🔍 Read the full analysis: An AI Model Edited Another: Astra’s Loop-Focused Repair on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
A November 2 blog post by four researchers listed with MIT CSAIL reports that GPT-6 Astra edited two weight matrices in Qwen3.5-0.8B without gradient descent. The edited model scored 48.1% on MATH-500, up from 7.9%, while the share of answers hitting the 32,768-token limit fell from 93.0% to 33.1%. The result is an early, unreviewed report, and the authors describe it as a repair of looping rather than a general improvement in mathematical ability.
Four researchers listed with MIT CSAIL report that OpenAI’s GPT-6 Astra edited two layers of the open-weight model Qwen3.5-0.8B, reducing repetitive outputs and raising its reported MATH-500 score from 7.9% to 48.1%. The November 2 account describes an iterative search using test feedback but no gradient descent; the authors call the findings early and say that fixing loops is not the same as making a model better at math.
The post, titled Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights, lists Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab as authors. Takehi and Teoh are identified as equal contributors. The report is a blog post, not a peer-reviewed paper, and the authors characterize the work as “very early results.”
According to the researchers’ published logs, Astra examined the model’s behavior and traced repeated output 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 weights. The edited model’s published files show differences in the two layers’ down-projection weight tensors; the post provides their hashes.
On MATH-500, the base model scored 7.9%, and 93.0% of answers hit the 32,768-token limit. After the 50-pass edit, the reported score was 48.1%, while 33.1% of answers still reached that limit. Smaller budgets had smaller effects: the 10-pass and 25-pass runs scored 12.3% and 14.1%, respectively. The researchers also report changes on GSM8K and IFEval, but say a separate test on Gemma 3 270M produced no improvement greater than about one point.
A Targeted Fix, Not New Math
The result points to a possible way to inspect and alter open model weights using another model, rather than relying on conventional training. If similar methods prove dependable, developers could have another option for correcting identifiable behaviors in models they can access and modify. That possibility matters most where model weights are available: the reported intervention directly changes the files that define the model.
The benchmark gains need to be read alongside the failure-rate data. The edit sharply reduced the share of answers running into the output limit, but one-third of MATH-500 responses still did so. The authors themselves caution that reducing loops is not equivalent to adding mathematical skill. The scores indicate improved performance on the reported evaluations, not proof of broad reasoning gains or reliability in other settings.
The work also illustrates a distinction behind claims that AI can improve AI “without training.” Here, that phrase means without gradient descent or loss-based weight fitting; it does not mean without substantial computation, examples or evaluation. Astra performed a guided search with repeated feedback, and the final weight edit produced a new model version that would need its own testing.
Popular right now
Ninja Flip Toaster Oven & Air Fryer, Large Capacity, Stainless, SP151
Breville BES870XL Barista Express Espresso Machine, Brushed Stainless Steel
Dyson W1 PencilWash Cordless Wet and Dry Floor Cleaner with Dock
As an affiliate, we earn on qualifying purchases.
How the Weight Search Worked
The researchers gave Astra access to a Python-based harness operating on the model’s weights. The setup used CPU processing, had no internet access, and allowed up to 300 code runs. Astra could not run the model freely: it had to use a limited number of forward passes to inspect outputs. Each pass could cover up to 512 problems or read up to 1,024 texts, according to the post.
The team tested budgets of 10, 25 and 50 passes. The highest-budget run took 20.3 hours and involved 158 turns, 80 Python runs and 50 forward passes, the report says. Astra had access to practice material outside the final test sets: 7,469 GSM8K problems and 7,275 MATH problems. The researchers say the benchmark questions used for final scores were not shown to Astra.
The target model’s tendency to loop was not presented as a newly discovered defect. The source says Qwen’s model card warns about thinking loops, and the base model showed the behavior in the researchers’ own testing: 85 of its first 96 practice answers reached the token limit. The same post describes eight experiments, including tasks where Astra failed to identify which of two fine-tuned Gemma models had been trained to flatter users.
““Very early results.””
— Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, in their blog post
What the Report Has Not Established
The results have not been peer-reviewed, and the source material provides no independent replication. It is not yet clear whether the same editing approach will work reliably on other models, tasks or evaluation sets. The reported scores come from particular benchmarks and settings; they do not establish how the edited model performs across real-world mathematical work or whether the change affects other behaviors.
The looping rate also remains substantial: 33.1% of edited-model answers still reached the maximum token limit on MATH-500. The post reports improvements on GSM8K and IFEval, but those figures do not by themselves show that the edit caused broadly transferable gains. The source does not describe a full evaluation of possible side effects from changing the two weight tensors.
OpenAI is not listed as an author, and the source says the researchers did not thank the company in the post and found no OpenAI statement about the experiment. The exact commercial or institutional relationship, if any, is not established by the material provided. The report identifies Astra as the model used, but its results should be understood as the researchers’ account of an experiment rather than an OpenAI endorsement.
Replication and Broader Testing
The immediate next step is independent testing of the edited model and the published procedure. Replication could check whether the reported benchmark changes hold under the same sampling settings and whether comparable edits can reduce loops without damaging other capabilities. The researchers’ post makes model files and logs available, giving other evaluators material to examine, though the supplied source does not announce a date for a paper or follow-up study.
Further work would also need to test the method on different model families and tasks, compare it with other repair approaches, and measure side effects beyond the reported benchmarks. Until those checks are available, the result is best treated as an early demonstration of a targeted weight edit, not a general recipe for improving AI models.
Key Questions
What did GPT-6 Astra change in Qwen3.5-0.8B?
According to the researchers, Astra identified a direction associated with looping and removed it from output columns in the model’s layers 15 and 19. The change affected 131,072 weights, about 0.016% of the model’s total weights.
Did the edited model stop looping?
No. On MATH-500, the share of answers reaching the 32,768-token limit fell from 93.0% to 33.1%, but roughly one-third still hit the limit.
What does “without training” mean in this report?
The researchers say they did not use gradient descent, backward passes or fit weights to a loss. Astra still conducted an iterative search using practice examples and repeated model evaluations, so the process involved substantial computation and feedback.
Has the result been peer-reviewed?
No. The source describes the work as a blog post published November 2, 2026, and says it is not peer-reviewed. The authors call the findings very early.
Does the result show that Astra made Qwen generally better at math?
That is not established. The reported benchmark scores improved, but the authors say that fixing loops is not the same as making a model better at math. Independent replication and broader testing are still needed.
Source: ThorstenMeyerAI.com
