Reconciling Fidelity and Expressiveness in Large Language Models

ACL ARR 2025 February Submission5910 Authors

Abstract: Grounding on external knowledge to generate responses is an effective method to mitigate hallucinations for Large Language Models (LLMs). However, current LLMs struggle to weave knowledge into responses seamlessly while ensuring fidelity like humans, often producing outputs that are either unsupported by external knowledge or overly verbose and unnatural. In this work, to break the trade-off between fidelity and expressiveness, we propose Collaborative Decoding (CoDe), which integrates the output probabilities with and without external knowledge based on their distribution divergence and model confidence to dynamically arouse relevant and reliable expressions from model's internal parameter. Additionally, a knowledge-aware reranking mechanism is designed to prevent the model from being overly confident in its prior parameter knowledge and from ignoring the given external knowledge. With extensive experiments, our plug-and-play CoDe achieved excellent performance in enhancing fidelity without sacrificing expressiveness on different LLMs and metrics, proving its effectiveness and generality.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: Large Language Models, Decoding Strategy, Text hallucination

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: english

Submission Number: 5910