Result
Table with columns: Model, MATH-500 accuracy| Model | MATH-500 accuracy |
|---|
| Qwen/Qwen3.5-0.8B | 6.5 % |
| This model | 32.0 % |
Evaluation settings: all 500 MATH-500 problems, 4 samples per problem (2,000 samples), thinking mode on with the model's chat template, the prompt suffix "Please reason step by step, and put your final answer within \boxed{}.", up to 32,768 new tokens, temperature 0.6, top_p 0.95, top_k 20, presence penalty 1.5, bf16 weights. The final boxed answer after the thinking block is checked against the reference with math-verify; samples that hit the token limit count as wrong. This was a single test run. Most of the base model's failures are samples that hit the token limit (93.5 % of base samples, 36.3 % for this model).
For comparison, MIT CSAIL's weight-editing agent reported 14.1% at 25 passes and 48.1% at 50 with a 131,072-weight edit.
Usage
Load and run it exactly like Qwen/Qwen3.5-0.8B, with the same recommended sampling settings.
License
The original model's license, Apache-2.0, applies. Base model by the Qwen team.
Citation
https://fermionresearch.com/research/mapping-every-neuron-in-qwen3-8b/