MIT Researchers Say GPT-6 Astra Fixed a Qwen Model by Editing 0.016% of Its Weights

The post, dated November 2 and written by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, drew more than 2,000 points on Hacker News. Within a day a widely read rebuttal argued that its framing misleads readers about how the results were achieved.

What the Researchers Report

In the main experiment, Astra worked inside a Python environment on Qwen3.5-0.8B, a model that often got stuck in reasoning loops. According to the authors, Astra traced the problem to a single direction written by the MLP blocks in layers 15 and 19. It then subtracted that direction from 64 output columns in each layer. That changed 131,072 weights, which the researchers say is 0.016% of the model, and no further training was done.

With a budget of 50 forward passes, the edited model's MATH-500 score rose from 7.9% to 48.1%, GSM8K from 13.8% to 67.1% and IFEval from 21.0% to 62.3%, the post says. The share of MATH-500 answers that hit the 32,768-token limit fell from 93.0% to 33.1%. With 10 or 25 passes the gains were smaller and the loops largely remained.

The post also covers smaller networks. Astra identified which Atari-style game each of four 17,253-weight policy networks had been trained on. It also repaired a stuck control policy by editing 3.5% of its weights. For three MinAtar games it wrote networks from scratch with no training data. Its Breakout network scored 517.4, against 25.3 for a PPO agent trained for 500 million steps. It fell short of PPO on Seaquest, scoring 53.1 against 88.9, and on Asterix, 53.2 against 55.2.

The authors describe the work as "very early results" and set out its limits. Astra could not pick out a sycophantic Gemma 3 270M model from its weights alone. The same editing method barely moved Gemma's MATH-500 score, which went from 6.8% to a peak of 7.7%. "Editing weights could remove an obstacle that hid an ability, but it could not add an ability the model lacked," they write. Logs are on GitHub and the edited Qwen model is on Hugging Face.

The Criticism

Matthew Honnibal, creator of the spaCy NLP library, published a response arguing that the post misleads. He notes that only a few small experiments, on networks of roughly 17,000 to 172,000 weights, involved reading raw weights without tools or running the network. The Qwen result relied on Python code and on running the model over batches of practice problems. "But those latter two aren't just working with raw weights, they write Python programs," he writes.

Honnibal says the Qwen fix is a genuine capability, one that a human machine-learning engineer could carry out and that earlier models would have failed. His objection is to presenting the experiments together as if they showed a single technique and a deep understanding of weights.

Why It Matters

For companies that build on open-weight models, a frontier model that can find and repair specific faults would point towards cheaper ways to debug, audit or adapt models without full retraining. The authors list auditing, unlearning and model merging as future directions but have not tested them. The debate over the post also shows that results like these need careful reading. The gains came from a small model with an obvious failure mode, used task-specific test budgets, and did not carry over to a second model family.