🔍 Read the full analysis: When GPT-6 Astra Edited Another Model, The Aim Was A Loop Fix on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A team of four researchers listed with MIT CSAIL used GPT-6 Astra to edit two weight matrices in the open Qwen3.5-0.8B model, reporting a rise in MATH-500 accuracy from 7.9% to 48.1%. The edit sharply reduced, but did not eliminate, answers that ran to the 32,768-token limit; the result comes from an early, non-peer-reviewed blog post and required extensive iterative testing.
Four researchers listed with MIT CSAIL report that OpenAI’s GPT-6 Astra edited two layers of the open Qwen3.5-0.8B model to reduce repetitive reasoning, raising its reported MATH-500 score from 7.9% to 48.1%. The change did not remove the problem entirely: 33.1% of edited-model answers still reached the benchmark’s 32,768-token limit, compared with 93.0% before the edit.
The results appear in a blog post titled Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights, published November 2 by researcher Rikiya Takehi. Takehi, Jayden Teoh, Alex Zhang and Omar Khattab are listed as authors, with Takehi and Teoh identified as equal contributors. The authors describe the work as “very early results”; the post is not peer-reviewed and, according to the source report, is not on arXiv.
According to the post and its logs, Astra used a Python environment to inspect and modify the model’s weights. The researchers prohibited gradient descent and backward passes, but the process still involved iterative testing and feedback. The highest-budget run lasted 20.3 hours and used 158 turns, 80 Python runs and 50 model forward passes. Each pass could evaluate up to 512 problems or process up to 1,024 texts. Astra was given practice questions, while the final benchmark questions were withheld.
The researchers say Astra traced looping to a direction associated with the MLPs in layers 15 and 19. It subtracted that direction from 64 output columns in each layer, changing 131,072 weights—about 0.016% of the model’s 0.8 billion parameters. The source report says the Hugging Face release shows differences in only the two corresponding down-projection weight tensors.
The reported MATH-500 results varied with the testing budget: 10 passes produced a 12.3% score, 25 passes produced 14.1%, and 50 passes produced 48.1%. The share of answers reaching the token cap was 87.2%, 86.4% and 33.1%, respectively, against the base model’s 93.0%. On GSM8K, the reported score rose from 13.8% to 67.1% in the 50-pass run; on IFEval, it rose from 21.0% to 62.3%.
A Repair, Not a New Math Ability
The result is relevant because it suggests a language model may help diagnose and alter the internal weights of another model, rather than merely recommend changes to its prompts or code. For organizations using open-weight models, targeted edits could offer a way to address specific failure patterns without repeating a full training run. The researchers’ logs also make the procedure easier for others to examine and attempt to reproduce.
But the reported gains should be read as a specific repair to a known failure, not proof that Astra broadly improved mathematical reasoning. Qwen’s model card already warns that the model can fall into thinking loops, and the researchers distinguish reducing that behavior from creating a new capability. The strongest result also depended on the largest test budget; 10- and 25-pass runs yielded much smaller improvements. A model edit is not automatically a reliable product fix: it changes the weights, and the altered model needs its own evaluation for performance and side effects.
The work also has a dual-use dimension. The post connects single-direction weight edits to research on removing refusal behavior from open chat models. That does not establish that this experiment changed safety behavior, but it shows why methods for directly editing weights may matter beyond benchmark scores.
Popular right now
Breville BES870XL Barista Express Espresso Machine, Brushed Stainless Steel
Vitamix Propel Series 750 Professional-Grade Blender, 4 Programs, Black
Ninja Foodi Air Fryer, 8 Qt DualZone, 6-in-1, Grey, DZ201
As an affiliate, we earn on qualifying purchases.
How the Weight Edit Worked
The experiment was one of eight in the researchers’ post, which explores whether a model can inspect and manipulate another neural network’s weights. The examples ranged from identifying what game a small network had learned to repairing a car-driving policy. The post also describes a failed test: Astra could not determine from weights alone which of two fine-tuned Gemma models had been trained to flatter users. The authors present the collection as early and limited, not as a general capability demonstration.
In the Qwen test, the researchers supplied practice material from GSM8K and MATH, but kept the final evaluation sets from Astra. The model itself could not freely run Qwen; it had to request a limited forward-pass budget to observe outputs. That setup matters when interpreting the phrase “without training”: no gradient-based training was used, but the edit was still the product of repeated experiments, programming and evaluation.
The source report says Astra was run as gpt-6-astra through the Codex command-line tool, with built-in tools disabled and a CPU-only, offline Python harness. OpenAI is not listed as a co-author, and the source report found no OpenAI statement about the work. The reported findings therefore come from the research team’s publication and logs, rather than an OpenAI announcement or independent peer-reviewed assessment.
““very early results””
— The researchers, in their blog post
Reproducibility and Side Effects
The published results have not been peer-reviewed, and the available source material does not establish whether independent researchers have reproduced the edits or benchmark scores. The results are reported by the authors; the logs and released model offer material for scrutiny, but do not by themselves settle questions about reproducibility or evaluation robustness.
It is also unclear whether the weight change affects other tasks, prompts or deployment settings in ways not captured by the listed benchmarks. The reported 50-pass result leaves a substantial fraction of MATH-500 answers at the token cap, so the looping behavior was reduced rather than eliminated. The source material does not provide a broad assessment of unintended changes or establish whether the method generalizes to larger models.
Finally, the work shows no evidence that GPT-6 Astra can routinely improve arbitrary models. The same post includes an experiment in which Astra could not infer a model’s training preference from its weights. The authors’ wider claims about model editing remain limited to the experiments described.
Independent Tests Will Matter
The immediate next step is independent evaluation of the released edited weights against the original model, using the same sampling settings and additional tasks. Researchers can check whether the reported benchmark changes hold up, whether looping falls under different prompts, and whether the edits cause regressions elsewhere. The blog post and model release provide a basis for that testing, but the source material does not identify a scheduled replication or formal peer-review process.
Further work would also need to establish how reliably this approach works across models and failure types, and how much testing is needed before an edit can be trusted. For now, the reported result is a promising but preliminary demonstration of targeted weight editing—not evidence that a model can generally train another model on its own.
Key Questions
What did GPT-6 Astra change?
According to the researchers, Astra identified a direction linked to repetitive reasoning and removed it from output columns in the MLP weight matrices of Qwen3.5-0.8B’s layers 15 and 19.
Did the edit eliminate Qwen’s looping?
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 run, so looping was reduced but remained common.
Was this done without training?
It was done without gradient descent or backward passes, according to the researchers. But “without training” does not mean without experimentation: the best run involved 158 turns, 80 Python runs and 50 forward passes.
Has the result been peer-reviewed?
The source report says the November 2, 2026 post is not peer-reviewed and is not on arXiv. The results are reported by the authors and have not been independently established in the material provided.
Did OpenAI conduct or endorse the experiment?
The four researchers are listed with MIT CSAIL, and GPT-6 Astra was the model used to perform the edit. The source report says OpenAI is not an author and found no OpenAI statement about the experiment.
Source: ThorstenMeyerAI.com
