🔍 Read the full analysis: A Loop Fix, Not A Smarter Model: GPT-6 Astra’s AI Edit on ThorstenMeyerAI.com
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
TL;DR
A November 2 blog post by four researchers affiliated with MIT CSAIL reports that GPT-6 Astra improved Qwen3.5-0.8B’s MATH-500 score from 7.9% to 48.1% by editing a small part of its weights, without gradient descent. The change reduced but did not eliminate answers that ran into the 32,768-token limit; the results are early, unreviewed, and required an extensive iterative search.
Four researchers affiliated with MIT CSAIL say they used OpenAI’s GPT-6 Astra to edit two weight tensors in the open model Qwen3.5-0.8B, raising its reported MATH-500 score from 7.9% to 48.1% without gradient descent. The edit targeted repetitive reasoning that often exhausted the model’s 32,768-token limit, but did not remove the problem: 33.1% of edited-model answers still hit that limit. The findings appeared in a November 2 blog post, not a peer-reviewed paper.
The post, “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights,” describes a tool-assisted search in which Astra examined model behavior and proposed changes to the weights. The researchers say it traced looping to a direction associated with the MLPs in layers 15 and 19, then subtracted that direction from 64 output columns in each layer. They report that 131,072 weights changed, or 0.016% of the model’s 0.8 billion weights. The edited model hosted on Hugging Face has changes to the down-projection tensors in those two layers.
The work did not use gradient descent, backward passes or fitting weights to a loss, according to the researchers’ published account. But “without training” does not mean a single prompt or a hands-off edit. The strongest run took 158 turns and 80 Python runs, with 50 forward passes over as many as 512 problems per pass, over 20.3 hours. Astra had access to practice problems, but not the final benchmark questions. The setup limited its code execution and ran the harness on a CPU without internet access.
The reported results rose with the number of forward passes. On MATH-500, runs with 10 and 25 passes scored 12.3% and 14.1%, respectively; the 50-pass run scored 48.1%. The share of answers reaching the token cap fell from 93.0% for the original model to 87.2%, 86.4% and 33.1%. On GSM8K, the reported score increased from 13.8% to 67.1% in the 50-pass run. On IFEval, it increased from 21.0% to 62.3%. These figures come from the researchers’ post and logs; the supplied material does not establish independent replication.
A Weight Edit Can Change Model Behavior
The result points to one possible use for capable AI systems: helping inspect and modify open-weight models, whose parameters can be accessed and changed. Rather than retraining the full model, the researchers report that Astra found a relatively small intervention linked to a specific failure mode. If similar techniques prove reliable, developers may be able to test targeted repairs with less compute than conventional model training requires.
That is not the same as showing that Astra broadly improves a model’s reasoning. The reported math gains came alongside a reduction in one known failure—repetitive output—and the authors themselves caution that fixing loops is not the same as making a model better at math. The score changes could reflect both fewer answers wasted in loops and other effects of the edit. The post does not establish which explanation accounts for how much of the increase.
The method also has safety and governance implications. A change to model weights can alter behavior without changing the surrounding software, and the researchers cite prior work showing that weight-direction edits can affect refusal behavior. That makes verification and version tracking important: a patched model is a distinct model variant, not merely the original model running with a different prompt.
Popular right now
Breville Barista Express Espresso Machine, Black Sesame
Ninja Flip Toaster Oven & Air Fryer, Large Capacity, Stainless, SP151
Ninja Luxe Café Mini Plus Espresso and Drip Coffee Machine with AutoFroth, Stainless Steel (ES351)
As an affiliate, we earn on qualifying purchases.
The Model’s Known Looping Problem
Qwen3.5-0.8B is a small open-weight model. Its model card warns that it can fall into thinking loops, according to the source account. In the researchers’ initial practice sample, 85 of 96 answers reached the 32,768-token limit. On MATH-500, the original model’s reported score was 7.9%, with 93.0% of answers reaching the cap.
The blog post covers eight experiments involving attempts to infer or alter information in model weights, including one unsuccessful attempt to distinguish which of two fine-tuned Gemma models had been trained to flatter users. The Qwen math result is one experiment, not evidence that every weight-editing task works. The authors describe the work as “very early results” and say its capabilities are exciting but limited.
GPT-6 Astra is described in OpenAI developer documentation as its most capable model for demanding work. The source says CNBC reported its rollout on September 3, 2026. OpenAI is not listed as an author of the research post, and the supplied material reports no OpenAI statement about the experiment. The authors are Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab; Takehi and Teoh are listed as equal contributors.
““Fixing loops, however, is not the same as making a model better at math.””
— Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, in the research blog post
Replication and Generality Remain Open
The reported benchmark results come from a blog post and accompanying logs. The work is not peer-reviewed, and the supplied account does not cite an independent reproduction. It is not yet clear whether the same approach will work on other models, different failure modes or evaluation sets, or whether the gains will hold under independent testing.
The source material does not provide enough detail to determine how much of the score increase comes specifically from reduced looping, rather than other behavioral changes caused by the edit. The edited model still hit the token limit in 33.1% of MATH-500 answers, so the loop was reduced, not eliminated. The account also says a model without loops did not improve in the reported trials, but does not establish a general rule about which models can benefit.
OpenAI’s role is also limited in the available record: Astra was used through the researchers’ harness, but OpenAI is not an author of the post and no company comment is reported. The post’s results should not be treated as an OpenAI-validated finding or as evidence that Astra can routinely edit models successfully.
Independent Tests Are the Next Check
The immediate next step is independent evaluation of the edited weights against the original model, using the same sampling settings and clearly documented test conditions. Replications could check whether the score gains persist across different math problems and whether the remaining token-limit failures fall further. Tests on additional models would show whether this was a model-specific repair or a more general weight-editing technique.
For users and developers, the available edited weights and published hashes offer a starting point for inspecting the change. Any deployment should keep the modified version identifiable and evaluate its behavior beyond the benchmarks reported in the post. The authors’ work is presented as preliminary; no follow-up study or release date is specified in the supplied material.
Key Questions
What did GPT-6 Astra change in Qwen3.5-0.8B?
The researchers say Astra identified a direction associated with repetitive output and subtracted it from weight columns in the model’s layers 15 and 19. They report changes to 131,072 weights, or 0.016% of the model’s total weights.
Did the edit eliminate Qwen’s reasoning loops?
No. The share of MATH-500 answers reaching the 32,768-token limit fell from 93.0% to 33.1% in the reported 50-pass run. A substantial portion still reached the limit.
Was the model improved without any training or work?
The researchers say they did not use gradient descent or fit the weights to a loss. However, Astra’s winning run involved 158 turns, 80 Python runs and 50 forward passes over 20.3 hours, so the process was an iterative search rather than a one-step edit.
Are the findings peer-reviewed?
No peer-reviewed paper is cited in the source material. The findings were published in a blog post on November 2, 2026, which the authors describe as very early results.
Did OpenAI conduct or endorse the experiment?
The researchers used GPT-6 Astra through their own harness, but OpenAI is not listed as an author of the post. The source material reports no OpenAI statement about the experiment.
Source: ThorstenMeyerAI.com
