Papers by Toni Liu
LLMs learn governing principles of dynamical systems, revealing an in-context neural scaling law (2024.emnlp-main)
Copied to clipboard
| Challenge: | LLaMA-2 language model is capable of in-context time series extrapolation without specific prompting or fine-tuning, revealing an in-constitution version of a neural scaling law. |
| Approach: | They propose an algorithm for extracting probability density functions of multi-digit numbers directly from Large language models (LLMs) LLaMA-2 is a language model trained on text and can extrapolate dynamical system time series without prompting or engineering . |
| Outcome: | The proposed model can extrapolate dynamical systems without prompting or engineering . it also achieves an in-context version of a neural scaling law . |