🔍 Read the full analysis: Could GPT-6 Astra Help Another Model Escape A Loop? on ThorstenMeyerAI.com
Get the little things that make your day 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 helped edit two layers of the open model Qwen3.5-0.8B without gradient descent. The model’s MATH-500 score rose from 7.9% to 48.1%, while answers hitting the 32,768-token limit fell from 93.0% to 33.1%. The work is an early, non-peer-reviewed report, and the loops remained common after the edit.
Four researchers affiliated with MIT CSAIL report that OpenAI’s GPT-6 Astra helped edit the weights of the open model Qwen3.5-0.8B, raising its MATH-500 score from 7.9% to 48.1% without gradient descent. The edit also reduced the share of answers that ran into the model’s 32,768-token limit from 93.0% to 33.1%, but did not eliminate the looping behavior.
The results appear in a blog post titled “Sparks of Weight-Space Programming,” published November 2 by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. The authors describe the findings as “very early results”. The post has not been peer-reviewed and is not on arXiv, according to the supplied source material.
In the reported experiment, Astra used a Python environment to inspect and modify the 24-layer model’s weights. The researchers prohibited gradient descent, backward passes and fitting weights to a loss. They say Astra identified a direction associated with repetitive outputs and subtracted it from output columns in the MLP layers numbered 15 and 19. The edit affected 131,072 weights, or 0.016% of the model’s 0.8 billion weights. The edited files on Hugging Face reportedly show changes only to those layers’ down-projection tensors.
The final MATH-500 benchmark was not shown to Astra during the search, according to the post. The team tested three search budgets. With 10 and 25 forward passes, MATH-500 scores reached 12.3% and 14.1%; with 50 passes, the score reached 48.1%. The token-limit rate fell to 87.2%, 86.4% and 33.1%, respectively. The researchers also report that the 50-pass version scored 67.1% on GSM8K, compared with 13.8% for the base model, and 62.3% on IFEval, compared with 21.0%.
A Targeted Fix, Not a New Math Skill
The experiment matters because it suggests that a model can help diagnose and alter another model’s behavior by working directly with its weights. For operators of open-weight models, this could offer a way to address a specific failure without running conventional gradient-based training. The reported edit was small relative to the model, though the search that found it was not effortless.
The scores need to be read alongside the failure rate. Even after the strongest edit, 33.1% of MATH-500 answers still reached the token cap. The authors themselves caution that “fixing loops” is not the same as making a model better at math. The benchmark gains are reported results from this setup, not evidence that Astra broadly improves models or that the method will work on other systems.
The method also has implications for model governance. The source post points to earlier work showing that removing a direction in a model’s weights can affect refusal behavior. An edit that changes a model’s output may be useful, but it can also alter safeguards or other behavior. A patch to open weights creates a modified model that should be checked and documented separately from its original version.
Products Worth Considering
Levoit EverestAir-P Smart Air Purifier for Large Rooms
Google Nest Learning Thermostat (4th Gen) with Nest Temperature Sensor (2nd Gen), Obsidian
Ring Alarm Pro 14-Piece Home Security Kit with Built-In eero Wi-Fi 6 Router
As an affiliate, we earn on qualifying purchases.
How the Weight Search Worked
Qwen3.5-0.8B is a small open model whose model card warns that it can fall into thinking loops, according to the source. In the researchers’ initial practice run, 85 of 96 answers reached the 32,768-token limit. That gave Astra examples of the behavior the team wanted to reduce.
The researchers’ harness gave Astra access to the model’s weights and a Python kernel, but not internet access or the ability to run the model freely. Each forward pass let it inspect outputs for up to 512 problems or up to 1,024 texts. The search used practice questions from GSM8K and MATH, while the final benchmark questions were withheld. The 50-pass run took a reported 20.3 hours and involved 158 turns and 80 Python runs; the source says the system allowed at most 300 code runs.
The post covers eight experiments, not all of them successful. In one test, Astra could not determine from weights alone which of two fine-tuned Gemma models had been trained to flatter a user. The reported math result is therefore one demonstration among mixed experiments, not a general proof that models can reliably inspect and rewrite other models.
““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 blog post
Questions About Repeatability and Scope
The results come from a non-peer-reviewed blog post, and the supplied material does not include an independent replication. It is not clear whether the same direction-based edit would work reliably across other model sizes, architectures, tasks or prompt settings. The report describes one strong result on Qwen3.5-0.8B, alongside weaker results at smaller search budgets and a separate experiment that failed.
The reported scores also do not establish why all the changes occurred. The authors associate the edit with reduced looping and improved benchmark performance, but the information provided does not show whether the gains persist across different sampling settings or longer evaluations. The post’s statement that the final test questions were withheld is relevant to test contamination, but independent testing would be needed to assess how robust the results are.
OpenAI is not listed as an author of the post, and the source material says it found no OpenAI statement about the experiment. The relationship, if any, between the researchers and OpenAI beyond their use of GPT-6 Astra is not specified here. The precise technical setup and benchmark reporting can be checked against the published logs and model files, but broader conclusions remain unsettled.
Independent Tests Are the Next Step
The immediate next step is independent evaluation of the edited model against the original, using additional math questions and varied sampling settings. That would help show whether the reported gains hold beyond the researchers’ test setup and whether loop reduction remains consistent. The published logs and Hugging Face model files provide material for other researchers to inspect and reproduce the edit.
Further work would also need to measure side effects, including performance on tasks beyond math and changes to safety behavior. Until such results are available, the finding is best treated as an early demonstration of weight editing with model assistance, not a validated general method for improving AI systems. No follow-up date or independent replication is specified in the source material.
Key Questions
Did GPT-6 Astra train Qwen3.5-0.8B?
Not through gradient descent, according to the researchers. Astra searched and edited the model’s weights using a Python harness, but the reported process involved repeated tests and feedback; “without training” does not mean without a substantial search effort.
Did the edit stop Qwen’s looping?
No. The reported share of MATH-500 answers reaching the 32,768-token limit fell from 93.0% to 33.1% in the 50-pass run. Loops remained in about one-third of answers under that measure.
How much did the MATH-500 score change?
The researchers report a rise from 7.9% to 48.1% for the 50-pass version. Smaller search budgets produced lower scores: 12.3% after 10 passes and 14.1% after 25 passes.
Has the result been peer-reviewed?
No. The supplied source describes a blog post published November 2, 2026, calls the results early, and says the work is not peer-reviewed and is not on arXiv.
Did OpenAI conduct or endorse the experiment?
The researchers used GPT-6 Astra, but the source says OpenAI is not an author of the post and reports no OpenAI statement about the experiment. It does not establish that OpenAI conducted or endorsed the work.
Source: ThorstenMeyerAI.com
