TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
A November 2, 2026, research blog post reports that GPT-6 Astra changed 0.016% of Qwen3.5-0.8B’s weights and raised its MATH-500 score from 7.9% to 48.1%. The experiment reduced, but did not eliminate, long reasoning loops; it is an early, non-peer-reviewed result that used extensive testing and iteration.
In “Sparks of Weight-Space Programming,” Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab describe eight experiments in which GPT-6 Astra inspected or changed neural network weights. The Qwen math test, experiment eight, drew attention because the model’s score rose substantially after a small edit. The authors call their findings “very early results” and say the capabilities remain limited. The post has not been peer reviewed, and the supplied material says it is not on arXiv.
The researchers report that GPT-6 Astra identified a direction associated with looping in the MLPs of layers 15 and 19, then subtracted that direction from 64 output columns in each layer. The edit affected 131,072 weights, or 0.016% of Qwen3.5-0.8B’s 0.8 billion weights. The edited model published on Hugging Face differs from the original in the two layers’ down-projection weight tensors, according to the post.
The improvement came with a substantial testing effort. The winning run used 158 turns, 80 Python runs and 50 forward passes over 20.3 hours. The researchers provided practice problems but withheld the final benchmark problems. The source says smaller budgets produced scores of 12.3% with 10 passes and 14.1% with 25. It also reports that the edit reduced the share of MATH-500 answers reaching the 32,768-token limit from 93.0% to 33.1%.
A Small Edit, a Big Score Shift GPT-6 Astra’s Weight Experiment
Researchers report that GPT-6 Astra changed a small share of Qwen3.5-0.8B’s weights and raised its MATH-500 score. The early result reduced repetitive reasoning, but did not eliminate it.
What changed
The edit targeted a repetitive reasoning pattern in two layers of Qwen3.5-0.8B.
Two model layers
The researchers identified looping linked to layers 15 and 19.
A targeted weight edit
They changed 131,072 weights without gradient descent.
How the search worked
Astra tested candidate edits in a constrained Python environment using practice problems while final benchmark problems stayed hidden.
Many initial answers ran to the token limit.
The search focused on a direction associated with looping.
The winning run used 158 turns over 20.3 hours.
What the result shows
The reported benchmark gain came alongside fewer answers exhausting their token budget.
- MATH-500 score: 7.9%
- 93.0% of answers hit the token limit
- MATH-500 score: 48.1%
- 33.1% of answers hit the token limit
What remains uncertain
The finding comes from one model and test setup, and the report has not been peer reviewed.
- Independent replication has not been established.
- The edit’s effects on other models, tasks, and behaviors remain unclear.
- Reducing loops does not prove a general improvement in math ability.
A Small Edit With Measurable Gains
The result points to a possible way to alter an open-weight model’s behavior without retraining all of its parameters with gradient descent. In this experiment, the edited share was 0.016% of the weights, yet the reported benchmark score increased by 40.2 percentage points. For developers able to inspect and modify model weights, that suggests targeted repairs may be worth investigating when a model has a specific, repeatable failure.
The size of the score change does not establish that Astra gave Qwen a new general ability to solve mathematics. The authors explicitly caution that “Fixing loops, however, is not the same as making a model better at math.” The test focused on a model whose own card reportedly warns of thinking loops. A benchmark improvement after reducing that failure is evidence of a repair in this setting, not proof that the same method will improve other models or tasks.
There are also practical limits. Finding the edit took a long, iterative process, and the behavior persisted in roughly one-third of test answers. Every weight patch creates a modified model whose behavior needs evaluation. The report raises a broader question for open-weight AI: if a small internal change can suppress an unwanted behavior, what other behaviors could such edits affect? The supplied material cites refusal removal as another documented use of single-direction edits, but it does not show that this experiment involved refusal changes.
Products Worth Considering
Apple Watch Series 12 GPS 46mm Smartwatch
Amazon Kindle Scribe Colorsoft, 64GB, 11-Inch Color Display with Premium Pen
Logitech MX Brio Ultra HD 4K Webcam, Graphite
As an affiliate, we earn on qualifying purchases.
How the Weight Search Worked
The researchers gave GPT-6 Astra access to a Python environment that could inspect and modify Qwen’s weights. According to the published logs described in the source, Astra ran through the Codex command-line tool at its highest reasoning setting. The harness was CPU-only, had no internet access and allowed up to 300 code runs. Astra could not freely run Qwen: each forward pass was a limited opportunity to observe its outputs, covering up to 512 problems or 1,024 texts.
The team barred gradient descent, backward passes and fitting the weights to a loss. It supplied 7,469 GSM8K practice problems and 7,275 MATH practice problems, while keeping the final test sets hidden. On the initial practice sample, 85 of 96 answers ran to the 32,768-token cap, according to the post. That gave the search a clear symptom to investigate: repeated reasoning that exhausted the answer limit.
The post describes eight experiments, not all of them successful. In one, Astra could not determine from weights alone which of two fine-tuned Gemma models had been trained to flatter users. The report situates the Qwen result within prior work on removing behaviors through weight edits, citing a 2024 paper by Arditi and colleagues on refusal. That history is relevant background, but it does not independently validate the new experiment’s results.
““Fixing loops, however, is not the same as making a model better at math.””
— Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, the report’s authors
What the Experiment Has Not Shown
The report is a blog post, not a peer-reviewed study, and the evidence described here does not establish whether other researchers can reproduce the result. It is also unclear how well the edit holds up on different math benchmarks, other model sizes or tasks beyond the tested setup. The remaining 33.1% token-limit rate shows that looping was reduced rather than eliminated.
The source says OpenAI did not author or acknowledge the post and provides no statement from the company about the experiment. It does not describe a deployment, a product change or a formal collaboration between OpenAI and the researchers. The report’s performance figures come from its authors; independent replication and fuller evaluation remain outstanding.
Replication Will Test Its Reach
The next meaningful step is independent evaluation of the published edit and its hashes against the original model, including checks on math performance, answer length and other behaviors. Researchers can also test whether the same search approach works under smaller budgets or transfers to other models. The authors’ own budget results suggest that more testing time mattered in this case.
The source material does not announce a follow-up date or a planned product release. For now, the public record described in the post consists of the researchers’ report, logs and edited model. Whether targeted weight editing becomes a repeatable technique will depend on replication and evidence that gains persist without introducing other unwanted changes.
Key Questions
What did GPT-6 Astra change?
According to the researchers, Astra subtracted a direction associated with looping from output columns in the MLPs of layers 15 and 19 of Qwen3.5-0.8B. The reported edit changed 131,072 weights.
How much did the MATH-500 score improve?
The report gives a score increase from 7.9% to 48.1%. The authors caution that reducing loops is not the same as adding a general math ability.
Were the loops eliminated?
No. The reported share of answers hitting the 32,768-token limit fell from 93.0% to 33.1%, so a substantial share still reached the cap.
Did the experiment use no training at all?
It did not use gradient descent, backward passes or fitting to a loss, according to the post. But the winning search involved 158 turns, 80 Python runs and 50 forward passes over 20.3 hours, so “without training” does not mean without extensive testing and iteration.
Has the result been peer reviewed or endorsed by OpenAI?
The source describes it as a blog post and says it has not been peer reviewed. It also says OpenAI was not an author and had made no statement about the experiment.
Source: Thorsten Meyer AI
