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 blog post by four researchers affiliated with MIT CSAIL reports that GPT-6 Astra reduced a looping problem in Qwen3.5-0.8B by editing a small number of weights in two layers. Its MATH-500 score rose from 7.9% to 48.1%, but 33.1% of answers still hit the token limit, and the work has not been peer-reviewed.
Four researchers affiliated with MIT CSAIL report that OpenAI’s GPT-6 Astra reduced repetitive outputs in the open-weight model Qwen3.5-0.8B by editing weights in two layers, without using gradient descent. In their November 2 blog post, they say the model’s MATH-500 score rose from 7.9% to 48.1%; the result is an early, unreviewed report, and many answers still ran into the model’s token limit.
The researchers — Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab — describe the experiment in a post titled “Sparks of Weight-Space Programming.” They say Astra examined practice problems and model outputs, then traced looping to a direction associated with the MLPs in layers 15 and 19. The edit changed 131,072 weights, about 0.016% of the model’s 0.8 billion weights, according to the post. The authors say the edited model’s public Hugging Face files show differences in the two layers’ down-projection tensors.
The reported change was substantial on the benchmark but did not eliminate loops. Before the edit, 93.0% of MATH-500 answers hit the 32,768-token limit; afterward, 33.1% still did. The reported score increase therefore came alongside a remaining failure mode. The authors caution that correcting loops is not the same as improving a model’s general mathematical ability.
The method also involved more than a handful of casual prompts. The published logs describe a 50-forward-pass run lasting 20.3 hours, with 158 turns and 80 Python runs. Astra had access to practice sets but not the final benchmark questions. The researchers prohibited gradient descent, backward passes and fitting weights to a loss. Runs with 10 and 25 forward passes produced lower MATH-500 scores of 12.3% and 14.1%, respectively, according to the post.
GPT-6 Astra Tweaked Another AI’s Weights to Address a Loop
Researchers report that a targeted weight edit reduced repetitive outputs in Qwen3.5-0.8B and raised its MATH-500 score. The early result is unreviewed, and looping remained.
A targeted repair
The researchers describe changing a direction associated with repetitive outputs in one open-weight model.
Qwen3.5-0.8B
The intervention focused on looping during math responses.
Layers 15 and 19
Astra modified MLP down-projection weights without gradient descent.
How the search worked
Astra used code and repeated forward-pass feedback to inspect outputs and test changes.
Practice questions came from GSM8K and MATH; final benchmark items were withheld.
The researchers linked repetitive outputs to a direction associated with the MLPs.
They tested the change on MATH-500 and reported a higher score alongside fewer limit hits.
What changed—and what remains
The reported improvement addresses a specific failure mode, not general mathematical ability.
- MATH-500 score rose from 7.9% to 48.1%.
- Fewer answers reached the token limit.
- 33.1% of answers still hit the limit.
- Effects on other tasks and behaviors remain unclear.
A Targeted Repair With Limits
The experiment suggests that a language model may help inspect and modify another model’s weights when given code access, a way to run the model, and repeated feedback. For developers working with open-weight models, that could offer another way to diagnose a specific failure without conventional gradient-based training. The edit is small relative to the full model, but the reported process still required extensive evaluation and human-built tooling.
The result does not establish that Astra can broadly improve models, or that a small edit reliably raises general reasoning ability. In the researchers’ other reported tests, edits to Gemma 3 270M shifted scores by no more than about a point, and Astra failed to identify which of two fine-tuned Gemma models had been trained to flatter users. The Qwen result is best read as a targeted intervention on a known weakness, with residual looping and limited evidence about transfer to other tasks.
Weight edits can also have consequences beyond the immediate benchmark. The post points to earlier research describing how a single direction in open chat models can be associated with refusal behavior. That makes the technique relevant to model reliability and control as well as performance: changing a narrow internal feature may alter behavior in ways that require careful evaluation. The Qwen experiment does not, by itself, show that the same approach is broadly safe or predictable.
Products Worth Considering
LG UltraGear 27GX704A-B 27-inch QHD OLED Gaming Monitor
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black
DJI Mic 3 Wireless Microphone Kit (2 Transmitters, 1 Receiver, Charging Case)
As an affiliate, we earn on qualifying purchases.
How the Weight Search Worked
Qwen3.5-0.8B is a small open-weight 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 limit, according to their post. The benchmark result addressed that specific behavior in a math setting; it does not cover all kinds of repetitive output or all uses of the model.
The researchers gave Astra a Python environment to inspect and change the 24-layer model’s weights. Their logs say the run was CPU-only, had no internet access and was limited to 300 code runs. Astra could request forward passes to observe outputs, with each pass covering up to 512 problems or 1,024 texts. The final benchmark items were withheld during the search, while practice questions came from GSM8K and MATH.
The post presents eight experiments and calls the findings “very early results.” It is not an arXiv paper and has not been peer-reviewed. The authors list MIT CSAIL affiliations; the post does not list OpenAI as an author or report an OpenAI statement on the work. The benchmark figures and method are claims made in the researchers’ publication and logs, rather than independently reviewed conclusions.
““Fixing loops, however, is not the same as making a model better at math.””
— The four researchers, in “Sparks of Weight-Space Programming”
Questions Beyond the Benchmark
The report does not establish whether the edit improves performance on other math datasets, other kinds of reasoning, or tasks outside math. It also does not show how stable the change is across different prompts or deployments. The available account describes one successful Qwen intervention, alongside experiments with smaller or unsuccessful effects.
It remains unclear how the edit affects other model behaviors, including whether any unintended changes appear outside the benchmark. The post reports that the looping rate fell, not that looping disappeared: one-third of answers still reached the token limit. Independent replication and peer review have not yet been reported in the source material.
Replication Will Test the Claim
The next useful evidence would be independent replication of the published edit and evaluation on additional, clearly described tasks. Researchers and model users can compare the released weight files with the original and test whether the reported reduction in looping persists under different prompts and settings.
The authors’ post does not specify a date for a peer-reviewed paper or further experiments. Until more results are available, the reported score change should be treated as an early benchmark finding about one model and one identified failure mode, not as evidence that AI systems can generally improve other models without training.
Key Questions
What did GPT-6 Astra change?
The researchers say Astra identified a direction associated with looping and modified weights in the MLP down-projection tensors of layers 15 and 19 in Qwen3.5-0.8B.
Did the edit remove the loops?
No. The reported share of MATH-500 answers hitting the token limit fell from 93.0% to 33.1%, so looping remained in a substantial portion of answers.
Did Astra train the model?
The researchers say they did not use gradient descent, backward passes or loss-based weight fitting. Astra did perform an iterative search using code and forward-pass feedback, so the process involved substantial testing.
Has the result been peer-reviewed?
No peer review is reported in the source material. The work appeared as a blog post on November 2, 2026, which the authors describe as preliminary.
Does the result show that Astra can improve any AI model?
No. The report describes one targeted Qwen intervention, while other experiments had limited or unsuccessful results. Broader effectiveness has not been established.
Source: Thorsten Meyer AI
