01 / AI · Analysis
Arithmetic circuits go into model weights without retraining
A paper compared writing arithmetic rules directly into model weights with fine-tuning on examples.

Researchers published a method for inserting arithmetic circuits into pretrained language models without retraining and tested it on six-digit calculations.
The paper puts the calculation rules in the model’s weights, the numerical values it learned during training. The method needs no additional training examples or separate code inserted while the model runs.
The researchers compared the circuits with fine-tuning on examples to see which method produced correct results more often.
Six-Digit Arithmetic Accuracy Reported by the Authors
| IBM Granite 4.2 3B | Circuit grafting | 99.17 | |
|---|---|---|---|
| LoRA | 42.2 | ||
| Alibaba Qwen 3.5 4B | Circuit grafting | 100 | |
| LoRA | 61.7 |
Calculation rules in model weights
The researchers built a program that turns rules for addition, subtraction and multiplication into values stored in the model’s weights.
According to the paper, the researchers widened the model’s internal numerical arrays to make room for the circuits. They also compensated for the resulting change in signal magnitude.
Accuracy on arithmetic responses
The paper compared the methods on IBM Granite 4.2 3B and Alibaba Qwen 3.5 4B. It measured how often the result at the start of a response matched the correct answer.
The baseline was LoRA, a fine-tuning method that used training examples in the study. In the researchers’ published results, circuit insertion scored higher than LoRA on both models.
The researchers included operations of up to eight digits in the scope of circuit generation. The abstract gave no separate accuracy figure for eight-digit operations.