GPT-6 Astra fixed a stuck loop in Qwen and everyone heard self-improvement

Before GPT-6 Astra touched it, Qwen3.5-0.8B ran 85 of its first 96 practice answers all the way to the 32,768-token limit. On MATH-500, 93% of its answers hit that cap. A model that never stops talking cannot score, so it scored 7.9%. Astra traced the repetition to one direction written by the MLPs in layers 15 and 19, subtracted it from 64 output columns in each, and changed 131,072 weights, or 0.016% of the model. The share of capped answers fell to 33% and MATH-500 rose to 48.1%. GSM8K went from 13.8% to 67.1% and IFEval from 21.0% to 62.3%.

I went into the post by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab at MIT CSAIL expecting a claim that an LLM had made a smaller model smarter by reasoning about its weights. What surprised me is that the authors say plainly that it did nothing of the kind. Their own write-up concedes that fixing loops lets the model finish answers it could already work out. Then they ran the same procedure on Gemma 3 270M, which loops on only 8.5% of answers, and no edit moved MATH-500 by more than about a point. Astra found ordinary arithmetic and algebra mistakes with no single component to switch off, and reported its own tiny gain as not established.

So the Qwen result is a bug fix, and a familiar style of bug fix at that. Deleting one direction from a model’s weights to remove a behaviour is an existing interpretability technique, and the authors cite it. The new part is that a general-purpose model found the direction itself with a notebook and 50 batches of forward passes. That is a real and interesting finding. It is also a long way from the recursive self-improvement some readers immediately saw in it. Astra edited Qwen, not itself, and nothing in the logs describes a model rewriting its own weights, and the post files self-improvement under speculative future uses.

The announcement on X drew 1.2 million views, and Ethan Mollick’s reply set the right terms: “Waiting for replication but this is a pretty big deal.” The replication part is where I think the coverage has been lazy. Look at the budget curve. With 10 forward passes MATH-500 reached 12.3% and 87.2% of answers still hit the cap. With 25 passes it was 14.1% and 86.4%. With 50 it jumped to 48.1%. That is a cliff. Either the search needs a threshold of evidence before the looping direction becomes visible, or the 50-pass run found it and the others did not. I could not find variance across repeated runs for this experiment in the post, and a cliff from one trajectory per budget tells you very little about how often the method works.

The account @1a3orn complained that Takehi’s announcement thread left out the Gemma failure. The sycophancy experiment got similar treatment in the retelling: from weights alone Astra failed in all four attempts and only succeeded once it was allowed to watch attention patterns on text it wrote. These are honest results, sitting in the post, and they define the method’s current reach much better than the headline numbers do.

The obvious reply is that this is the most reproducible AI claim of the month. The authors released every prompt, every line of code Astra ran, and exact diffs of the Experiment 8 edits against the public base weights. Ahsen Khaliq at Hugging Face rebuilt the edited model from the logs and re-ran its evaluation. I agree that this openness deserves credit, and it means I trust that the checkpoint scores what they say. But re-scoring a checkpoint confirms the checkpoint. It says nothing about whether Astra reliably finds such edits, how it compares with a human using standard ablation tools, or whether a free decoding change gets most of the way. The authors say they are still building a benchmark and harness, which is an admission that the measurement does not exist yet.

That last comparison is one any reader with a GPU can run this week. Qwen’s own model card warns about these loops. Take the base Qwen3.5-0.8B, sample MATH-500 with whatever anti-repetition settings its model card recommends, and count how many answers still hit 32,768 tokens and how many come back correct. If the unedited model with sensible decoding lands anywhere near 48.1%, the headline result is a sampling problem that Astra solved the hard way, and the 0.016% figure should be read as a cleaner patch for a bug that already had a workaround.

Prompted by Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights, rikiyat.github.io.