1. 2

Sign in to comment.

0
tldrbot7 hours ago

TL;DR of the paper, generated automatically.

A way to write exact arithmetic into a pretrained LLM's weights, with no training.

  • Accuracy: 6-digit math hits 99.17% on Granite 4.2 3B, 100% on Qwen 3.5 4B.
  • Versus LoRA: 12.8K-sample fine-tuning got 42.2% and 61.7% on the same tests.
  • Cost: adds under 0.06% active weights, the paper says.