-0.8B, raising its MATH-500 score from 7.9% to 48.1% without gradient descent. The change targeted a repetitive-answer failure: the share of answers that hit the benchmark’s 32,768-token limit fell from 93.0% to 33.1%, but did not disappear. The findings appeared in a blog post on November 2 and have not been peer-reviewed.
The post, titled “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights,” describes eight experiments. Its authors are Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab; Takehi and Teoh are identified as equal contributors. They characterize the work as “very early results”. The post is not a peer-reviewed paper, and the authors say the capabilities remain limited.
For the Qwen experiment, the researchers gave Astra access to the model’s weights through a custom Python setup. According to their published logs, Astra was run as gpt-6-astra through the Codex command-line tool at its highest reasoning effort. The setup was CPU-only and had no internet access; built-in Codex tools were disabled. Researchers supplied practice problems and limited the number of model forward passes Astra could request. The final benchmarks were not shown to Astra.
The reported edit changed the down_proj weights in layers 15 and 19 of Qwen3.5-0.8B. The authors say it removed a direction associated with looping from 64 output columns in each layer, changing 131,072 weights—about 0.016% of the model’s 0.8 billion weights. The Hugging Face version, as described in the source report, differs from the original in those two tensors.
A targeted fix with wider implications
The result matters because it suggests that an AI model may be able to help inspect and alter the internal weights of another model, rather than only produce text or code around it. If the approach can be reproduced and tested more broadly, it could offer developers a way to diagnose specific behaviors in open-weight models. The claim is narrower than the viral framing of an AI system teaching another model a new skill: in this experiment, Astra appears to have addressed an existing failure mode that prevented the small model from completing answers.
The performance gains also came with substantial testing. The successful run used 50 forward passes, 80 Python runs and 158 turns over 20.3 hours, according to the source report. A forward pass let Astra inspect the model’s responses across batches of problems. With budgets of 10 and 25 passes, the reported MATH-500 scores were only 12.3% and 14.1%. The difference indicates that the result depended on iterative search and feedback, not a one-step edit.
Weight editing also has safety implications. The source post points to earlier research showing that a single-direction edit can remove refusal behavior from open chat models. That does not establish that Astra’s technique will work reliably across systems, but it highlights why modifications should be evaluated for both intended and unintended effects. A changed weight file is a distinct model version, and its behavior requires separate testing.
Products Worth Considering
Apple MacBook Pro with M5 Max 18-core CPU 20-core GPU 14.2-inch Display 48GB Memory 2TB SSD Space Black
Apple MacBook Pro 16.2-inch with M5 Pro, 18-Core CPU, 20-Core GPU, 48GB Memory, 1TB SSD, Wi-Fi 7, Silver
Apple MacBook Pro 14.2-inch with M5 Max, 18-core CPU, 32-core GPU, 36GB Memory, 2TB SSD, Wi-Fi 7, Silver
As an affiliate, we earn on qualifying purchases.
How the Qwen looping problem was tested
Qwen3.5-0.8B is a small open model, and its model card warns that it can get caught in thinking loops, according to the source report. In the researchers’ initial practice sample, 85 of 96 answers reached the 32,768-token limit. The authors compared looping and non-looping outputs, then used the model’s weights to identify a direction they associated with repetition.
The reported MATH-500 figures were 7.9% for the base model and 48.1% after the 50-pass edit. On the same benchmark, answers reaching the token cap declined from 93.0% to 33.1%. On GSM8K, the reported 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 in the blog post, not independently verified findings. The source also says smaller edits did not produce comparable gains on the other tested model, Gemma 3 270M.
The researchers barred gradient descent, backward passes and fitting weights to a loss function. In that limited sense, the change was made “without training”; it was not made without evaluation or computational work. Astra repeatedly examined outputs and modified the weights through code. The distinction matters when describing what the experiment demonstrates.
““Fixing loops, however, is not the same as making a model better at math.””
— The researchers, in “Sparks of Weight-Space Programming”
Reproducibility and side effects remain open
The reported benchmark improvements have not been established through peer review or independent replication. The source material identifies the post and published logs but does not provide a third-party evaluation of the edited model. It is also not clear whether the same approach would work on larger models, other model families, or different reasoning failures.
The edit reduced, but did not eliminate, answers reaching the token limit: 33.1% still hit the cap after the successful run. The available account also does not establish whether gains hold across broader test sets or whether the weight change affects other behaviors outside the reported benchmarks. The source says the benchmarks were withheld from Astra, but does not by itself settle questions about test-set contamination or independent validation.
OpenAI is not listed as an author of the post, and the source report says it found no statement from the company about the experiment. The work therefore should not be presented as an official OpenAI project or endorsement. The full practical cost, repeatability and safeguards for applying such edits also remain to be assessed.
Independent tests will determine reach
The immediate next step is for other researchers to reproduce the edit and evaluate the resulting model on the same benchmarks and additional tests. That work could establish whether the reported gains persist, whether looping falls under different prompting and sampling conditions, and whether other capabilities change as a side effect.
Further experiments would also need to clarify how much effort the method requires across models. In this report, the 50-pass run substantially outperformed the 10- and 25-pass runs, while the source says edits to Gemma 3 270M barely changed scores. Until results are replicated and the altered models are tested beyond the reported tasks, the finding remains a specific preliminary repair, not evidence of a general method for making models more capable.
Key Questions
What did GPT-6 Astra change?
According to the researchers’ blog post, Astra changed the down_proj weight tensors in layers 15 and 19 of Qwen3.5-0.8B. The authors linked the removed direction to repetitive, looping answers.
Did the edited model stop looping?
No. The reported share of MATH-500 answers reaching the 32,768-token limit fell from 93.0% to 33.1%. That is a reduction, not elimination.
Was this done without training?
The researchers did not use gradient descent, backward passes or loss-based weight fitting. But Astra used iterative tests and code to inspect and edit the weights, including 50 forward passes in the run with the largest reported MATH-500 gain.
Are the results independently verified?
The source describes a blog post published November 2, 2026, with logs and a model on Hugging Face. It says the work has not been peer-reviewed; independent replication is not established in the supplied material.
Does this prove AI can generally improve other AI models?
No. The reported experiment concerns one small model and a targeted looping problem. The authors call the results very early, and the account does not show that the method works broadly across models or reliably improves general capability.
Source: ThorstenMeyerAI.com
