🔍 Read the full analysis: GPT-6 Astra And The AI Loop Fix That Wasn’t An IQ Jump 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 edited two layers of the open Qwen3.5-0.8B model without gradient descent. The model’s reported MATH-500 score rose from 7.9% to 48.1%, but 33.1% of edited-model answers still hit the token limit. The result is an early, non-peer-reviewed report of a targeted repair, not evidence of a general jump in intelligence.
Four researchers affiliated with MIT CSAIL say they used OpenAI’s GPT-6 Astra to edit two layers of the open-weight model Qwen3.5-0.8B, raising its reported MATH-500 score from 7.9% to 48.1% without gradient descent. The change also reduced the share of answers that ran into the benchmark’s 32,768-token limit, from 93.0% to 33.1%—meaning the looping behavior was reduced, not eliminated.
The report, titled Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights, appeared on November 2 as a blog post by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. The authors describe the work as very early results. It is not a peer-reviewed paper, and the source says it has not been posted on arXiv. OpenAI is not listed as an author, and the post does not report an OpenAI statement about the experiment.
According to the researchers’ logs, Astra identified a direction associated with repetitive outputs in the model’s layers 15 and 19, then subtracted it from output columns in those layers. The edited Hugging Face model reportedly differs from the original in the two corresponding down-projection weight tensors. The change affected 131,072 weights, or 0.016% of the 0.8-billion-parameter model, according to the post.
The reported MATH-500 results depended on the amount of search effort. With 10 and 25 forward passes, the score reached 12.3% and 14.1%; at 50 passes it reached 48.1%. The 50-pass run also involved 158 turns, 80 Python runs and 20.3 hours, according to the published logs. For that run, 33.1% of answers still hit the token limit. The authors also report improved results on GSM8K and IFEval, but these figures come from the same early report and should not be treated as independent confirmation.
A Targeted Repair, Not New Intelligence
The result matters because it suggests that an AI system may be able to help diagnose and alter a model’s internal behavior when its weights are accessible. In this experiment, the reported change was narrow: researchers targeted repetitive output that interfered with completing math problems. The score gains are notable within the reported test setup, but they do not establish a broad increase in reasoning ability or prove that the edited model is more capable across tasks.
The distinction between repairing a failure mode and adding a new skill is central. The source notes that Qwen’s model card already warns of thinking loops, and the researchers themselves caution that fixing loops is not the same as making a model better at math. The test results are consistent with a model becoming more able to finish answers; they do not, by themselves, show that the underlying mathematical reasoning improved independently of that repair.
For developers and organizations, the method also has a practical limit: it relies on access to editable model weights. A successful patch creates a modified model that must be evaluated and managed as its own version. The report’s findings do not establish that the technique will work on closed models or reliably transfer to other model families.
Products Worth Considering
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Waterproof, Dustproof, USB‑C, Up to 12 hr Battery, Black
LG UltraGear 34GX90SA-W 34-inch WQHD Curved OLED Gaming Monitor
Apple Watch Series 8 GPS, 41mm, Midnight Aluminum Case with Midnight Sport Band (S/M)
As an affiliate, we earn on qualifying purchases.
How the Weight Edit Was Tested
The researchers prohibited gradient descent and backward passes, but “without training” does not mean the model was changed without substantial testing or feedback. Astra could inspect and edit the weights through a researcher-built Python environment, while the harness allowed a limited number of code runs and forward passes. The logs describe 50 forward passes in the highest-budget run; each could evaluate up to 512 problems or process up to 1,024 texts. The benchmarks used for final scoring were withheld from Astra, according to the post.
Instead, the system received practice material from GSM8K and MATH, separate from the final benchmark sets. Early tests showed extensive looping: 85 of the first 96 practice answers reached the token limit, the authors report. They say Astra compared looping and non-looping outputs and used those observations to identify the direction it then removed from the weights.
The post describes eight experiments, not all of which succeeded. In one, Astra could not determine from the weights alone which of two fine-tuned Gemma models had been trained to flatter users. The contrast helps frame the Qwen result as a specific demonstration, rather than evidence that a model can generally infer or rewrite any behavior from weights.
“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 report
What the Scores Cannot Yet Show
The findings are based on a single early report rather than peer-reviewed research. The source material does not provide independent replications, and it remains unclear whether the same kind of edit will work reliably on other models, tasks or looping patterns. Nor do the reported benchmark scores settle whether the model’s reasoning improved, or whether it mainly became less likely to spend its output budget repeating itself.
The loops also remain measurable after editing: 33.1% of MATH-500 answers still reached the 32,768-token cap in the best-reported run. The source gives benchmark scores and token-limit rates, but these results alone do not establish performance in real-world use or describe the impact of the edit on other behaviors. The source also does not report an OpenAI response or indicate that Astra’s developer participated in the work.
Replication and Broader Model Tests
The immediate next step is independent testing of the edited model and the researchers’ logs, including checks of the same benchmarks and other tasks. Further work would need to establish whether the change preserves unrelated capabilities, whether it can be applied to other open-weight models, and how often a similar search succeeds under smaller budgets. The reported 10- and 25-pass results—12.3% and 14.1% on MATH-500—show that the strongest reported gain required the largest tested budget.
For now, readers can inspect the published logs and the edited model on Hugging Face, as described in the source report. No follow-up paper, outside replication or OpenAI comment is included in the supplied material, so the method’s broader reliability and status remain open questions.
Key Questions
Did GPT-6 Astra train Qwen3.5-0.8B?
The researchers say they did not use gradient descent or backward passes. Astra instead used test feedback and Python access to edit selected weights. Calling this “without training” needs that qualification: the process involved an iterative search, forward passes and substantial testing.
Did the edit remove Qwen’s looping behavior?
No. The reported share of MATH-500 answers hitting the 32,768-token limit fell from 93.0% to 33.1%, but loops remained in about one-third of answers under the reported test conditions.
Did Qwen gain general mathematical intelligence?
The report does not establish a general intelligence gain. It records a higher MATH-500 score after a targeted edit intended to reduce repetitive output. The authors caution that fixing loops is not the same as making a model better at math.
Has the result been peer-reviewed?
No peer review is reported in the source material. The work was published as a November 2, 2026 blog post, which the authors describe as very early results.
Can the method be used on any AI model?
That has not been shown. The experiment used an open-weight model whose parameters could be edited. The report does not establish that the method works on closed models or reliably applies across different model families.
Source: ThorstenMeyerAI.com
