Abstract
Factuality is a crucial requirement in information-seeking dialogue: the system should respond to the user's queries so that the responses are meaningful and aligned with the knowledge provided to the system. However, most modern large language models suffer from hallucinations, that is, they generate responses not supported by or contradicting the knowledge source. To mitigate the issue and increase faithfulness of information-seeking dialogue systems, we introduce BeInfo, a simple yet effective method that applies behavioural tuning to aid information-seeking dialogue. Relying on three standard datasets, we show that models tuned with BeInfo become considerably more faithful to the knowledge source both for datasets and domains seen during BeInfo-tuning, as well as on unseen domains, when applied in a zero-shot manner. In addition, we show that the models with 3B parameters (e.g., Flan-T5) tuned with BeInfo demonstrate strong performance on data from real 'production' conversations and outperform GPT-4 when tuned on a limited amount of such realistic in-domain dialogues.
Introduction: The Challenge of Hallucination
Large Language Models (LLMs) are notorious for "hallucinating"—generating responses that sound confident but are factually incorrect or unsupported by the provided context. While Retrieval-Augmented Generation (RAG) helps by providing external knowledge to the model, it is not a complete solution. Models often fail to be selective (picking the right info from a noisy source) or adequate (admitting when the provided source does not contain the answer).
In high-stakes environments like customer service or technical support, these failures are deal-breakers. This paper introduces BeInfo, a method that treats faithfulness as a behavioral trait that can be instilled into a model regardless of the specific facts it knows.
Methodology: Behavioural Fine-Tuning (BeInfo)
Instead of just fine-tuning a model to answer questions, BeInfo tunes it to exhibit specific "faithful" behaviours through two primary strategies:
- Selectivity Training: The authors augment training data by adding "distractors"—randomly sampled, irrelevant facts—to the ground-truth knowledge source. This forces the model to ignore noise and pinpoint the specific information relevant to the user's query.
- Adequacy Training: They intentionally include "unanswerable" dialogues where the provided knowledge source is irrelevant to the query. The model is trained to respond with a clarification request rather than making something up.
The study explores three training regimes: General-only (tuning on public datasets for zero-shot transfer), Task-only (tuning directly on domain-specific data), and Full BeInfo (a two-stage process combining general behavioural tuning with task-specific refinement).
Experimental Setup and Datasets
The research utilizes four primary datasets for evaluation:
- FaithDial: Designed for faithful, knowledge-grounded dialogue.
- TopiOCQA: Conversational QA involving complex topic switching.
- DoQA: A domain-specific FAQ dataset (Cooking, Travel, Movies) used for testing zero-shot generalization.
- HOTEL-200: A new internal evaluation set consisting of 200 anonymized real-world conversations in the hotel reservation domain to test production-ready performance.
Key Results
BeInfo-tuned models showed significant improvements in faithfulness across all benchmarks. A key finding is that smaller models (3B parameters) can be more reliable than massive off-the-shelf models when properly tuned.
In the real-world HOTEL-200 case study, the Flan-T5 XL model (3B) using the "Full BeInfo" regime achieved a GPT4-Eval score of 4.81, surpassing GPT-4's own performance as a responder (4.63) on a 7-point faithfulness scale. Furthermore, on the DoQA Travel domain, BeInfo-tuned models achieved near-perfect K-Precision, effectively eliminating "out-of-thin-air" hallucinations.
Conclusion and Future Work
BeInfo demonstrates that faithfulness is a behavioral skill that transfers to entirely new domains. By training for selectivity and adequacy, developers can deploy efficient, low-latency models that rival the factuality of the largest LLMs. Future research will explore the use of "hard distractors" (semantically similar but incorrect facts) and the integration of BeInfo with the retrieval components of RAG systems to handle end-to-end errors.
Summary updated