🔍 Read the full analysis: AI Repairs AI: GPT-6 Astra Edited Another Model’s Weights. A Loop Fix, Not 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 listed with MIT CSAIL reports that GPT-6 Astra edited 131,072 weights in two layers of Qwen3.5-0.8B without gradient descent. The model’s MATH-500 score rose from 7.9% to 48.1%, but 33.1% of answers still hit the token limit, and the result has not been peer-reviewed.
Four researchers listed with MIT CSAIL report that OpenAI’s GPT-6 Astra edited the weights of the open model Qwen3.5-0.8B, reducing repetitive reasoning loops and raising its score on the MATH-500 benchmark from 7.9% to 48.1%. The November 2 blog post describes an iterative, test-guided search rather than gradient-descent training; the result is an early report, not a peer-reviewed finding.
The researchers say Astra identified a direction associated with looping responses and subtracted it from output columns in two model layers. The change affected 131,072 weights, or 0.016% of the model’s 0.8 billion parameters. The edited model posted on Hugging Face differs from the original in the two specified weight tensors, according to the source account’s inspection of the files and published hashes.
The reported MATH-500 score rose as the researchers increased the number of model-evaluation passes: 12.3% after 10 passes, 14.1% after 25 and 48.1% after 50, compared with 7.9% for the base model. The share of answers reaching the 32,768-token limit fell from 93.0% to 33.1% in the 50-pass run. That means looping was reduced, not eliminated.
The final run used 158 turns, 80 Python runs and 50 forward passes, and took 20.3 hours, according to the published logs described in the source. The researchers prohibited gradient descent, backward passes and fitting weights to a loss. Astra was given practice problems, not the final benchmark questions. Results also improved on GSM8K and IFEval in the 50-pass run, while the smaller 10- and 25-pass budgets produced much more modest MATH-500 scores.
A Targeted Repair, Not New Math Ability
The result matters as a demonstration of a possible way to inspect and patch open-weight models using another model as a programming assistant. If the method proves reliable, developers could use it to address a specific failure mode without running conventional gradient-descent training. The reported edit is small relative to the model, while the score change is large enough to merit independent replication.
But the headline gain should not be read as evidence that Astra gave Qwen a general new mathematical ability. The authors’ interpretation, as reported in the source, is that the edit addressed a known weakness: Qwen3.5-0.8B was prone to thinking loops, and those loops often prevented it from producing a final answer. The authors put the distinction plainly: “Fixing loops, however, is not the same as making a model better at math.” Removing a failure mode can raise a benchmark score without establishing broader reasoning gains.
The technique also has trade-offs for organizations that use open models. An edited weight file is a modified model, so users need to track which version they deploy and test whether a targeted patch changes other behaviors. The source notes prior work using single-direction weight edits to remove refusals from open chat models, illustrating that such changes may affect safety behavior as well as performance.
Products Worth Considering
Anker SOLIX C1000 Gen 2 Portable Power Station
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Waterproof, Dustproof, USB‑C, Up to 12 hr Battery, Black
EcoFlow Delta 3 Ultra 3072Wh Portable Power Station – 3600W AC Output
As an affiliate, we earn on qualifying purchases.
How the Weight Edit Was Tested
The report, titled Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights, describes eight experiments. Its authors are Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab; Takehi and Teoh are identified as equal contributors. The post lists the authors with MIT CSAIL and calls the work “very early results.” It is not a peer-reviewed paper, and the source says it was not on arXiv at the time of writing.
For the Qwen experiment, Astra ran through the Codex command-line tool at the highest reasoning setting, with built-in tools disabled. A researcher-built setup gave it a CPU-based Python environment to examine the model’s weights and run code, without internet access. The team allowed up to 300 code runs and supplied practice material from GSM8K and MATH. Astra could not directly run Qwen: each forward pass used to observe its outputs was a limited resource, covering up to 512 problems or 1,024 texts.
The source reports that Qwen’s model card already warned of a tendency to enter thinking loops. In the researchers’ initial practice sample, 85 of 96 answers ran to the token limit. The final test scores came from benchmarks withheld from Astra, according to the report. OpenAI’s documentation describes GPT-6 Astra as its most capable model for demanding work, but the source says OpenAI was not an author of the post and had not commented on this experiment.
““Very early results.””
— The researchers, describing the report
Replication and Broader Effects Remain Open
The result remains a report from a blog post rather than a peer-reviewed study. The source provides logs and identifies the edited tensors, but independent replication has not been established. It is not yet clear whether the same procedure will work consistently across model families, tasks or starting conditions, or how stable the improvement will be under different evaluation settings.
The remaining 33.1% token-limit rate shows that the looping behavior was not fully removed. The reported benchmark gains also do not, by themselves, establish that the model learned more general mathematical reasoning. The source describes one experiment in which Astra could not identify which of two fine-tuned Gemma models had been trained to flatter users, and reports that edits to a much smaller Gemma 3 270M model moved scores by no more than about a point. Those results suggest the approach’s effects may depend on the model and problem.
It is also unclear whether the edits changed behaviors beyond those measured in the reported benchmarks, including safety-related responses. The source reports no statement from OpenAI on the experiment, and does not establish that OpenAI participated in or endorsed the work.
Independent Tests Are the Next Check
The immediate next step is for other researchers to reproduce the weight edit and benchmark results, including the reduction in token-limit failures. Further tests could check whether the edited model retains its gains on new math problems, whether results hold across sampling settings and model sizes, and whether the patch alters unrelated capabilities or safety behavior.
The source does not provide a schedule for a peer-reviewed paper or further experiments. Until more evidence is available, the reported result is best treated as a specific early demonstration: one model used to edit another model’s weights, with a substantial benchmark improvement in a task where the original frequently looped, but with important limits still visible.
Key Questions
What did GPT-6 Astra change?
According to the researchers’ post, Astra subtracted a direction associated with looping from output columns in two layers of Qwen3.5-0.8B, changing 131,072 weights.
Did the researchers train Qwen with gradient descent?
No. The report says gradient descent, backward passes and fitting weights to a loss were prohibited. Astra used an iterative search with Python code and model-evaluation passes, so “without training” does not mean the work involved no testing or effort.
Did the edit eliminate Qwen’s loops?
No. The reported share of MATH-500 answers that reached the 32,768-token limit fell from 93.0% to 33.1% in the 50-pass run, but a substantial fraction still hit the limit.
Is the result peer-reviewed?
No. The source describes the work as a November 2, 2026 blog post with published logs and calls it an early result. It was not peer-reviewed at the time covered by the report.
Does the higher score prove Qwen gained general math ability?
No. The authors say that fixing loops is not the same as making a model better at math. The score increase is consistent with fewer answers being lost to repetitive outputs, but broader reasoning gains have not been established.
Source: ThorstenMeyerAI.com
