[Paper Review] Should We Fine-Tune or RAG? Evaluating Different Techniques to Adapt LLMs for Dialogue

Simone Alghisi, Massimo Rizzoli+3 more|arXiv (Cornell University)|Jun 10, 2024

Speech and dialogue systems4 citations

TL;DR

This paper evaluates fine-tuning and in-context learning with Retrieval-Augmented Generation (RAG) across four dialogue types—open-domain, knowledge-grounded, task-oriented, and question answering—using Llama2 Chat and Mistral Instruct. Results show no universal best technique; performance depends on base model and dialogue type, with human evaluation critical to avoid misleading automatic metrics.

ABSTRACT

We study the limitations of Large Language Models (LLMs) for the task of response generation in human-machine dialogue. Several techniques have been proposed in the literature for different dialogue types (e.g., Open-Domain). However, the evaluations of these techniques have been limited in terms of base LLMs, dialogue types and evaluation metrics. In this work, we extensively analyze different LLM adaptation techniques when applied to different dialogue types. We have selected two base LLMs, Llama-2 and Mistral, and four dialogue types Open-Domain, Knowledge-Grounded, Task-Oriented, and Question Answering. We evaluate the performance of in-context learning and fine-tuning techniques across datasets selected for each dialogue type. We assess the impact of incorporating external knowledge to ground the generation in both scenarios of Retrieval-Augmented Generation (RAG) and gold knowledge. We adopt consistent evaluation and explainability criteria for automatic metrics and human evaluation protocols. Our analysis shows that there is no universal best-technique for adapting large language models as the efficacy of each technique depends on both the base LLM and the specific type of dialogue. Last but not least, the assessment of the best adaptation technique should include human evaluation to avoid false expectations and outcomes derived from automatic metrics.

Motivation & Objective

• To systematically evaluate fine-tuning and in-context learning across multiple dialogue types using consistent metrics and protocols.
• To assess the impact of external knowledge—both retrieved via RAG and gold-standard—on response generation quality.
• To compare performance across two base LLMs: Llama2 Chat and Mistral Instruct.
• To highlight the limitations of automatic metrics by demonstrating discrepancies with human evaluation outcomes.
• To provide explainability via integrated gradients to analyze input contribution in response generation.

Proposed method

• Fine-tuned Llama2 Chat and Mistral Instruct on four dialogue datasets: Open-Domain, Knowledge-Grounded, Task-Oriented, and Question Answering.
• Applied in-context learning using few-shot demonstrations without model parameter updates.
• Integrated Retrieval-Augmented Generation (RAG) with top-3 retrieved knowledge passages to ground responses.
• Used gold knowledge (ground-truth) as a strong baseline to evaluate retrieval quality and model performance.
• Employed automatic metrics (BLEU-4, KF1, ROUGE-L) and a standardized human evaluation protocol for correctness, appropriateness, and validity.
• Applied integrated gradients to attribute input token contributions and analyze generation errors.

Experimental results

Research questions

• RQ1 Which adaptation technique—fine-tuning or in-context learning—yields superior response quality across different dialogue types?
• RQ2 How does the inclusion of retrieved or gold external knowledge affect model performance in each dialogue type?
• RQ3 To what extent do automatic metrics align with human judgments in evaluating dialogue response quality?
• RQ4 How do different base LLMs (Llama2 Chat and Mistral Instruct) perform under the same adaptation techniques and dialogue tasks?
• RQ5 What are the key error patterns in LLM-generated responses, and how can they be explained via input attribution?

Key findings

• Fine-tuning outperformed in-context learning in 3 out of 4 dialogue types (KGD, TOD, QA), with the highest F1 and BLEU-4 scores on gold knowledge.
• For Task-Oriented Dialogues, fine-tuned models achieved 45.49 KF1 with gold knowledge, significantly outperforming in-context learning (28.79 KF1).
• In Open-Domain Dialogues, in-context learning with no knowledge achieved 95% correctness, while fine-tuning dropped to 65%, indicating task-specific model behavior.
• RAG with retrieved knowledge improved performance over no knowledge in most cases, but recall was low (e.g., 69% for KGD paragraph retrieval), limiting gains.
• Human evaluation revealed that automatic metrics like ROUGE-L can mislead: in-context learning generated longer, more detailed responses but scored lower due to short ground-truths.
• Gold knowledge consistently improved performance across models and dialogue types, with fine-tuned Mistral Instruct achieving 85% correctness in QA with gold knowledge.