🔍 Read the full analysis: AI Model Repair: GPT-6 Astra Edited Another Model’s Weights on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 2 blog post by four researchers listed with MIT CSAIL reports that GPT-6 Astra modified two weight tensors in the open Qwen3.5-0.8B model without gradient descent. The edited model scored 48.1% on MATH-500, up from 7.9%, but 33.1% of answers still reached the token limit; the result is an early, non-peer-reviewed report.
Four researchers listed with MIT CSAIL report that OpenAI’s GPT-6 Astra edited two internal weight tensors in the open model Qwen3.5-0.8B, raising its MATH-500 score from 7.9% to 48.1% without gradient descent. The November 2 blog post describes an iterative search using test feedback—not an automatic fix requiring no evaluation—and says the model’s repetitive outputs fell but did not disappear.
The experiment appears in a post titled Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights, by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. The authors describe the work as “very early results”. It is a blog post, not a peer-reviewed paper, and the source says it was not posted on arXiv. OpenAI is not listed as a co-author, and the post reports no statement from the company about the experiment.
According to the published logs, GPT-6 Astra was run through the Codex command-line tool at the highest reasoning setting. The researchers’ own harness gave it CPU access to the model weights and a Python kernel, with no internet connection and a limit of 300 code runs. Astra could request forward passes to inspect model outputs; the researchers ran versions with budgets of 10, 25 and 50 passes. Each pass could cover up to 512 problems or 1,024 texts. The final benchmark questions were withheld from Astra, while practice material came from GSM8K and MATH.
The researchers say Astra identified a direction associated with repetitive outputs in layers 15 and 19, then subtracted it from 64 output columns in each layer. The change affected 131,072 weights, or 0.016% of the model’s stated 0.8 billion weights. The post says the edited Hugging Face model differs from the original only in the two layers’ down-projection tensors, and publishes their hashes. On MATH-500, the 50-pass version scored 48.1%, while 33.1% of its answers still reached the 32,768-token limit. The base model scored 7.9%, with 93.0% reaching that limit.
A Weight Edit, Not a New Math Skill
The result offers an example of a language model helping to change another model’s parameters without using gradient descent. That is relevant to developers working with open-weight models: if a weakness can be isolated inside a model, a targeted edit may offer another way to alter its behavior. The report also makes clear that the process required substantial evaluation and repeated tool use, rather than a single prompt or effortless repair.
The score increase is substantial in the reported MATH-500 test, but the authors caution against treating it as a general gain in mathematical ability. They write: “Fixing loops, however, is not the same as making a model better at math.” The model’s baseline problem was that it often kept generating until it hit the output limit. Reducing that behavior can make more answers usable and improve a benchmark score without showing that the model acquired a broader reasoning capability. The remaining 33.1% limit-hit rate also means the defect was reduced, not removed.
The approach has practical limits. It depends on access to editable weights, so it does not directly apply to closed models whose parameters are unavailable. A modified checkpoint is also a distinct model version that would need its own testing for reliability, safety and performance on tasks beyond the reported benchmarks. The post’s results are a case study, not evidence that the method will work consistently across models or problems.
Products Worth Considering
Bose SoundLink Plus Portable Bluetooth Speaker, Wireless Outdoor Speaker, Up to 20 Hours Battery Life, Waterproof and Dustproof, Black
DJI Mic 3 Wireless Microphone Kit (2 Transmitters, 1 Receiver, Charging Case)
DJI Osmo Mobile 8 Gimbal with Mic Mini Transmitter, Infinity Black
As an affiliate, we earn on qualifying purchases.
How the Researchers Tested the Edit
The source describes eight experiments in the blog post, including tasks involving small networks and a car-driving policy. One reported attempt did not succeed: Astra could not determine from weights alone which of two fine-tuned Gemma models had been trained to flatter users. The Qwen math experiment was the eighth and is the one highlighted by the source as attracting online attention.
Qwen3.5-0.8B’s model card warns that the model is prone to thinking loops, according to the post. In the researchers’ initial practice sample, 85 of 96 answers reached the 32,768-token limit. The authors supplied practice problems from GSM8K and MATH, while withholding the benchmark questions used for final evaluation. That separation is intended to keep Astra from directly tuning against the final test items, although the report remains an early experiment rather than an independent validation.
The amount of tool use mattered. At 10 and 25 forward passes, the reported MATH-500 scores were 12.3% and 14.1%, respectively, compared with 48.1% at 50 passes. The post says the winning run took 158 turns, 80 Python runs and 20.3 hours. On GSM8K, the 50-pass version scored 67.1%, up from a base score of 13.8%; on IFEval, it scored 62.3%, compared with 21.0% for the base model. The source also reports that a model without loops did not improve in the same way, and that edits to Gemma 3 270M changed scores by no more than about one point.
““Fixing loops, however, is not the same as making a model better at math.””
— The researchers, in their blog post
How General the Repair Is
The post does not establish whether the same weight-editing process will work reliably on other models, other kinds of failure or independent test sets. The source provides logs and a model checkpoint, but reports no peer review or independent replication. It is also unclear how the edit affects other tasks, including areas not included in the listed evaluations.
The benchmark figures describe results under the researchers’ stated setup; the source does not provide evidence here about performance in routine deployment or across a wider range of prompts and sampling conditions. The edit left 33.1% of MATH-500 answers at the token cap, so the looping behavior remained present. The source also gives no OpenAI comment, and does not establish that OpenAI participated in or endorsed the research.
Replication and Broader Testing
The immediate next step is independent testing of the published checkpoint and logs, including checks of the reported scores and an evaluation of tasks beyond math. Researchers would also need to establish whether the same procedure can identify and repair other specific model failures, and whether the resulting edits preserve performance and safety elsewhere.
The source does not announce a journal submission, a formal follow-up paper or an OpenAI response. For now, readers can treat the result as a documented but preliminary demonstration: GPT-6 Astra was used to edit a small open model’s weights, with reported gains on selected benchmarks and a substantial residual loop rate. Any broader claims will depend on further validation.
Key Questions
What did GPT-6 Astra change?
According to the researchers’ logs, Astra identified a direction associated with looping and changed the down-projection weight tensors in layers 15 and 19 of Qwen3.5-0.8B. The post reports 131,072 changed weights.
Did the model’s looping stop?
No. The reported share of MATH-500 answers that reached the 32,768-token limit fell from 93.0% to 33.1% in the 50-pass version, so the behavior was reduced, not eliminated.
Was this done without any training or testing?
The researchers say they did not use gradient descent or fit the weights to a loss. But the process did use iterative testing and tool calls: the winning run reportedly took 158 turns and 80 Python runs, with 50 forward passes.
Has the result been peer-reviewed?
No peer review is reported in the source. The result was published as a blog post on November 2, 2026, and the authors characterize it as very early work.
Does this show that AI can generally improve other AI models?
It shows one reported edit to one small open model, with gains on selected evaluations. The source does not establish that the method generalizes across models or failures, and the authors warn that fixing loops is not the same as improving mathematical ability broadly.
Source: ThorstenMeyerAI.com
