Abstract
Factual faithfulness is a crucial requirement in information-seeking dialogue: the system should respond to the user queries so that the responses are meaningful and aligned with the knowledge provided to the system. However, most modern large language models (LLMs) suffer from hallucinations, that is, they generate responses not supported by or even contradicting the knowledge source. To mitigate the issue and increase faithfulness of information-seeking dialogue systems supported by the LLMs, we introduce BeInfo, a simple yet effective method that applies ‘behavioural tuning’ on the LLMs to aid information-seeking dialogue. Relying on three standard information seeking dialogue 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 present a ‘real-life’ case study on conversations with real users, showcasing that the models with 3B parameters (e.g., Flan-T5) tuned with BeInfo demonstrate strong performance on data from real ‘production’ conversations: when tuned on a limited amount of such realistic in-domain dialogues, they surpass much larger LLMs used ‘off-the-shelf’, both on automatic and human evaluation metrics.
Introduction
Large Language Models (LLMs) are increasingly deployed in information-seeking dialogue systems, such as customer service bots and virtual assistants. While Retrieval-Augmented Generation (RAG) helps ground these models, they still suffer from two primary failure modes: poor selectivity (failing to distinguish relevant from irrelevant facts) and a lack of response adequacy (hallucinating answers when the knowledge source is insufficient).
Researchers at PolyAI argue that standard fine-tuning focuses too much on content and not enough on the process of navigating external knowledge. To address this, they introduced BeInfo, a behavioural fine-tuning method designed to instill factual discipline in dialogue models.
Methodology
BeInfo (Behavioural Fine-Tuning for Information-Seeking) modifies the training process to explicitly teach models how to behave when presented with external data:
- Selectivity Training: During fine-tuning, ground-truth knowledge is bundled with irrelevant "distractor" facts. This forces the model to identify and extract only the pertinent information.
- Adequacy Training: The training set includes "unanswerable" examples where the knowledge source does not contain the answer. The model is trained to respond with a clarification request rather than generating a hallucination.
The study explores three regimes: General BeInfo (tuning on broad datasets), Task-specific BeInfo (tuning on the target domain), and Full BeInfo (a two-stage approach combining both).
Experimental Setup and Datasets
The framework was evaluated on three standard datasets: FaithDial (factual consistency), TopiOCQA (conversational QA), and DoQA (multi-domain). Additionally, a real-world case study was conducted using HOTEL-200, a dataset of 200 anonymized dialogues from a production-level hotel reservation system.
Key Insights and Results
Behavior over Scale
A striking finding is that behavioural preparation can compensate for a lack of raw parameters. A 3B-parameter model (Flan-T5 XL) tuned with BeInfo was more faithful to facts than off-the-shelf models many times its size. In the hotel domain case study, the Full BeInfo model achieved a GPT4-Eval score of 4.81/6.0, surpassing the baseline GPT-4 score of 4.63.
Zero-Shot Generalization and Hallucination Reduction
Models tuned on broad domains showed immediate improvements in faithfulness when tested on entirely different domains (e.g., Cooking or Travel) without further training. On the DoQA dataset, BeInfo improved the K-Precision (the proportion of the response supported by the knowledge source) of Flan-T5 XL from 94.12 to 100.0.
Furthermore, BeInfo shifted "terrible" hallucination failures into safe "clarification" responses. This transition is critical for production environments where providing no answer is far safer than providing a false one.
Conclusion
BeInfo demonstrates that specialized behavioural tuning is a highly effective strategy for improving the reliability of information-seeking dialogue systems. By prioritizing process over scale, smaller models can achieve production-grade performance that rivals or exceeds much larger general-purpose LLMs. Future work will focus on scaling this method to larger models and refining selectivity through the use of more complex "hard negative" distractors.
Summary updated