🔍 Read the full analysis: Another Model’s Weights, Edited By GPT-6 Astra: What Changed? 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 Qwen3.5-0.8B model, raising its MATH-500 score from 7.9% to 48.1%. The edit used iterative testing without gradient descent, but the result is not peer-reviewed, and 33.1% of edited-model answers still hit the benchmark’s 32,768-token limit.
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, raising its score on the MATH-500 benchmark from 7.9% to 48.1%. The reported gain came from an iterative search that removed a direction associated with repetitive outputs, not from gradient descent; the edited model still reached its output limit on 33.1% of answers, 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 eight experiments and provides logs. Its authors are Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab; Takehi and Teoh are identified as equal contributors. The researchers call the work “very early results.” It is a blog post, not a peer-reviewed paper, and the source says it was not posted on arXiv.
For the Qwen experiment, the researchers gave Astra access to the model’s weights through their own Python-based setup. The logs describe a CPU-only environment without internet access, with a limit of 300 code runs. Astra could use practice questions but was not shown the final benchmark questions. The researchers prohibited gradient descent, backward passes and fitting weights to a loss. They tested search budgets of 10, 25 and 50 forward passes, each of which allowed the model to answer up to 512 problems.
According to the post, the base model hit its 32,768-token limit on 93.0% of MATH-500 answers. At 10 and 25 passes, its reported score rose only to 12.3% and 14.1%; at 50 passes it reached 48.1%, while the token-limit rate fell to 33.1%. The authors say Astra identified a direction associated with looping in the MLPs of layers 15 and 19 and subtracted it from 64 output columns in each layer. The edit affected 131,072 weights, or 0.016% of the model’s stated 0.8 billion weights. The published Hugging Face version, the post says, differs from the original in those two layers’ down_proj tensors.
A Targeted Fix, Not New Math Skill
The result matters because it suggests that a capable model may be able to help diagnose and alter the behavior of another model’s open weights, without a conventional gradient-descent training run. That could offer researchers and model operators another way to address a specific failure, such as repetitive output. The experiment also makes the limits visible: the model edit did not eliminate looping, and the strongest result required substantially more testing than the smaller search budgets.
The researchers themselves caution against reading the score increase as a new mathematical ability. “Fixing loops, however, is not the same as making a model better at math,” they write. In this case, the base model’s answers often consumed the available output budget before completing a response. Reducing those loops may have allowed more answers to finish, contributing to a higher benchmark score. The published figures do not, on their own, establish how much of the gain reflects improved problem-solving rather than fewer incomplete answers.
The approach also has practical limits for companies and developers. It applies to models whose weights can be accessed and changed; it cannot be used in the same way on a closed model’s inaccessible parameters. And a weight edit creates a modified model version that needs its own testing for accuracy, safety and unintended effects. The source notes that similar single-direction edits have been studied to remove refusal behavior from open chat models, illustrating that a targeted edit can change behaviors beyond the intended task.
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 with M5 Max 18-core CPU 20-core GPU 14.2-inch Display 48GB Memory 2TB SSD Space Black
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 Weight Search Worked
The experiment is part of a broader set of demonstrations in which the authors asked Astra to inspect or alter neural-network weights. The post describes attempts ranging from identifying what game a small network had learned to repairing a failing car-driving policy. Not every task succeeded: the authors report that Astra could not determine, from weights alone, which of two fine-tuned Gemma models had been trained to flatter users.
The Qwen result was run with GPT-6 Astra through the Codex command-line tool at the highest reasoning setting described in the logs. The researchers’ harness, rather than built-in Codex tools, provided access to the weights and test feedback. They supplied practice material from GSM8K and MATH and reserved the final benchmarks from Astra. The base and edited models used the same stated sampling settings for the reported comparisons.
The post also reports changes on other benchmarks. On GSM8K, the score rose from 13.8% for the base model to 67.1% after 50 passes; on IFEval, it rose from 21.0% to 62.3%. These are results reported by the authors, not independent replications. The source says smaller budgets produced much less improvement, and reports that a separate test on Gemma 3 270M found no edit that changed the score by more than about a point.
““Fixing loops, however, is not the same as making a model better at math.””
— Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, in their blog post
What the Benchmark Cannot Settle
The findings have not been independently verified in the source material, and the blog post has not undergone peer review. It is also unclear whether other researchers can reproduce the result under the same setup, or whether the change will hold across different question sets, sampling choices and model uses. The reported benchmark scores alone do not establish that the edit improves mathematical reasoning independently of its effect on answer length and looping.
The post does not provide evidence that the modified model is more reliable across general tasks or safe for deployment without further evaluation. The source also says OpenAI was not an author, did not receive thanks in the post and had made no statement about the experiment. The relationship, if any, between OpenAI and the researchers’ use of Astra is not described beyond the tool and model named in the logs.
Replication and Broader Testing
The next useful steps are independent replication and tests that separate completed-answer accuracy from changes in output length and loop frequency. Further evaluation could check whether the edit works across more mathematical problems, whether it affects unrelated capabilities, and whether a similar search succeeds on other models. The researchers’ post describes the work as preliminary; it does not announce a follow-up paper, a release schedule or an official OpenAI assessment.
For now, the edited Qwen model and published logs provide material for other researchers to inspect. Until follow-up testing is available, the reported score increase should be treated as a promising result from one disclosed experiment, not as evidence that AI can generally improve models without training.
Key Questions
What changed in Qwen3.5-0.8B?
The researchers say GPT-6 Astra subtracted a direction associated with repetitive outputs from weight columns in two layers. The published version differs from the original in the down_proj tensors of layers 15 and 19.
Did the edit remove the model’s loops?
No. The reported share of MATH-500 answers that hit the 32,768-token limit fell from 93.0% to 33.1%, but loops were still reported in about a third of answers.
Was the model trained with gradient descent?
The researchers say they prohibited gradient descent, backward passes and fitting weights to a loss. Astra used iterative searches with code and forward-pass feedback; the 50-pass run took 20.3 hours according to the source.
Has the result been peer-reviewed?
No peer review is reported in the source. The result appeared in a November 2, 2026 blog post that the authors describe as presenting “very early results.”
Does the score prove the model gained mathematical reasoning ability?
No. The score is a reported benchmark result, and the authors caution that fixing loops is not the same as improving math ability. The available information does not isolate how much of the gain came from fewer answers running into the output limit.
Source: ThorstenMeyerAI.com
