🔍 Read the full analysis: GPT-6 Astra Edited AI Weights To Fix A Loop, Not Boost IQ 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 GPT-6 Astra edited two layers of Qwen3.5-0.8B to reduce repetitive output. The model’s reported MATH-500 score rose from 7.9% to 48.1%, 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 loops were reduced rather than eliminated.
Four researchers affiliated with MIT CSAIL report that OpenAI’s GPT-6 Astra edited two layers of the open-weight model Qwen3.5-0.8B to reduce repetitive reasoning loops, raising its reported score on the MATH-500 benchmark from 7.9% to 48.1%. The change was made without gradient descent, but it did not eliminate the problem: 33.1% of edited-model answers still reached the 32,768-token limit, according to the researchers’ November 2 blog post.
The post, titled “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights,” describes the result as an early experiment. Its authors are Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, listed with MIT CSAIL; Takehi and Teoh are identified as equal contributors. The post is not a peer-reviewed paper and, according to the source material, was not posted on arXiv.
The researchers say Astra identified a direction associated with looping in the model’s intermediate representations and subtracted it from output weights in layers 15 and 19. The edit affected 131,072 weights, or 0.016% of the model’s 0.8 billion weights. The published Hugging Face version, as described in the source, differs from the original in those two layers’ down-projection weight tensors.
The highest-budget run used 50 forward passes and took 20.3 hours. The authors’ logs also record 158 turns and 80 Python runs. Astra used practice problems that were separate from the final test sets; the researchers say the benchmark questions were not shown to it. They prohibited gradient descent and backward passes, while giving Astra a Python environment to inspect and edit the weights.
A Repair, Not a New Math Ability
The result matters because it suggests that a model may help diagnose and modify another model’s behavior when the latter’s weights are accessible. For teams running open-weight models, targeted edits could offer a way to address specific failure modes without a conventional training run. That possibility is different from a general-purpose system acquiring new reasoning skills: in this case, the main change was reducing a known tendency to repeat output until a token limit was reached.
The remaining failure rate is central to interpreting the score. Even after the edit, 33.1% of MATH-500 answers hit the limit, compared with 93.0% before. The headline score increase is a reported benchmark result, not proof that the model can reliably solve a broader range of math problems. The source also notes that smaller editing budgets produced much smaller gains, while a model without the looping issue did not show a comparable improvement in the reported tests.
There is also a deployment trade-off: editing the weights creates a modified model whose behavior must be checked separately. Any organization adopting such a patch would need to test whether it improves the targeted behavior without damaging performance elsewhere. The blog’s experiment is evidence of a possible technique, not an independent validation of its reliability or safety.
Products Worth Considering
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-inch with M5 Max Chip, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
Apple MacBook Pro 14 2021 with M1 Max, 64GB RAM, 2TB SSD, Silver
As an affiliate, we earn on qualifying purchases.
How the Qwen Experiment Was Run
The experiment was the eighth of eight demonstrations in the researchers’ post, which also describes other attempts to inspect or change neural-network weights. The source reports one unsuccessful test: Astra could not determine from the weights alone which of two fine-tuned Gemma models had been trained to flatter users. The authors characterize the overall work as “very early results” and say the capabilities remain limited.
Qwen3.5-0.8B is a small model, and the source says its model card warns that it is prone to thinking loops. On the researchers’ practice questions, 85 of the first 96 answers from the base model reached the 32,768-token cap. The team then used examples of looping and non-looping outputs to guide Astra’s search for a weight edit.
The reported MATH-500 scores increased with the editing budget: 12.3% after 10 passes, 14.1% after 25, and 48.1% after 50, compared with 7.9% for the base model. The share reaching the token limit was 87.2%, 86.4% and 33.1%, respectively, versus 93.0% for the base. The researchers also report that GSM8K scores rose from 13.8% to 67.1% in the 50-pass run, and IFEval scores from 21.0% to 62.3%. These are figures reported in the blog post, not independently reviewed findings.
““Fixing loops, however, is not the same as making a model better at math.””
— The researchers, in “Sparks of Weight-Space Programming”
Limits of the Reported Results
The figures come from the researchers’ blog post and logs; the work has not been peer-reviewed, according to the source material. It is not yet clear whether independent teams can reproduce the edit and results, or whether the same method will work on other model sizes and architectures.
The edited model continued to reach the token cap in roughly one-third of MATH-500 answers. The source material does not establish how the weight change affects performance across a wider range of tasks, nor whether it introduces other unwanted behaviors. It also reports no statement from OpenAI about the experiment; the researchers used GPT-6 Astra through the Codex command-line tool, but OpenAI is not listed as a co-author.
Reproduction and Broader Testing
The next evidence to watch for is independent replication, broader evaluation of the edited model and publication of further technical analysis. Researchers and model users would need to check whether the lower loop rate holds across different prompts and settings, and whether the gains persist outside the benchmarks reported in the post.
The authors’ post also frames weight editing as a developing capability, with several experiments showing mixed results. No follow-up timetable is given in the supplied material. For now, the report supports a narrow conclusion: under the researchers’ setup, Astra helped identify and remove a weight direction linked to looping in one small model, with improved benchmark scores but substantial residual failures.
Key Questions
What did GPT-6 Astra change?
According to the researchers, Astra subtracted a direction associated with repetitive output from weights in layers 15 and 19 of Qwen3.5-0.8B. The edit affected 131,072 weights, or 0.016% of the model’s total weights.
Did the researchers train the model?
They say they did not use gradient descent, backward passes or weight fitting to a loss. Astra did, however, conduct an iterative search using test feedback, Python runs and model forward passes. In this report, “without training” means without gradient descent, not without substantial experimentation.
Were the loops eliminated?
No. The reported share of MATH-500 answers reaching the 32,768-token limit fell from 93.0% to 33.1% in the 50-pass run. The researchers describe the change as a reduction in looping, not its complete removal.
Has the result been independently verified?
The supplied source describes the findings as a blog-post report and says they have not been peer-reviewed. It does not provide independent replication results.
Source: ThorstenMeyerAI.com
