The Information Propagation Hypothesis: Optimizing Information Flow in Large Language Models

ACL ARR 2025 February Submission1895 Authors

Abstract:

Large language models (LLMs) have shown remarkable success in various tasks, yet their internal mechanisms remain inadequately understood.
This paper investigates these mechanisms by analyzing how input query information propagates within task-specific spaces.
Specifically, we propose a prompt-pair detection method that constructs a task-specific label space and projects hidden representations onto it to examine information propagation during the understanding, generation, and decision-making stages.
Our findings reveal that LLMs compress and decompress query information into hidden representations near the task-specific label space during the understanding and generation stages.
In the decision-making stage, labels with distributions similar to the query are predicted, but these labels do not always match the true labels, leading to errors.
To address this, we analyze the query distribution and find that queries tend to cluster around semantically similar queries, regardless of proximity to the true label. Based on this, we propose a similarity-based voting method (SiV) that aggregates votes from semantically similar queries to improve prediction accuracy, mitigating errors caused by relying solely on label similarity.
Extensive experiments show that SiV enhances both accuracy and speed, while also enabling incremental updates without training.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: large language models, model interpretability, emotional recognition, questions and answers, topic classification

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 1895