🔍 Read the full analysis: The Weight Edit That Fixed A Loop, Not Model Intelligence 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 affiliated with MIT CSAIL report that OpenAI’s GPT-6 Astra improved Qwen3.5-0.8B’s MATH-500 score from 7.9% to 48.1% by editing weights in two layers, without gradient descent. The result appears in a November 2 blog post, not a peer-reviewed paper; looping fell sharply but remained common, and the edit required substantial testing.
Four researchers affiliated with MIT CSAIL report that OpenAI’s GPT-6 Astra raised the Qwen3.5-0.8B model’s MATH-500 score from 7.9% to 48.1% by editing weights in two layers, without using gradient descent. The change targeted a repetitive-output problem, not a demonstrated new capacity for mathematical reasoning: 33.1% of edited-model answers still reached the 32,768-token limit, according to the team’s November 2 blog post.
The researchers—Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab—published their account under the title “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights.” They describe the work as “very early results”. It is a blog post with published logs, not a peer-reviewed paper, and the source says it is not on arXiv.
According to those logs, Astra identified a direction associated with repetitive outputs in the model’s multilayer perceptrons at layers 15 and 19. It subtracted that direction from 64 output columns in each layer, changing 131,072 weights—reported as 0.016% of the model’s 0.8 billion weights. The edited Hugging Face model has differences in the two layers’ down-projection weight tensors, the post says.
The largest reported result followed a budget of 50 forward passes, alongside 158 turns and 80 Python runs over 20.3 hours. The harness gave Astra access to the model’s weights and practice questions, but not the final benchmark questions. Gradient descent, backward passes and fitting weights to a loss were prohibited. MATH-500 results rose with the larger budgets: 12.3% after 10 passes, 14.1% after 25, and 48.1% after 50; the share of answers hitting the token cap fell from 93.0% to 87.2%, 86.4% and 33.1%, respectively.
A Repair, Not a Reasoning Breakthrough
The reported result points to a possible way of diagnosing and modifying open-weight models without conventional gradient-based training. If reproducible, targeted edits could help developers address specific failure modes in models they can inspect and run themselves. The team also reports gains beyond MATH-500: on GSM8K, the score rose from 13.8% to 67.1% after 50 passes; on IFEval, it rose from 21.0% to 62.3%.
Those figures do not establish that the model acquired a broad improvement in mathematical ability. The source authors explicitly caution that “fixing loops” is not the same as making a model better at math. The benchmark gains are reported by the researchers, and the remaining token-limit rate shows that the targeted behavior was reduced rather than removed.
The method also depends on access to editable model weights. It does not describe a way to patch a closed model whose weights are unavailable. And a modified checkpoint is, in practical terms, a new model version: developers would need to test its behavior and decide whether the trade-off is acceptable before deploying it.
Popular right now
De'Longhi Rivelia Fully Automatic Espresso Machine with Milk Frother, White
iRobot Roomba 105 Robot Vacuum with LiDAR Navigation
Roborock Robot Vacuum and Mop Combo with Self-Empty Dock
As an affiliate, we earn on qualifying purchases.
How the Weight Search Worked
The experiment was the eighth of eight tests in the blog post, which also describes attempts to infer information from weights and repair a failing car-driving policy. One attempt did not work: Astra could not tell from weights alone which of two fine-tuned Gemma models had been trained to flatter users, according to the researchers.
For the Qwen test, Astra could not freely run the model. The team’s harness allowed it to inspect and edit the weights and use a CPU-based Python environment, while model behavior had to be checked through a limited number of forward passes. Each pass could cover up to 512 problems or 1,024 texts. Practice material came from GSM8K and MATH; the final benchmark questions were withheld from Astra, the post says.
The approach is distinct from gradient descent, but “without training” does not mean the change required no evaluation or iteration. The 10- and 25-pass runs produced much smaller MATH-500 gains, and the source reports that an experiment on Gemma 3 270M did not improve its score by more than about a point. The results therefore suggest that performance depends on the model, target behavior and search budget.
““very early results””
— The researchers, in “Sparks of Weight-Space Programming”
Questions About Repeatability Remain
The result has not yet been described in a peer-reviewed paper, and the supplied source offers no independent replication. It is not clear whether other researchers will reproduce the score changes using the same model, evaluation settings and editing process, or whether the method will transfer to other model sizes and tasks.
The source reports benchmark scores and logs, but the provided material does not establish how stable the gains are across repeated evaluation runs or how the edit affects a wider range of tasks. Looping also persisted in one-third of MATH-500 outputs after the largest-budget edit. OpenAI’s role is also limited in the available account: the researchers used GPT-6 Astra, but the source says OpenAI was not an author and had issued no statement about the experiment.
Replication and Broader Testing
The next useful evidence would be independent replication, with full benchmark settings and repeated runs, followed by tests of whether the edited model retains performance outside the reported tasks. The researchers’ blog post supplies logs and identifies the edited model, giving other teams material to examine, but the source does not announce a follow-up paper or a specific replication schedule.
For now, the finding is best read as a reported demonstration that a model can help search for a targeted change to open weights under constrained conditions. Whether the technique can reliably repair other failures—and whether the gains outweigh unwanted effects—remains open.
Key Questions
What changed in Qwen3.5-0.8B?
The researchers report removing a direction linked to repetitive output from weight tensors in layers 15 and 19. They say the edit changed 131,072 weights, or 0.016% of the model’s total weights.
Did the edited model stop looping?
No. The reported share of MATH-500 answers reaching the 32,768-token limit fell from 93.0% to 33.1% after the largest-budget edit, so the behavior was reduced but not eliminated.
Was this training?
The researchers say they did not use gradient descent, backward passes or loss-based weight fitting. Astra still used an iterative search with practice questions, Python runs and 50 forward passes in the strongest reported run.
Has the result been peer-reviewed?
Not according to the supplied source. The findings were published as a November 2, 2026 blog post with logs; the source says the work is not peer-reviewed and is not on arXiv.
Can the method be used on any AI model?
The report does not show that. The method requires access to editable weights, and the source describes mixed results on other tests, including little change on Gemma 3 270M. Transfer to other models and tasks remains unconfirmed.
Source: ThorstenMeyerAI.com
