🔍 Read the full analysis: GPT-6 Astra Adjusted Another Model’s Weights—but Didn’t Raise Its IQ 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 reports that GPT-6 Astra edited two weight tensors in the open Qwen3.5-0.8B model without gradient descent. The modified model scored 48.1% on MATH-500, up from 7.9%, but 33.1% of answers still hit the token limit; the work is an early, non-peer-reviewed result.
Researchers say OpenAI’s GPT-6 Astra edited 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 reduced, but did not eliminate, repetitive answers: 33.1% of edited-model answers still reached the benchmark’s 32,768-token limit, compared with 93.0% before the edit.
The result appears in Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights, a blog post published November 2 by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, who are listed with MIT CSAIL. The authors describe their work as “very early results”. It is not a peer-reviewed paper, and the source says it is not on arXiv.
According to the researchers’ logs, Astra identified a direction associated with repetitive outputs in layers 15 and 19, then subtracted it from 64 output columns in each layer. The edit changed 131,072 weights, or 0.016% of the model’s stated 0.8 billion weights. The edited Hugging Face model has changes in the two corresponding down-projection weight tensors, according to the source’s inspection.
The reported improvement followed an iterative search, not a single prompt or a weight update learned through gradient descent. The winning run used 50 forward passes, 80 Python runs and 158 turns over 20.3 hours, with access to practice problems but not the final benchmark questions. The report says MATH-500 scores were 12.3% after 10 passes and 14.1% after 25, before reaching 48.1% after 50. The share of answers hitting the token limit fell from 93.0% to 87.2%, 86.4% and 33.1%, respectively.
The post also reports gains on other benchmarks after the 50-pass run: GSM8K rose from 13.8% to 67.1%, while IFEval rose from 21.0% to 62.3%. These are results reported by the authors; the supplied material does not describe independent replication. The authors also report that, on another model, Gemma 3 270M, none of the edits changed scores by more than about a point.
What the Weight Edit Changes
The finding matters because it suggests a capable model may help diagnose and alter the behavior of an open-weight model without conventional fine-tuning. If the approach proves repeatable, it could offer developers another way to address specific failures in models they can inspect and modify. The reported edit targeted a failure mode already identified in Qwen’s model card: repetitive thinking loops.
But the headline score should not be read as evidence that Astra gave Qwen a broad new mathematical ability. The authors explicitly distinguish fixing loops from making a model better at math. A model that spends less of its output budget repeating itself has more room to reach an answer, which could improve benchmark performance without establishing that its underlying mathematical reasoning became more capable. The remaining 33.1% token-limit rate shows the behavior was reduced rather than removed.
The method also depends on access to weights and substantial testing. The researchers’ use of 50 forward passes and hours of tool-assisted analysis complicates claims that the model was improved “without training” if that phrase suggests no meaningful effort or evaluation. The narrower claim is that the run avoided gradient descent and loss-based weight fitting.
Products Worth Considering
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
LG UltraGear 34GX90SA-W 34-inch WQHD Curved OLED Gaming Monitor
Bose SoundLink Flex Bluetooth Speaker (2nd Gen) – Portable Outdoor Speaker with Hi‑Fi Audio, Waterproof and Dustproof, USB‑C, Up to 12 Hours
As an affiliate, we earn on qualifying purchases.
How the Researchers Tested Astra
The experiment was one of eight in the researchers’ post, which examined whether a language model could inspect or alter neural-network weights. For the Qwen test, their harness gave Astra a Python kernel to work with the model’s weights. The reported setup was CPU-only and offline, with a maximum of 300 code runs. Astra could not run the model directly; it used forward passes to examine outputs, with each pass covering up to 512 problems or 1,024 texts.
The final MATH-500 and other benchmark questions were withheld from Astra, according to the post. The researchers instead supplied practice problems, including 7,469 from GSM8K and 7,275 from MATH. The authors say they prohibited backward passes, gradient descent and fitting weights to a loss. That boundary is central to their description of the approach as weight-space programming rather than conventional model training.
The source says GPT-6 Astra ran through the Codex command-line tool at its highest reasoning setting, with built-in Codex tools disabled. OpenAI’s documentation describes Astra as its most capable model for demanding work, and the source reports that CNBC covered its rollout on September 3, 2026. OpenAI is not listed as an author, and the supplied material says the company did not comment on the experiment.
““Fixing loops, however, is not the same as making a model better at math.””
— The researchers, in “Sparks of Weight-Space Programming”
What the Result Does Not Establish
The findings have not been peer reviewed or independently replicated in the material provided. It is not clear whether the same edit would produce comparable gains across repeated runs, different sampling settings or other model versions. The reported numbers describe particular benchmark tests and do not by themselves show how the edited model performs on broader mathematical tasks.
It also remains unclear how reliably Astra can identify useful edits outside this example, or how much human design and oversight the process needs. The source describes a 20.3-hour run with a defined harness and substantial test access, not an autonomous system that can generally upgrade models. The edit’s durability and possible effects on capabilities beyond the measured benchmarks are not established in the supplied report.
The source reports no statement from OpenAI about Astra’s use. It also does not establish that the result has been reproduced by the model’s developer or an independent research group. The reported remaining loop rate means the approach should not be described as having eliminated repetition.
Replication and Broader Model Tests
The immediate next test is whether independent researchers can reproduce the edit and benchmark changes using the published logs and the edited model’s weights. Replication across runs would help establish whether the result depends on a particular search path or evaluation setup. Testing on additional open-weight models would show whether the method applies beyond this small Qwen model.
Further evaluation would also need to measure more than accuracy and token-limit rates. Researchers could check whether the edit affects other tasks, whether looping falls under different prompts, and whether performance holds on fresh test sets. The supplied source does not give a schedule for a peer-reviewed paper or further experiments, so no next publication or independent validation date is confirmed.
Key Questions
What did GPT-6 Astra change?
According to the researchers, Astra altered 131,072 weights across two layers of Qwen3.5-0.8B to reduce a direction associated with repetitive outputs.
Did the edited model stop looping?
No. The reported share of MATH-500 answers reaching the 32,768-token limit fell from 93.0% to 33.1%, so the behavior remained in a substantial portion of answers.
Was this done without training?
The researchers say they did not use gradient descent, backward passes or loss-based weight fitting. However, the best run involved 50 forward passes, 80 Python runs and 20.3 hours of iterative search, so “without training” does not mean without testing or computational work.
Has the result been independently verified?
The supplied source describes a November 2 blog post by researchers listed with MIT CSAIL and calls the work early. It says the result is not peer reviewed; it provides no independent replication.
Does the score prove Qwen became better at math?
Not by itself. The authors say the edit addressed a looping weakness and caution that fixing loops is not the same as improving mathematical ability. The benchmark gains are reported, but broader reasoning improvements have not been established in the supplied material.
Source: ThorstenMeyerAI.com
