🔍 Read the full analysis: Can One AI Repair Another? GPT-6 Astra’s Loop Fix Explained on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 2 blog post by four researchers listed with MIT CSAIL reports that GPT-6 Astra identified and edited a direction in two layers of Qwen3.5-0.8B. The model’s MATH-500 score rose from 7.9% to 48.1%, while answers hitting the token limit fell from 93.0% to 33.1%. The result is an early, non-peer-reviewed report: it used extensive iterative testing, and the researchers say reducing loops is not the same as teaching the model mathematics.
Four researchers listed with MIT CSAIL report that OpenAI’s GPT-6 Astra identified and removed a direction in two layers of the open-weight model Qwen3.5-0.8B, reducing repetitive answers and raising its MATH-500 score. In a blog post published November 2, 2026, they say the score rose from 7.9% to 48.1%; the share of answers that ran into the 32,768-token limit fell from 93.0% to 33.1%, meaning the looping was reduced, not eliminated.
The researchers describe the work in “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights.” It is a blog post with published logs, not a peer-reviewed paper, and the authors call the results “very early.” The experiment did not use gradient descent or train the model against a loss function. Instead, Astra examined model outputs and weights, proposed changes, and used test feedback to guide an iterative search.
According to the post, Astra traced the repetition to a direction associated with the MLPs in layers 15 and 19. It subtracted that direction from 64 output columns in each layer, changing 131,072 weights, or 0.016% of the model’s 0.8 billion weights. The researchers say the published Hugging Face version differs from the original in the two layers’ down-projection weight tensors.
The best reported result used a budget of 50 forward passes. The run took 158 turns and 80 Python executions over 20.3 hours, according to the logs. The researchers say Astra could use practice questions but was not shown the final benchmark questions. On MATH-500, runs with 10 and 25 passes scored 12.3% and 14.1%, respectively, compared with 7.9% for the base model. The 50-pass version scored 48.1% and still hit the token limit on 33.1% of answers.
A Repair, Not a New Math Skill
The result points to one possible use of capable models: helping inspect and modify the internal weights of models that are available to researchers or organizations. If the reported method proves repeatable, it could offer a way to address a specific failure without running conventional gradient-descent training. That matters for teams maintaining smaller models, where a narrow defect can make outputs costly or unusable.
But the headline score needs the right interpretation. The source post says the change addressed a known weakness—repetitive loops—and explicitly cautions that “fixing loops” is not the same as making a model better at math. A model that spends fewer answers repeating itself has more opportunity to produce a correct answer, which can raise a benchmark score without showing that it acquired new mathematical reasoning ability. The remaining 33.1% token-limit rate also shows that the defect persisted.
The experiment is relevant only where the weights can be accessed and edited. It does not establish that one closed model can repair another model whose weights are unavailable, nor that the approach is ready for routine deployment. A modified set of weights is, in practical terms, a changed model that needs its own testing for quality, safety and unintended effects.
Products Worth Considering
Logitech MX Mechanical Wireless Illuminated Keyboard, Clicky
Bose SoundLink Flex Bluetooth Speaker (2nd Gen) – Portable Outdoor Speaker with Hi‑Fi Audio, Waterproof and Dustproof, USB‑C, Up to 12 Hours
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
As an affiliate, we earn on qualifying purchases.
How the Weight Edit Was Tested
Qwen3.5-0.8B is an open-weight model whose model card, the researchers say, warns that it is prone to thinking loops. In their initial sample, 85 of 96 practice answers reached the 32,768-token cap. That gave the researchers a concrete failure to investigate rather than a general request to improve an already capable model.
The published logs describe a restricted setup: Astra ran as gpt-6-astra through the Codex command-line tool at its highest reasoning setting, while built-in Codex tools were disabled. A custom harness provided a Python kernel for examining the weights and running the model on a CPU, with no internet access and a maximum of 300 code runs. Practice material included 7,469 GSM8K questions and 7,275 MATH questions. Each forward pass could process up to 512 problems or 1,024 texts, according to the post.
The post covers 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 also connect the Qwen edit to earlier research on removing refusal behavior by deleting a direction from model weights, including a 2024 paper by Arditi and colleagues. That precedent provides context for the technique, but it does not independently validate this math result.
““Fixing loops, however, is not the same as making a model better at math.””
— The researchers, in “Sparks of Weight-Space Programming”
What the Benchmark Does Not Establish
The reported figures come from a single early experiment presented in a blog post, rather than a peer-reviewed study. The source material does not provide independent replication of the Qwen edit or evidence that the same approach will reliably improve other models and tasks. The researchers also report that a loop-free model did not improve in their tests, while experiments on Gemma 3 270M produced no change greater than about a percentage point.
It remains unclear how much of the higher math score reflects better mathematical reasoning and how much follows from preventing answers from consuming their output budget on repetition. The reported drop in token-limit failures is substantial, but one-third of edited-model answers still hit that limit. The available account also does not establish how the edit affects performance across a broader range of prompts, languages or deployment settings.
OpenAI is not listed as an author, and the post does not thank the company or include a statement from it about the experiment. The result is therefore a researcher-reported use of GPT-6 Astra, not an OpenAI announcement or an endorsement of the method by the company.
Replication and Broader Model Tests
The immediate next test is whether independent researchers can reproduce the weight edit and its benchmark results using the published logs and edited model. Further evaluation would need to check whether the lower loop rate holds on tasks beyond the reported benchmarks and whether the edit introduces other changes in model behavior.
The source post does not announce a follow-up study, a release schedule or plans to incorporate the edit into an official Qwen model. For now, the next milestones are independent replication, broader safety and capability testing, and clearer evidence about whether score gains persist outside the tested setup.
Key Questions
What did GPT-6 Astra change in Qwen3.5-0.8B?
According to the researchers’ logs, Astra identified a direction associated with looping in layers 15 and 19 and subtracted it from 64 output columns in each layer. The edit changed 131,072 weights, which the authors say is 0.016% of the model’s weights.
Did the edit eliminate Qwen’s looping?
No. The share of MATH-500 answers that reached the 32,768-token limit fell from 93.0% to 33.1% in the 50-pass run. The researchers report a reduction, not a complete removal.
Did Astra train the model?
The researchers say they did not use gradient descent, backward passes or weight fitting against a loss. Astra instead used an iterative search with model outputs and Python access to the weights. Calling this “without training” does not mean the process required no testing or substantial effort.
Is the result peer-reviewed?
No peer-reviewed paper is identified in the source material. The result appears in a blog post dated November 2, 2026, which the authors describe as early work.
Does the higher score prove that Qwen gained math ability?
No. The authors caution that repairing loops is not the same as improving mathematical ability. The reported benchmark score increased, but the experiment does not establish how much of that gain came from improved reasoning rather than fewer answers being consumed by repetition.
Source: ThorstenMeyerAI.com
