🔍 Read the full analysis: Can A Weight Edit Fix An AI Loop Without Making It Smarter? 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 November 2 blog post by four researchers listed with MIT CSAIL reports that OpenAI’s GPT-6 Astra edited two weight tensors in the open Qwen3.5-0.8B model without gradient descent. On MATH-500, the edited model scored 48.1%, up from 7.9%, while answers hitting the token limit fell from 93.0% to 33.1%. The work is an early, unreviewed result: it used 50 model-evaluation passes over 20.3 hours, and the loop behavior remained in a substantial share of answers.
Four researchers listed with MIT CSAIL report that OpenAI’s GPT-6 Astra edited a small number of weights in the open model Qwen3.5-0.8B, raising its MATH-500 score from 7.9% to 48.1% without gradient descent. The change also reduced answers that ran into the benchmark’s 32,768-token limit from 93.0% to 33.1%—a marked reduction, but not an end to the looping behavior.
The results appear in “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights,” a blog post published November 2 by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. Takehi and Teoh are identified as equal contributors. The post has published logs, but is not a peer-reviewed paper; its authors describe the work as “very early results.” OpenAI is not listed as an author, and the source material reports no OpenAI statement about the experiment.
According to the researchers’ logs, Astra used a Python environment to inspect and edit the model’s weights, with access to practice problems but not the final benchmark questions. Gradient descent and backward passes were prohibited. The strongest run used 50 forward passes, 80 Python runs and 158 turns over 20.3 hours. A forward pass let the system inspect outputs on up to 512 problems or 1,024 texts. With only 10 or 25 passes, the MATH-500 scores were 12.3% and 14.1%, respectively.
The researchers say Astra traced repetitive 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 published edited model reportedly differs from the original in only those two layers’ down-projection tensors.
At the 50-pass setting, the researchers also report that GSM8K accuracy rose from 13.8% to 67.1%, while IFEval rose from 21.0% to 62.3%. These are results reported by the blog post, using its described evaluation setup; they should not be read as independently validated performance figures.
A Repair, Not a New Math Ability
The result matters because it suggests a way to change an open model’s behavior after training, without running a conventional training process. A targeted edit may be useful when a model has a specific failure—such as repeating itself—that makes it hard to use. Reducing that failure can also improve benchmark scores because a model that stops consuming its output budget on loops has more opportunity to answer the question.
That distinction limits what the result establishes. The researchers explicitly caution that fixing loops is not the same as making a model better at math. The reported gains may reflect the removal of a barrier that prevented Qwen from completing answers, rather than the addition of deeper mathematical knowledge. The remaining 33.1% token-limit rate shows that the problem was reduced, not removed.
The method also required substantial iterative work and access to the model’s raw weights. Its relevance is therefore greatest for teams that can inspect and alter open-weight models, not users of services that expose only a chatbot interface. And a weight patch produces a modified model that needs its own testing: changed behavior in one benchmark does not establish reliability across other tasks or settings.
Products Worth Considering
Apple Mac mini Desktop Computer with M4 chip, 10‑core CPU and GPU, 24GB Unified Memory, 512GB SSD, Gigabit Ethernet
Apple MacBook Pro 16-inch (2021) with M1 Max, 64GB RAM, 1TB SSD
Apple MacBook Pro 14 2021 with M1 Max, 64GB RAM, 2TB SSD, Silver
As an affiliate, we earn on qualifying purchases.
The Looping Problem in Qwen
Qwen3.5-0.8B is a small open model whose model card warns of a tendency to enter thinking loops, according to the source post. In the researchers’ initial practice set, 85 of the first 96 answers reached the 32,768-token ceiling. On MATH-500, the unedited model scored 7.9%, with 93.0% of answers hitting that limit.
The post describes eight experiments involving weight inspection or editing, including attempts to identify training behavior and repair a car-driving policy. One test did not succeed: Astra could not determine from weights alone which of two fine-tuned Gemma models had been trained to flatter users. This mixed set of outcomes matters: the Qwen result is one experiment within a broader set of early demonstrations, not evidence that a model can reliably rewrite any other model’s capabilities.
The authors distinguish this approach from training by ruling out gradient descent and fitting weights to a loss. But “without training” can be misleading if taken to mean effortless or resource-free. Astra used repeated evaluations, Python analysis and a sizable run budget to choose and test its edit. The report also references earlier research on removing refusal behavior from models by editing a direction in their weights, illustrating that targeted behavioral edits have prior research context.
““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 November 2 blog post
How Broad the Fix Really Is
The report does not establish whether the edit improves performance across a wider range of math problems, prompts or deployment conditions. The final benchmark questions were withheld from Astra, according to the logs, but independent replication and peer review are not reported in the supplied material. It is also unclear how stable the change would be across different sampling settings or subsequent model updates.
The looping rate remained high: 33.1% of MATH-500 answers still hit the token limit after the largest run. The source gives scores for GSM8K and IFEval as well, but the excerpt does not provide enough detail to assess all evaluation choices or determine whether other capabilities were affected. Nor does it establish that the same approach will work consistently on other models; the cited Gemma 3 270M tests reportedly showed changes of no more than about a point.
The post’s claims about Astra’s actions are based on the researchers’ logs and released model files. The supplied source says OpenAI did not comment on the experiment. The exact practical costs and reproducibility of the method outside the reported setup remain open questions.
Replication and Wider Testing
The immediate next step is for other researchers to inspect the published logs and edited model and test whether they can reproduce the results. The edited weights and reported hashes offer a way to compare the patched model with the original, but independent confirmation is not included in the source material. Broader evaluations would need to test whether the edit reduces loops across tasks without harming other model behavior.
The authors frame the work as early research and outline multiple experiments rather than a finished method for routine model repair. No follow-up paper, peer-review timeline or further release is specified in the supplied material. For now, the central result remains a reported proof of concept: a model used iterative feedback to make a narrow weight edit, with large gains on selected benchmarks and a substantial failure rate still present.
Key Questions
Did GPT-6 Astra train Qwen3.5-0.8B?
Not through gradient descent, according to the researchers. They say Astra edited weights directly after inspecting model outputs and practice problems. The process still involved iterative testing and analysis, so “without training” does not mean without substantial work.
Did the edit eliminate Qwen’s looping?
No. On MATH-500, the share of answers reaching the 32,768-token limit fell from 93.0% to 33.1% in the 50-pass run. The reported figures show a reduction, not elimination.
How much of the model’s weights changed?
The researchers report changing 131,072 weights across two layers, about 0.016% of the model’s 0.8 billion weights. The released model reportedly differs from the original in the two layers’ down-projection tensors.
Has the result been peer reviewed?
Not according to the source material. The findings were published in a November 2, 2026 blog post, which the authors call very early results. Independent replication is not reported in the supplied material.
Does the score increase prove Qwen became better at math?
No. The authors caution that removing a loop is not the same as adding mathematical ability. The edit may have helped the model finish more answers, and the report does not establish broader improvements across tasks or conditions.
Source: ThorstenMeyerAI.com
