🔍 Read the full analysis: How One Model’s Weights Were Edited To Address An AI Loop 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
Four researchers affiliated with MIT CSAIL report that GPT-6 Astra edited 0.016% of Qwen3.5-0.8B’s weights, raising its MATH-500 score from 7.9% to 48.1%. The result comes from an early, non-peer-reviewed blog post; 33.1% of edited-model answers still reached the benchmark’s 32,768-token limit, and the method used 158 turns and 80 Python runs.
Four researchers affiliated with MIT CSAIL report that OpenAI’s GPT-6 Astra modified two layers of the open-weight model Qwen3.5-0.8B, raising its score on the MATH-500 benchmark from 7.9% to 48.1%. The November 2 post describes an iterative weight-editing method without gradient descent, but the improvement did not eliminate looping: 33.1% of edited-model answers still hit the 32,768-token limit.
The researchers — Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab — published their account, “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights,” with logs. The post says Astra traced repeated output to one direction associated with the model’s MLPs in layers 15 and 19, then subtracted that direction from 64 output columns in each layer. The change affected a reported 0.016% of the model’s 0.8 billion weights.
The reported MATH-500 score rose as the researchers gave Astra more opportunities to inspect results: 12.3% after 10 forward passes, 14.1% after 25 and 48.1% after 50. The base model reached the 32,768-token cap on 93.0% of answers; the edited model did so on 33.1%. The authors also report GSM8K scores rising from 13.8% to 67.1% after 50 passes, and IFEval scores rising from 21.0% to 62.3%.
“Fixing loops, however, is not the same as making a model better at math,” the researchers write. Their logs describe a substantial search: the winning run took 158 turns, 80 Python runs and 50 forward passes, over 20.3 hours. Astra received practice questions but not the final benchmark questions. The account says gradient descent, backward passes and fitting weights to a loss were prohibited; this means the work avoided that training procedure, not that it required no testing or computation.
A Repair, Not a New Math Skill
The result matters because it suggests a language model may help identify a localized defect in another model’s weights, offering a possible route for targeted repairs to open-weight systems. If reproducible, such edits could let developers address a specific failure without retraining every parameter. The reported change was confined to two tensors, and the edited version has been posted on Hugging Face, where its weight differences can be inspected.
But the evidence supports a narrower conclusion than the headline score alone might suggest. The authors frame the work as reducing a known looping behavior, not creating general mathematical ability. The residual token-limit rate remained high, and the reported benchmark gains depended on a large search budget. In lower-budget runs, MATH-500 scores were 12.3% and 14.1%; the post also says a loop-free model did not improve in the same way.
For organizations using open weights, a successful patch could be useful, but it also creates a modified model that must be evaluated and maintained. The source notes that related single-direction edits have been studied for removing refusal behavior from open chat models. That connection cuts both ways: weight editing may support repairs, but it may also alter safeguards. The current experiment does not establish how reliably the approach transfers to other models or tasks.
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
Dell XPS 13 9350 Copilot+ Laptop, 13.4-inch 120Hz Display, Intel Core Ultra 7 256V, 16GB RAM, 512GB SSD, Ubuntu Linux 24.04 LTS
As an affiliate, we earn on qualifying purchases.
How the Weight Search Worked
Qwen3.5-0.8B is a small open model, and its model card warns that it can fall into thinking loops. In the researchers’ initial practice sample, 85 of 96 answers reached the token cap. That behavior made the experiment a targeted intervention: the team asked Astra to improve math performance, while the immediate problem observed in the model was repeated output that consumed its response budget.
According to the published logs, Astra ran as gpt-6-astra through the Codex command-line tool at the highest reasoning setting. The researchers’ harness provided a CPU-only Python environment with access to the model’s weights, no internet connection and a limit of 300 code runs. Astra could not directly run the model; it requested forward passes, each of which could process up to 512 problems or read up to 1,024 texts. The researchers supplied 7,469 GSM8K and 7,275 MATH practice problems, while withholding the final test questions.
The post reports eight experiments, not all successful. In one, Astra could not identify which of two fine-tuned Gemma models had been trained to flatter users. The authors describe the broader results as “very early” and say the capabilities remain limited. The article is a blog post rather than a peer-reviewed paper, and the supplied source says it was not posted on arXiv.
“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 post
Reproducibility and Remaining Loops
The report has not been peer reviewed, and the source material does not establish whether independent researchers have replicated the edit or the benchmark scores. Although the post and logs are public and the edited weights are available, independent verification remains open. The source also gives no evidence that the same method will work consistently across model sizes or architectures.
It is not clear from the supplied account how much of the score increase came specifically from suppressing loops versus other effects of the weight change, or how the edit performs across a broader range of prompts and deployment settings. The benchmark results are reported for particular evaluation setups; they do not establish real-world accuracy. The edited model still hit the token cap on 33.1% of MATH-500 answers, so the loop was reduced rather than removed.
The post describes a no-gradient-descent process, but the phrase “without training” can obscure the work involved. Astra used repeated testing and feedback, and the strongest result required 50 forward passes over 20.3 hours. OpenAI’s involvement is also limited in the supplied reporting: the post used GPT-6 Astra, but OpenAI was not listed as an author, and the source reports no statement from the company about the experiment.
Independent Tests of the Edit
The next useful test is independent replication using the posted logs, original weights and edited model. Researchers can check whether the edit consistently lowers looping while preserving other capabilities, and whether the benchmark gains hold under different prompts and sampling settings. Tests across more models would show whether the technique is specific to this Qwen model or can address other localized failures.
The authors’ post provides artifacts that may support those checks, including the edited model and hashes for the changed tensors. No follow-up study or formal review is identified in the supplied material. Until such work is available, the reported score gains should be treated as early experimental results, not as evidence that models can generally improve themselves through weight editing.
Key Questions
What did GPT-6 Astra change?
According to the researchers’ post, Astra identified a direction linked to looping and subtracted it from output columns in layers 15 and 19 of Qwen3.5-0.8B. The change affected a reported 0.016% of the model’s weights.
How much did the MATH-500 score improve?
The researchers report an increase from 7.9% to 48.1% after the 50-pass run. This is a result from their reported benchmark setup, not a guarantee of performance on other math questions or in deployment.
Did the edit stop the model from looping?
No. The share of MATH-500 answers that reached the 32,768-token limit fell from 93.0% to 33.1%, according to the post, but looping remained in a substantial share of answers.
Was this done without training?
The researchers say they used no gradient descent, backward passes or loss-based fitting. However, the work involved iterative testing: the strongest reported run used 158 turns, 80 Python runs and 50 forward passes.
Has the result been independently verified?
The supplied source describes a blog post with logs and an edited model, not a peer-reviewed paper. It does not report independent replication, so the result’s reproducibility remains unsettled.
Source: ThorstenMeyerAI.com
