Probing Language Models on Their Knowledge Source

TL;DR

This paper introduces a probing framework to understand how large language models prioritize internal versus external knowledge during inference, revealing that mid-layer activations are key indicators of knowledge source selection.

Contribution

The paper presents a novel probing method to analyze knowledge source prioritization in LLMs, highlighting the importance of mid-layer activations in conflict resolution.

Findings

Mid-layer activations predict knowledge source selection.

Models show different behaviors based on input relations.

Framework works across various LLM sizes.

Abstract

Large Language Models (LLMs) often encounter conflicts between their learned, internal (parametric knowledge, PK) and external knowledge provided during inference (contextual knowledge, CK). Understanding how LLMs models prioritize one knowledge source over the other remains a challenge. In this paper, we propose a novel probing framework to explore the mechanisms governing the selection between PK and CK in LLMs. Using controlled prompts designed to contradict the model's PK, we demonstrate that specific model activations are indicative of the knowledge source employed. We evaluate this framework on various LLMs of different sizes and demonstrate that mid-layer activations, particularly those related to relations in the input, are crucial in predicting knowledge source selection, paving the way for more reliable models capable of handling knowledge conflicts effectively.