Paper Image

How language models process machine-generated prompts

Published on:

24 October 2023

Primary Category:

Computation and Language

Paper Authors:

Corentin Kervadec,

Francesca Franzon,

Marco Baroni

Bullets

Key Details

Machine prompts like 'Antarctica sequelsStationrough' outperform sensible human prompts

Human and machine prompts differ in perplexity, attention, activations

Machine prompts activate units responding to non-linguistic items

Results suggest machine prompts use different internal pathways

Understanding unnatural language processing is key for model controllability

AI generated summary

How language models process machine-generated prompts

This paper investigates how language models respond to machine-generated prompts that lack clear semantic meaning, compared to human-written prompts. The authors find differences in perplexity, attention patterns, knowledge neuron activation, and associated vocabulary items when processing human vs. machine prompts. Their analysis suggests machine prompts trigger distinct internal pathways not specialized for natural language.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up