Large Language Models (LLMs) have made it possible for recommendation systems to interact with users in open-ended conversational interfaces. In order to personalize LLM responses, it is crucial to elicit user preferences, especially when there is limited user history. One way to get more information is to present clarifying questions to the user. However, generating effective sequential clarifying questions across various domains remains a challenge. To address this, we introduce a novel approach for training LLMs to ask sequential questions that reveal user preferences. Our method follows a two-stage process inspired by diffusion models. Starting from a user profile, the forward process generates clarifying questions to obtain answers and then removes those answers step by step, serving as a way to add ``noise'' to the user profile. The reverse process involves training a model to ``denoise'' the user profile by learning to ask effective clarifying questions. Our results show that our method significantly improves the LLM's proficiency in asking funnel questions and eliciting user preferences effectively.

As large language models (LLMs) are increasingly relied on in AI systems, predicting when they make mistakes is crucial. While a great deal of work in the field uses internal representations to interpret model behavior, these representations are inaccessible when given solely black-box access through an API. In this paper, we extract features of LLMs in a black-box manner by using follow-up prompts and taking the probabilities of different responses as representations to train reliable predictors of model behavior. We demonstrate that training a linear model on these low-dimensional representations produces reliable and generalizable predictors of model performance at the instance level (e.g., if a particular generation correctly answers a question). Remarkably, these can often outperform white-box linear predictors that operate over a model's hidden state or the full distribution over its vocabulary. In addition, we demonstrate that these extracted features can be used to evaluate more nuanced aspects of a language model's state. For instance, they can be used to distinguish between a clean version of GPT-4o-mini and a version that has been influenced via an adversarial system prompt that answers question-answering tasks incorrectly or introduces bugs into generated code. Furthermore, they can reliably distinguish between different model architectures and sizes, enabling the detection of misrepresented models provided through an API (e.g., identifying if GPT-3.5 is supplied instead of GPT-4o-mini). Large language models (LLMs) have demonstrated strong performance on a wide variety of tasks (Radford et al.), leading to their increased involvement in larger systems. For instance, they are often used to provide supervision (Bai et al., 2022; Sam et al., 2024), as tools in decision-making (Benary et al., 2023; Sha et al., 2023), or as controllers on agentic frameworks (Xi et al., 2023; Robey et al., 2024). Thus, it is crucial to understand and predict their behaviors, especially in high-stakes settings. However, as with any deep network, it is difficult to understand the behavior of such large models (Zhang et al., 2021). For instance, prior work has studied input gradients or saliency maps (Simonyan et al., 2013; Zeiler & Fergus, 2014; Pukdee et al., 2024)) to attempt to understand neural network behavior, but this can fail to reliably describe model behavior (Adebayo et al., 2018; Kindermans et al., 2019; Srinivas & Fleuret, 2020). Other work has studied the ability of transformers to represent certain algorithms (Nanda et al., 2022; Zhong et al., 2024) that may be involved in their predictions. One promising direction in understanding LLMs (or any other multimodal model that understands natural language) is to leverage their ability to interact with human queries. Recent work has demonstrated that a LLM's hidden state contains low-dimensional features of model truthfulness or harmfulness (Zou et al., 2023a).

A key distinguishing feature of conversational recommender systems over traditional recommender systems is their ability to elicit user preferences using natural language. Currently, the predominant approach to preference elicitation is to ask questions directly about items or item attributes. These strategies do not perform well in cases where the user does not have sufficient knowledge of the target domain to answer such questions. Conversely, in a shopping setting, talking about the planned use of items does not present any difficulties, even for those that are new to a domain. In this paper, we propose a novel approach to preference elicitation by asking implicit questions based on item usage. Our approach consists of two main steps. First, we identify the sentences from a large review corpus that contain information about item usage. Then, we generate implicit preference elicitation questions from those sentences using a neural text-to-text model. The main contributions of this work also include a multi-stage data annotation protocol using crowdsourcing for collecting high-quality labeled training data for the neural model. We show that our approach is effective in selecting review sentences and transforming them to elicitation questions, even with limited training data. Additionally, we provide an analysis of patterns where the model does not perform optimally.

With the aid of an intelligent system, a group of people (the key group) faces a hiring decision about many candidates who are characterized by attributes, such as experiences, technical skills, communication skills, etc. The goal is to help the key group make a group decision without directly eliciting their full preferences over all candidates, which is often infeasible given the vast number of candidates. Instead, the intelligent system may ask fellow employees (the regular group) about their preferences in order to learn about the key group's preferences. How can the intelligent system decide which member in the regular group to ask and which question should be asked? This example illustrates the preference elicitation problem, which has been widely studied in the field of recommender systems [Loepp et al., 2014], healthcare [Erdem and Campbell, 2017, Weernink et al., 2014], marketing [Huang and Luo, 2016], stable matching [Drummond and Boutilier, 2014, Rastegari et al., 2016], etc. Most previous works studied a special case of the aforementioned scenario, in which the regular group is the key group. The objective of preference elicitation is to achieve some goal using as few samples (data) as possible. A common approach is to adaptively ask questions that maximize expected information gain, measured by some information criteria. Moreover, most previous work focused on a specific type of elicitation questions, e.g.

Not all questions are asked with the same intention. Humans tend to address the implicit meaning of the question (that contributes to its pragmatic force), which requires knowledge of the context and a degree of common ground, more so than addressing the explicit propositional content of the question. Is recognizing the pragmatic force in today's human-robot dialogue systems worth the trouble? We focus on display questions (questions to which the asker already knows the answer) and argue that there are realistic human-robot interaction scenarios in existence today that would benefit from the deeper intention recognition. We also propose a method for obtaining display question annotations by embedding an elicitation question into the dialogue. The preliminary study of our robot receptionist shows that at least 16.7% of interactions with the embedded elicitation question include a display question.