🔍 Read the full analysis: What Did GPT-6 Astra’s Edit Change In Another AI Model? on ThorstenMeyerAI.com
Get the latest gadgets 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 research blog reports that GPT-6 Astra used a search process to edit two layers in the open-weight Qwen3.5-0.8B model, raising its MATH-500 score from 7.9% to 48.1%. The result is not peer-reviewed, required 50 forward passes and 20.3 hours, and left 33.1% of test answers hitting the model’s token limit.
Researchers report that OpenAI’s GPT-6 Astra raised the open-weight Qwen3.5-0.8B model’s score on MATH-500 from 7.9% to 48.1% by editing weights in two layers, without using gradient descent. The November 2 blog post describes a targeted attempt to reduce the model’s repetitive reasoning loops; after the edit, 33.1% of MATH-500 answers still reached the 32,768-token limit, so the behavior was reduced rather than removed.
The work appears in a post titled “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights,” published by Rikiya Takehi on November 2. Four authors are listed with MIT CSAIL: Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab; Takehi and Teoh are identified as equal contributors. The post is not a peer-reviewed paper, and the authors call the findings “very early results.”
According to the researchers’ logs, Astra was given access to the model’s weights through a Python environment and could request forward passes to observe the model’s outputs. The team barred gradient descent, backward passes and fitting weights to a loss. Astra used a reported 50 forward passes, 80 Python runs and 158 turns in a process that took 20.3 hours. The final benchmarks were not shown to Astra during the search, the authors say.
The reported edit altered the down-projection weight tensors in layers 15 and 19 of the 24-layer model. The post says the change affected 131,072 weights, or 0.016% of the model’s 0.8 billion weights. In the authors’ account, Astra identified an internal direction associated with looping and subtracted it from output columns in those layers. The edited model and logs are available on Hugging Face and in the published material, according to the source.
The results depended heavily on the search budget. On MATH-500, runs with 10 and 25 forward passes reached 12.3% and 14.1%, respectively, compared with 48.1% at 50 passes. The share of responses hitting the token limit fell from 93.0% in the base model to 87.2%, 86.4% and 33.1% across those runs. The researchers also report that GSM8K rose from 13.8% to 67.1% and IFEval from 21.0% to 62.3% in the 50-pass run.
A Targeted Repair, Not a New Math Skill
The result matters because it offers a reported example of a model making a small, direct change to another model’s weights, rather than merely suggesting new training data or code. If reproducible, this kind of method could give teams working with open-weight systems another way to diagnose and patch specific failure patterns.
But the distinction between repair and capability gain is central. The researchers say the intervention addressed a known weakness: Qwen’s model card warns that it can fall into thinking loops. They also write, “Fixing loops, however, is not the same as making a model better at math.” The jump in benchmark accuracy may largely reflect the model completing answers it previously failed to finish, not a general increase in mathematical ability.
The reported figures also do not establish that this approach works broadly. In the source account, tests on Gemma 3 270M produced no edit that changed scores by more than about a point, and a model without loops did not improve through the same process. The Qwen result is therefore a specific demonstration, not evidence that one model can reliably upgrade any other model.
Products Worth Considering
Apple Watch Series 8 GPS, 41mm, Midnight Aluminum Case with Midnight Sport Band (S/M)
Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W
Apple Watch SE (2nd Gen) GPS 40mm with Starlight Aluminum Case and Sport Band (S/M)
As an affiliate, we earn on qualifying purchases.
How the Qwen Loop Was Measured
The experiment focused on Qwen3.5-0.8B, a relatively small open-weight model, and its behavior on math problems. Before the edit, the researchers report that 85 of the first 96 practice answers ran to the model’s maximum output length of 32,768 tokens. On the final MATH-500 evaluation, 93.0% of base-model responses reached that limit, while its reported accuracy was 7.9%.
The team used practice questions from GSM8K and MATH, while withholding the final benchmark sets from Astra, according to the post. That setup was intended to let Astra inspect behavior and propose an edit without directly optimizing against the final test questions. The authors say the harness ran on a CPU, had no internet access and limited code execution to 300 runs.
The post describes eight experiments, not all of which succeeded. In one, Astra could not determine from weights alone which of two fine-tuned Gemma models had been trained to flatter users. The Qwen math repair was the experiment that produced the headline result. OpenAI’s developer documentation describes GPT-6 Astra as its most capable model for demanding work, and CNBC reported a rollout on September 3, 2026. The source says OpenAI was not an author of the research post and had made no statement about this experiment.
““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 reported results come from a blog post rather than a peer-reviewed paper, and the source characterizes them as preliminary. The post provides logs and the edited model, but independent replication and outside evaluation are not established by the material provided. It is also not clear whether other researchers or model developers have reproduced the same edit and scores.
The token-limit figures show that the looping behavior remained common even after the best reported run: one-third of MATH-500 responses still hit the maximum length. The source does not establish how well the edit performs across other math tests, prompts, sampling settings or deployment conditions. Nor does the single successful example establish that Astra can consistently diagnose and improve different models. The source reports no OpenAI comment on the experiment.
Independent Tests Will Set the Scope
The next useful step is independent replication of the reported edit and benchmark results, including checks that the scores hold under different evaluation settings and that the reduced looping does not come with other performance costs. The published Hugging Face model and logs give other researchers material to inspect, but the source does not announce a scheduled peer-reviewed paper or a formal validation effort.
For now, the finding is best treated as a promising, narrow repair demonstration. Whether weight-space editing can be repeated reliably, scaled to other models or applied without unwanted side effects remains open. Further results from the authors or independent teams would clarify whether this approach is a practical tool for model maintenance or a one-off success on a model with a particularly visible failure mode.
Key Questions
What did GPT-6 Astra change in Qwen3.5-0.8B?
According to the researchers, Astra identified a direction associated with repetitive looping and subtracted it from output columns in the model’s down-projection weights in layers 15 and 19. The edit affected 131,072 weights, which the post says is 0.016% of the model’s total weights.
Did the edit eliminate Qwen’s reasoning loops?
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 remaining rate means loops were reduced, not eliminated.
Did GPT-6 Astra train the model?
The researchers say they did not use gradient descent, backward passes or loss-based weight fitting. Astra nevertheless carried out an iterative search, using test feedback and many forward passes, so “without training” does not mean the result required no computation or experimentation.
Is the 48.1% MATH-500 score independently verified?
The source reports the score in a research blog post with logs and an edited model, but the post is not peer-reviewed. The material provided does not establish independent replication.
Does this show that AI can improve any other AI model?
No. The result concerns one Qwen model and one targeted looping problem. The source also reports that tests on Gemma 3 270M produced no edit that changed scores by more than about a point, and that a model without loops did not improve.
Source: ThorstenMeyerAI.com
