🔍 Read the full analysis: What Changed When GPT-6 Astra Edited A Qwen Model’s Weights? on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
In an experiment published by ThorstenMeyerAI.com on Nov. 5, GPT-6 Astra edited one matrix in the Qwen3-4B-Instruct-2507 model. On a frozen synthetic test of copying special characters into JSON exactly, the edited model answered 70 of 120 cases correctly, up from 49; a separate table-row task remained at zero of 30. The result has not received an independent review, and 50 of 120 final test cases were still wrong.
GPT-6 Astra, operating as a Codex agent, edited one matrix in the open Qwen3-4B-Instruct-2507 model and raised its exact score from 49 to 70 of 120 cases on a frozen test of copying special characters into JSON, according to an experiment report published Nov. 5 by ThorstenMeyerAI.com. The improvement was limited to that task: a separate test of extracting table rows remained at zero of 30, and the experiment has not yet received an independent review.
The task tested whether the model could place short text containing tabs, line breaks, quotation marks or backslashes into a JSON field without changing the text. A response counted as correct only if a JSON parser could decode it and reproduce the original characters exactly. In JSON strings, characters such as a tab must be represented with an escape sequence such as \t, rather than copied as a raw tab.
The report says the edit was a calculated rank-one update to the output projection matrix in the final multilayer perceptron (MLP) layer. It did not use gradient training, and the other 397 stored tensors remained bit-identical to the original model. The final test was frozen before the edited model was evaluated, according to the report.
On that test, the edited model produced 21 more exact answers than the original, with 21 cases fixed and none made incorrect. The result is also reported as a gain of 17.5 percentage points. However, 50 of the 120 final cases were still wrong. The report says clearer prompting alone had not improved performance on an earlier screening set: both the original prompt and a more explicit escaping prompt scored 16 of 32 cases.
Infographic 01 · The measured result
Where the edit helped, and where it did not
Qwen3-4B-Instruct-2507, original versus one edited matrix, on synthetic texts that must survive the trip into JSON without losing a character.
Final test · 120 cases by group
- Texts with tabsthe target failureOriginal19/30Edit30/30
- Line breaksalready perfectOriginal30/30Edit30/30
- Mixed special characterstabs, quotes, backslashesOriginal0/30Edit10/30
- Second table rowstill unsolvedOriginal0/30Edit0/30
Development, validation and controls
- Developmentused to fit the edit · not independentOriginal24/40Edit40/40
- Validationnew cases · 9 wins, 0 lossesOriginal10/40Edit19/40
- Random directioncontrol on validation · same strengthOriginal10/40Edit10/40
- Plain-text tabscontrol · real tabs must surviveOriginal10/10Edit10/10
Exact answer = after normal JSON parsing, the text field matches the source text character for character. Same prompt, temperature 0, seed and output limit for both models; saved and reloaded edit, no hooks, no grammar constraints, no repair. Final test, validation and controls frozen before any independent output of the edited model. Fifty of 120 final answers are still wrong.
Sources: our experiment records, 5 Nov 2026 (Lexassay research, internal): experiment 3 result report and final result by GPT-6 Astra working as a Codex agent (OpenAI). Synthetic test cases, not published.
Powered by Thorsten Meyer AI
A Narrow Gain in Weight Editing
The result is a small but measurable example of an AI system changing another model’s stored weights rather than relying only on a revised prompt or conventional retraining. For developers, the practical question is whether targeted edits can reliably address specific failure modes while leaving unrelated behavior intact. This experiment offers evidence for improvement on one tightly defined behavior, not a general upgrade to Qwen.
The limits matter as much as the gain. The table-row extraction task stayed at 0 of 30, and the JSON test used synthetic cases. The report does not establish whether the change helps with real-world text, other models, or broader software tasks. Nor does a result on one test show that the model is more dependable overall: 50 final answers remained incorrect.
The experiment also follows an earlier unsuccessful attempt by the same agent. In that test, it edited an attention head while trying to improve extraction of fields from synthetic business messages; the original and edited versions each scored 97 of 200 on the frozen final test. The new result therefore marks progress on a selected failure, not a consistent record of successful weight edits.
Popular right now
Breville Bambino Espresso Machine, Brushed Stainless Steel
TCL 43-Inch 4K UHD Smart LED Roku TV (43S435, 2021 Model)
De’Longhi Classic Matte Black Espresso Machine with Milk Frother
As an affiliate, we earn on qualifying purchases.
How the Test Followed Earlier Trials
The report places this work in a week-long series about AI systems inspecting and changing neural-network weights. In the first experiment, Anthropic’s Claude edited the local model Ornith to make it finish more legal answers. The number of completed answers rose from four of 15 to 12 of 15, but the legal score remained zero, according to the site. GPT-6 Astra later reviewed that work and identified reported problems including a resume bug, a command-line fault and side effects from weight conversion.
Before the Qwen JSON experiment, GPT-6 Astra had also tried editing that model for extraction of five fields from synthetic business messages. Both versions scored 97 of 200 on the frozen test. The report says the two versions produced the same 47 invalid answers, mostly involving raw tab characters in JSON strings. It identifies that earlier failure as the lead for the later, more focused test. The third experiment was therefore targeted at a known error, rather than an exhaustive search for model weaknesses.
ThorstenMeyerAI.com says the site owner assigned the task on Nov. 5 and that GPT-6 Astra, with its reasoning effort set to extra high, planned the experiment, inspected the weights, generated tests, computed the edit and produced the report. Claude did not take part in the experiment; according to the source, a Claude agent later researched and drafted the article and checked the reported figures against stored results.
“The starting hint for experiment 3 “came from the previous attempt.””
— ThorstenMeyerAI.com experiment report
Independent Checks Still Pending
The result is reported by the site that commissioned and documented the experiment, and the source says no second AI has independently reviewed it. The reported frozen test and stored results provide details about how the score was measured, but the supplied material does not include an external replication or review of the method and outcome.
It is also unclear how well the edit would generalize beyond the synthetic cases, whether it would affect other tasks not tested, or whether similar gains would hold across other models and test sets. The report does not provide evidence that the edit improves general JSON handling broadly. Its stated finding is restricted to the specified exact-copying test; performance on the separate table-row task did not improve.
Replication and Broader Testing
The next meaningful checks would be an independent review of the experiment records and a replication on fresh, held-out cases, ideally including examples closer to real software inputs. Testing should also check for unintended effects on other model behaviors, since the report’s unchanged tensor count does not by itself establish that the edited model behaves identically elsewhere.
ThorstenMeyerAI.com has not announced a date for an independent review or a follow-up test in the supplied material. Until those checks are reported, the current finding remains a single experiment on a narrow synthetic task, with a measurable score gain and substantial remaining errors.
Key Questions
What did GPT-6 Astra change?
According to ThorstenMeyerAI.com, the agent applied a calculated rank-one update to the output projection matrix in the final MLP layer of Qwen3-4B-Instruct-2507. The report says the other 397 stored tensors were bit-identical to the original.
What improved in the test?
The model had to copy text containing special characters into JSON without losing or changing any character. Its exact score rose from 49 of 120 to 70 of 120 on the frozen final test, a reported gain of 17.5 percentage points.
Did the edit improve every task?
No. The separate test of extracting a table row remained at 0 of 30. The report also says 50 of the 120 final JSON cases were still incorrect after the edit.
Has the result been independently verified?
Not yet, according to the source material. ThorstenMeyerAI.com says the experiment has not been checked by a second AI; the supplied report also does not describe an external replication.
Was this ordinary model training?
The report says the experiment did not use gradient training. It describes a calculated change to one matrix in the model’s weights, designed around a specific failure involving JSON escaping.
Source: ThorstenMeyerAI.com
