FaithDial: A Faithful Benchmark for Information-Seeking Dialogue (TACL, MIT Press, Dec 2022)
- Source: TACL article, DOI 10.1162/tacl_a_00529
- Venue: Transactions of the Association for Computational Linguistics (TACL), MIT Press, December 2022
- Category: Evaluation of search engines / dialogue systems
- The authors manually revised the Wizard of Wikipedia (WoW) dataset: responses that hallucinated, were vague, uncooperative, or off-topic were rewritten to be faithful to the knowledge source while remaining engaging.
- Crowdworkers were shown the evidence passage and asked to fix problematic bot utterances, yielding a corrected evaluation set of dialogues.
- The benchmark is accompanied by human evaluation protocols: crowdworkers interact with models and assess faithfulness, knowledgeability, and engagement directly.
- Common automatic metrics for grounded dialogue (e.g., token-overlap and entailment scores against references) correlate weakly with human judgments of faithfulness — better reference scores do not mean fewer hallucinations.
- Models fine-tuned on FaithDial are preferred by human evaluators over the same models trained on the original WoW data, particularly on knowledgeability and faithfulness.
- Even the best models still hallucinate substantially; the authors frame faithfulness evaluation as requiring direct human assessment rather than reference-based metrics alone.
- The benchmark is derived from Wikipedia-domain dialogues, so cross-domain and multilingual generalization is not covered.
- Human evaluation is costly; the paper itself acknowledges that scalable automatic faithfulness metrics remain an open problem.
- Evaluation of Retrieval-Augmented Generation: A Survey (arXiv:2405.07437)
- A Dataset of Information-Seeking Questions and Answers Anchored in Research Abstracts (arXiv:2105.03011)
- ARES: An Automated Evaluation Framework for RAG (arXiv:2311.09476)
- AI Search Has A Citation Problem (CJR, Mar 2025)
Overview
FaithDial is a benchmark designed to measure how faithful information-seeking dialogue systems are — that is, whether their responses actually stay grounded in the provided knowledge rather than hallucinating unsupported content.
Motivation
Knowledge-grounded chatbots routinely produce responses that are generic, evasive, or contradicted by the evidence. Existing dialogue benchmarks inherit this problem: training and test data themselves contain hallucinations, making it difficult to measure model faithfulness meaningfully. The FaithDial paper addresses this by building a cleaner, human-verified benchmark.
Benchmark construction
Key findings
Relevance to search and RAG evaluation
FaithDial is frequently cited in work on hallucination detection, retrieval-augmented generation (RAG) evaluation, and conversational search. Its central lesson — that automatic metrics can reward unfaithful outputs — applies broadly to evaluating LLM-based search and RAG pipelines, where citation accuracy and grounding checks are increasingly treated as process-level metrics alongside task success rates.