arxiv.org/abs/2506.08966
14.06.2025 14:44
👍 1
🔁 0
💬 0
📌 0
arxiv.org/abs/2506.08966
Thought language models are bad with numbers?
So did we.
❌ Turns out we were wrong.
Models just use their own special representation for numeric values.
Similar to binary, closer to sinusoidal.
If we interpret this space right, we can recover values they represent with 100% accuracy!
#llms #nlp