🔍 Read the full analysis: How AI-to-AI Weight Editing Tackled A Model Loop on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Four researchers associated with MIT CSAIL report that GPT-6 Astra edited two layers of Qwen3.5-0.8B without gradient descent, raising its MATH-500 score from 7.9% to 48.1%. The model still hit the 32,768-token limit on 33.1% of answers, and the result comes from an early, non-peer-reviewed blog post rather than a paper. The work used extensive iterative testing, so “without training” does not mean without substantial computation or evaluation.
Four researchers associated with MIT CSAIL say OpenAI’s GPT-6 Astra improved a small open-weight model’s math benchmark performance by directly editing its weights, without gradient descent. In their November 2 blog post, the team reports that Qwen3.5-0.8B rose from 7.9% to 48.1% on MATH-500 after Astra identified and removed a direction linked to repetitive answers; the edit reduced, but did not eliminate, those loops.
The experiment is part of “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights,” published 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 it is not posted on arXiv, according to the source account. Takehi and Teoh are identified as equal contributors.
The team reports that Astra traced the repetition 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, or 0.016% of the model’s 0.8 billion parameters. The edited Hugging Face model reportedly differs from the original only in the two layers’ down-projection weight tensors; the researchers published hashes for those tensors.
The benchmark results came after a substantial search process. According to the published logs, Astra made 158 turns, ran Python 80 times and used 50 model forward passes over 20.3 hours. The team prohibited gradient descent and backward passes, but Astra received test feedback through practice problems and repeated model evaluations. The final benchmarks were not shown to Astra, the authors say.
A Repair, Not a New Math Ability
The reported gain matters because it suggests a language model can help inspect and modify another model’s weights, potentially offering a way to address specific failures in open-weight systems. Here, the change targeted a known problem: Qwen’s model card warns that the model can enter thinking loops. The authors themselves caution that “Fixing loops, however, is not the same as making a model better at math.”
The distinction matters for interpreting the score increase. The base model frequently spent its answer budget repeating itself, so reducing that behavior gave it more opportunity to produce a usable answer. After editing, 33.1% of MATH-500 answers still reached the 32,768-token cap, compared with 93.0% before. The experiment therefore supports a claim about improving performance by repairing a failure mode, not a general claim that Astra taught the smaller model mathematics.
The method also has limits for organizations using closed systems: it requires access to the target model’s weights. Any direct weight patch produces a modified model that must be tested and governed as a separate version. The source account also notes that related single-direction edits have been studied as a way to remove refusal behavior from open chat models, underscoring that targeted weight changes can have consequences beyond the intended benchmark.
Popular right now
Ninja Luxe Café Mini Plus Espresso and Drip Coffee Machine with AutoFroth, Stainless Steel (ES351)
Apple AirPods Max 2 Wireless Over-Ear Headphones – Midnight
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Hazelnut Taupe
As an affiliate, we earn on qualifying purchases.
How the Weight Search Worked
Astra did not simply receive the weights and return a finished model. The researchers ran it through the Codex command-line tool at its highest reasoning setting and built a custom harness that gave it a Python kernel for inspecting and editing weights. The setup was CPU-only and offline, with a limit of 300 code runs. Astra could not run Qwen directly: each forward pass used to observe its behavior consumed part of a limited budget.
For practice, the harness supplied 7,469 problems from GSM8K and 7,275 from MATH, which the authors say were separate from the final test sets. The researchers compared budgets of 10, 25 and 50 forward passes. On MATH-500, the corresponding scores were 12.3%, 14.1% and 48.1%; the share reaching the token limit was 87.2%, 86.4% and 33.1%. On GSM8K, the 50-pass result was 67.1%, up from a base score of 13.8%. The authors’ post covers eight experiments, including attempts that did not succeed, rather than presenting this result as a standalone paper.
The improvement was not consistent across the smaller budgets or every model tested. A separate Gemma 3 270M experiment reportedly showed no score change greater than about one point. In the Qwen trial, 10 and 25 passes produced only modest MATH-500 gains, while the larger 50-pass run produced the striking result. That pattern makes the size of the evaluation budget part of the story, not a minor detail.
““Fixing loops, however, is not the same as making a model better at math.””
— The researchers, in “Sparks of Weight-Space Programming”
Limits of the Reported Evidence
The result has not been validated in a peer-reviewed publication, and the available account is based on the researchers’ blog post, logs and released model. The source material does not establish whether independent teams have reproduced the edit or whether the benchmark gains hold across other sampling settings, math datasets or real-world use. The researchers call the findings preliminary.
It is also unclear how broadly the method generalizes. The reported Qwen improvement followed 50 forward passes and 20.3 hours of work, while smaller budgets yielded much lower scores and the Gemma test showed little change. The post does not establish a reliable rule for predicting which models or failure modes can be addressed this way.
The source account says OpenAI did not author the post, was not acknowledged in it and had issued no statement about the experiment. The team’s use of GPT-6 Astra is described through its logs; that does not amount to an endorsement by OpenAI. The published information also does not show whether the edit affects other behaviors or how the modified model performs beyond the reported evaluations.
Replication and Broader Testing
The immediate next step is independent replication of the released edit and its reported benchmarks, with the same evaluation settings and clear reporting of token-limit failures. Further tests would need to establish whether the score gains persist on additional math tasks and whether the edit changes unrelated model behaviors.
The researchers have not announced a timetable for a peer-reviewed paper or a larger validation study in the supplied material. Until those results are available, the finding is best treated as a promising but narrow demonstration: an AI system helped patch a specific failure in an open-weight model, using repeated testing and direct access to its weights.
Key Questions
What did GPT-6 Astra change?
According to the researchers, Astra removed a direction linked to repetitive output from selected weights in layers 15 and 19 of Qwen3.5-0.8B. The edit changed 131,072 weights, which the post describes as 0.016% of the model’s total.
Did the researchers train Qwen with gradient descent?
No gradient descent, backward passes or fitting weights to a loss were allowed, according to the published account. But Astra used iterative searches, Python runs and model evaluations; “without training” does not mean the process required no testing or computation.
Did the edit stop the model from looping?
No. The reported share of MATH-500 answers that reached the 32,768-token limit fell from 93.0% to 33.1%. The researchers report a substantial reduction, not the elimination, of looping.
Has the result been peer-reviewed?
The source describes the work as a blog post published November 2, 2026, not a peer-reviewed paper. The authors call the findings “very early results,” and independent replication is not established in the available material.
Can this approach be used on any AI model?
The report does not show that it works reliably across models. The Qwen result depended on a large evaluation budget, while a reported Gemma 3 270M test showed no change greater than about one point. The method also requires access to the model’s weights.
Source: ThorstenMeyerAI.com
