Interpretable Next-token Prediction via the Generalized Induction Head

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Kim, Eunji, Mantena, Sriya, Yang, Weiwei, Singh, Chandan, Yoon, Sungroh, Gao, Jianfeng
Format: Preprint
Veröffentlicht: 2024
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866915585667891200
author Kim, Eunji
Mantena, Sriya
Yang, Weiwei
Singh, Chandan
Yoon, Sungroh
Gao, Jianfeng
author_facet Kim, Eunji
Mantena, Sriya
Yang, Weiwei
Singh, Chandan
Yoon, Sungroh
Gao, Jianfeng
contents While large transformer models excel in predictive performance, their lack of interpretability restricts their usefulness in high-stakes domains. To remedy this, we propose the Generalized Induction-Head Model (GIM), an interpretable model for next-token prediction inspired by the observation of "induction heads" in LLMs. GIM is a retrieval-based module that identifies similar sequences in the input context by combining exact n-gram matching and fuzzy matching based on a neural similarity metric. We evaluate GIM in two settings: language modeling and fMRI response prediction. In language modeling, GIM improves next-token prediction by up to 25%p over interpretable baselines, significantly narrowing the gap with black-box LLMs. In an fMRI setting, GIM improves neural response prediction by 20% and offers insights into the language selectivity of the brain. GIM represents a significant step toward uniting interpretability and performance across domains. The code is available at https://github.com/ejkim47/generalized-induction-head.
format Preprint
id arxiv_https___arxiv_org_abs_2411_00066
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Interpretable Next-token Prediction via the Generalized Induction Head
Kim, Eunji
Mantena, Sriya
Yang, Weiwei
Singh, Chandan
Yoon, Sungroh
Gao, Jianfeng
Computation and Language
Artificial Intelligence
Machine Learning
While large transformer models excel in predictive performance, their lack of interpretability restricts their usefulness in high-stakes domains. To remedy this, we propose the Generalized Induction-Head Model (GIM), an interpretable model for next-token prediction inspired by the observation of "induction heads" in LLMs. GIM is a retrieval-based module that identifies similar sequences in the input context by combining exact n-gram matching and fuzzy matching based on a neural similarity metric. We evaluate GIM in two settings: language modeling and fMRI response prediction. In language modeling, GIM improves next-token prediction by up to 25%p over interpretable baselines, significantly narrowing the gap with black-box LLMs. In an fMRI setting, GIM improves neural response prediction by 20% and offers insights into the language selectivity of the brain. GIM represents a significant step toward uniting interpretability and performance across domains. The code is available at https://github.com/ejkim47/generalized-induction-head.
title Interpretable Next-token Prediction via the Generalized Induction Head
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2411.00066