🔍 Read the full analysis: GPT-6 Astra Worked On Another AI’s Weights To Tackle A Loop 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 weight tensors in the open-weight Qwen3.5-0.8B model to address repetitive outputs. The reported MATH-500 score rose from 7.9% to 48.1%, but 33.1% of answers still reached the token limit, and the result has not been peer-reviewed.
Four researchers affiliated with MIT CSAIL report that OpenAI’s GPT-6 Astra made a small edit to the weights of open model Qwen3.5-0.8B, reducing repetitive answers and raising its MATH-500 score from 7.9% to 48.1%. The November 2 blog post describes an experimental method without gradient descent, but it is not peer-reviewed, and the edit did not eliminate looping: 33.1% of edited-model answers still reached the benchmark’s 32,768-token limit.
The post, titled Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights, was written 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” and say its capabilities are exciting but limited. It has not been posted on arXiv or peer-reviewed, according to the supplied reporting.
In the Qwen experiment, Astra used a researcher-built Python environment to inspect and edit the model’s weights. According to the published logs, it was run as gpt-6-astra through the Codex command-line tool at its highest reasoning effort. Built-in Codex tools were disabled; the environment ran on a CPU without internet access. Astra could use practice questions but was not shown the final benchmark sets. The researchers barred gradient descent, backward passes and fitting weights to a loss.
The reported edit targeted the down_proj tensors in layers 15 and 19. The authors say Astra identified a direction associated with looping and subtracted it from 64 output columns in each layer, changing 131,072 weights, or 0.016% of the model’s 0.8 billion parameters. The published model files reportedly show differences only in those two tensors. The reported MATH-500 score rose to 12.3% after 10 forward passes, 14.1% after 25, and 48.1% after 50; the corresponding shares of answers hitting the token limit were 87.2%, 86.4% and 33.1%, compared with 7.9% accuracy and 93.0% at the limit for the base model.
A Targeted Repair, Not New Training
The result matters because it suggests a language model may help diagnose and modify an open-weight model directly, rather than only generate code or instructions for a conventional training process. If replicated, this kind of targeted intervention could offer researchers and model operators another way to address specific failures, such as repetitive generation, in systems whose weights they can access.
But the reported gains should not be read as proof that Astra independently taught Qwen mathematics. The authors’ own framing, quoted in the supplied report, is that “Fixing loops, however, is not the same as making a model better at math.” A reduction in a failure mode can allow a model to finish more answers, lifting benchmark performance without establishing a new underlying mathematical capability. The remaining 33.1% rate of answers hitting the token limit also shows the issue was reduced, not removed.
The method also required repeated testing and meaningful compute time, despite the shorthand “without training.” The supplied report says the strongest run took 158 turns and 20.3 hours, with 80 Python runs and 50 forward passes. A forward pass could evaluate as many as 512 problems. The improvement was not immediate: runs with smaller budgets achieved much lower MATH-500 scores. This makes the experiment relevant as a research technique, not yet evidence of a low-cost, reliable repair process for deployed models.
Products Worth Considering
Logitech MX Keys S Wireless Keyboard, Graphite
Amazon Kindle Scribe (64GB) with Premium Pen, Metallic Jade
Apple Watch Series 12 GPS 46mm Smartwatch
As an affiliate, we earn on qualifying purchases.
How the Qwen Loop Was Tested
Qwen3.5-0.8B is a small open-weight model, and its model card warns that it is prone to thinking loops, according to the supplied source. In the researchers’ initial practice run, 85 of 96 answers reached the 32,768-token cap. The experiment focused on this specific weakness rather than attempting a broad upgrade across tasks.
The researchers gave Astra practice material from outside the final tests: 7,469 GSM8K questions and 7,275 MATH questions. The final evaluation included MATH-500, GSM8K and IFEval. The source reports that the 50-pass version scored 67.1% on GSM8K, up from 13.8%, and 62.3% on IFEval, up from 21.0%. Those results are reported by the blog authors, not independently validated in a peer-reviewed study.
The post describes eight experiments, including tasks involving other models and a driving policy. One attempt to identify which of two fine-tuned Gemma models had been trained to flatter a user failed: Astra could not tell from the weights alone. The source also reports limited results from other tests, including no improvement greater than about a point in the cited Gemma 3 270M runs. These examples place the Qwen result within an early set of demonstrations rather than a general capability claim.
““very early results””
— The authors of “Sparks of Weight-Space Programming”
Replication and Limits Remain Open
The figures come from a researchers’ blog post and its logs, not a peer-reviewed paper or independent replication. The supplied reporting says the edited model and weight hashes are available on Hugging Face, but that does not by itself establish that the evaluation will reproduce across other environments, sampling choices or model versions.
It is also not clear whether the approach will work reliably on larger models or on failures other than looping. The source describes mixed results in other experiments and says smaller forward-pass budgets produced modest gains. The reported comparison therefore applies to this model, setup and evaluation; it does not establish a general method for improving AI systems.
OpenAI is not listed as an author, and the post does not thank the company. The supplied source says it found no statement from OpenAI about the experiment. The relationship between the researchers and the company, beyond using GPT-6 Astra, is not established in the material provided.
Independent Tests Will Matter
The next meaningful test is whether other researchers can reproduce the weight edit and benchmark results using the published logs and model files. Independent evaluation could check whether the gains hold with consistent sampling, whether the lower loop rate persists on new prompts, and whether the edit causes regressions on tasks beyond those reported.
The authors characterize the work as early, and the supplied material gives no announced timetable for a peer-reviewed paper or follow-up release. Until further evaluation is available, the result is best understood as a reported demonstration of a targeted edit to one open-weight model, not a confirmed general-purpose way to improve model capability.
Key Questions
What did GPT-6 Astra change?
According to the researchers’ logs, Astra edited the down_proj weights in Qwen3.5-0.8B’s layers 15 and 19, subtracting a direction the authors associated with repetitive looping.
Did the edit stop Qwen from looping?
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 source describes a reduction, not elimination.
Was the model trained?
The authors say they did not use gradient descent, backward passes or loss-based fitting. However, Astra used repeated tests and forward passes to guide edits, so “without training” does not mean without experimentation or computational effort.
Are the results independently verified?
The supplied reporting describes the results as coming from a November 2 blog post and published logs. The post is not peer-reviewed, and independent replication is not established in the source material.
Does the MATH-500 score prove Qwen gained new math ability?
No. The authors caution that fixing loops is not the same as making a model better at math. The reported score reflects this model and test setup; it does not by itself establish a broad increase in mathematical ability.
Source: ThorstenMeyerAI.com
