Hand-written neural net weights
Lead: The MIT CSAIL result is a small fix to a model that was already trained. It is not a way to build a model by hand. I can do the same kind of fix on a small open model on this Mac. To make a big model from nothing, you still have to train it.
What MIT reported (secondary coverage, paper not read)
- Editor model
- GPT-6 Astra
- Target model
- Qwen3.5-0.8B
- Weights changed
- 131,072 of 0.8 billion (0.016%)
- MATH-500
- 7.9% → 48.1% (replication: 51.0%)
- Answers that hit the token limit
- 93.0% → 33.1%
- Gemma 3 270M
- no clear gain (it rarely loops)
Reports say the edit removed one "loop direction" from 64 columns in two MLP layers. A co-author said the prompt also told Astra to check for repetitive loops. Critics call the method standard ML engineering. The study has not been peer reviewed.
How the loop fix works
This is directional ablation. Researchers used it before, for example to remove the refusal direction from chat models.
Can I do it?
Yes, on a small model. First, download Qwen 0.8B (about 1.6 GB) and run it on math problems. Then find the loop direction, edit the weights, and compare scores. A full MATH-500 run on a laptop is slow, so a first test uses about 50 problems.
Which models gain, and when to train instead
No public results exist for models other than Qwen and Gemma. This table is my inference from how the fix works. It is not measured data.
| Model state | Expected gain from the loop fix |
|---|---|
| Loops on most answers (small reasoning models, for example Qwen3.5-0.8B) | Large |
| Loops sometimes | Small to medium |
| Rarely loops (Gemma 3 270M, most large models) | None |
| Fails because it does not know the math | None. An edit adds no skill. |
Run the model on 50 problems. Count the answers that hit the token limit. A high count means the fix can help. A low count means it cannot.
| Goal | Best method |
|---|---|
| Build a capable model from nothing | Train it. |
| Fix one narrow fault in a trained model | Edit the weights. |
| Add a broad new skill | Fine-tune it. |