Recent research has focused on developing non-task-oriented dialogue agents with high anthropomorphism, aiming to sustain users’ dialogue motivation over prolonged periods. Previous studies have introduced chatbots that deliberately incorporate short response delays to enhance anthropomorphism and user satisfaction. However, in the context of long-term casual conversations, the behavior of these chatbots, which reply within a few seconds, may lead to a perception of the chatbot as always on standby, lacking its own schedule or personal life, thus appearing less humanlike. To address this issue, we propose a non-real-time chatbot that interacts with users while introducing deliberate response delays on a larger time scale, ranging from a few minutes to several hours. A five-day dialogue experiment confirmed that the non-real-time chatbot was perceived as more humanlike and encouraged users to engage in extensive dialogue. These insights are invaluable for designing chatbots that can maintain casual conversations over prolonged periods.

Existing Theory of Mind (ToM) benchmarks diverge from real-world scenarios in three aspects: 1) they assess a limited range of mental states such as beliefs, 2) false beliefs are not comprehensively explored, and 3) the diverse personality traits of characters are overlooked. To address these challenges, we introduce ToMATO, a new ToM benchmark formulated as multiple-choice QA over conversations. ToMATO is generated via LLM-LLM conversations featuring information asymmetry. By employing a prompting method that requires role-playing LLMs to verbalize their thoughts before each utterance, we capture both first- and second-order mental states across five categories: belief, intention, desire, emotion, and knowledge. These verbalized thoughts serve as answers to questions designed to assess the mental states of characters within conversations. Furthermore, the information asymmetry introduced by hiding thoughts from others induces the generation of false beliefs about various mental states. Assigning distinct personality traits to LLMs further diversifies both utterances and thoughts. ToMATO consists of 5.4k questions, 753 conversations, and 15 personality trait patterns. Our analysis shows that this dataset construction approach frequently generates false beliefs due to the information asymmetry between role-playing LLMs, and effectively reflects diverse personalities. We evaluate nine LLMs on ToMATO and find that even GPT-4o mini lags behind human performance, especially in understanding false beliefs, and lacks robustness to various personality traits.

To improve user engagement during conversations with dialogue systems, we must improve individual dialogue responses and dialogue impressions such as consistency, personality, and empathy throughout the entire dialogue. While such dialogue systems have been developing rapidly with the help of large language models (LLMs), reinforcement learning from AI feedback (RLAIF) has attracted attention to align LLM-based dialogue models for such dialogue impressions. In RLAIF, a reward model based on another LLM is used to create a training signal for an LLM-based dialogue model using zero-shot/few-shot prompting techniques. However, evaluating an entire dialogue only by prompting LLMs is challenging. In this study, the supervised fine-tuning (SFT) of LLMs prepared reward models corresponding to 12 metrics related to the impression of the entire dialogue for evaluating dialogue responses. We tuned our dialogue models using the reward model signals as feedback to improve the impression of the system. The results of automatic and human evaluations showed that tuning the dialogue model using our reward model corresponding to dialogue impression improved the evaluation of individual metrics and the naturalness of the dialogue response.

Facilitating self-disclosure without causing discomfort remains a difficult task—especially for AI systems. In real-world applications such as career counseling, wellbeing support, and onboarding interviews, eliciting personal information like concerns, goals, and personality traits is essential. However, asking such questions directly often leads to discomfort and disengagement. We address this issue with RaPSIL (Rapport-aware Preference-guided Self-disclosure Interview Learner), a two-stage LLM-based system that fosters natural, engaging conversations to promote self-disclosure. In the first stage, RaPSIL selectively imitates interviewer utterances that have been evaluated by LLMs for both strategic effectiveness and social sensitivity. It leverages LLMs as multi-perspective judges in this selection process. In the second stage, it conducts self-play simulations, using the Reflexion framework to analyze failures and expand a database with both successful and problematic utterances. This dual learning process allows RaPSIL to go beyond simple imitation, improving its ability to handle sensitive topics naturally by learning from both successful and failed utterances. In a comprehensive evaluation with real users, RaPSIL outperformed baselines in enjoyability, warmth, and willingness to re-engage, while also capturing self-descriptions more accurately. Notably, its impression scores remained stable even during prolonged interactions, demonstrating its ability to balance rapport building with effective information elicitation. These results show that RaPSIL enables socially aware AI interviewers capable of eliciting sensitive personal information while maintaining user trust and comfort—an essential capability for real-world dialogue systems.

Long-term chatbots are expected to develop relationships with users. The major trend in this field’s recent long-term chatbot studies is to train systems with virtual long-term chat data called Multi-Session Chat (MSC), which collects text chat from multiple sessions of crowd workers playing the roles of speakers with defined personas. However, no investigation has attempted to determine whether such virtual long-term chat can successfully simulate relationship-building between speakers. To clarify the difference between an actual long-term intimacy process and an MSC intimacy process, this study collects real long-term chat and MSC in Japanese and compares them in terms of speech form and dialogue acts. The results of analyzing these factors suggest that MSC have an unnatural tendency to behave as if they have a close relationship with non-polite speech levels compared to actual long-term chats, but also as if they have a shallow relationship with more questions than real long-term chats.

In recent years, several high-performance conversational systems have been proposed based on the Transformer encoder-decoder model. Although previous studies analyzed the effects of the model parameters and the decoding method on subjective dialogue evaluations with overall metrics, it is not analyzed enough how the differences of fine-tuning datasets affect the user's detailed impressions. In addition, the Transformer-based approach has mostly been verified for English, not for such languages as Japanese that have large inter-language distances. In this study, we developed large-scale Transformer-based Japanese dialogue models and Japanese chit-chat datasets and examined their effectiveness. We analyzed the relationships between users' multifaceted impressions and fine-tuning datasets.

Tsunehiro Arimoto, Ryuichiro Higashinaka, Kou Tanaka, Takahito Kawanishi, Hiroaki Sugiyama, Hiroshi Sawada, Hiroshi Ishiguro. Proceedings of the 21th Annual Meeting of the Special Interest Group on Discourse and Dialogue. 2020.

Research on conversational dialogue systems has attracted attention for achieving dialogue systems that can build social relationships with users. Although such systems are required for continuing long conversations with users to build relationships, they sometimes make sentences that are not related to the dialogue context, causing the dialogue to easily break down. The cause of the difficulty is simple. Unlike task-oriented dialogues, unexpected user utterances whose meaning the system cannot capture are frequently spoken in conversational dialogues, since the dialogue domain is much less restricted and the dialogue goal is less obvious. In this paper, we propose a novel strategy for dialogue systems through which two robots coordinate to create long conversations by avoiding dialogue breakdowns. If one of the two robots accepts user utterances with backchannel, the responsibility of responding to the user utterances is resolved and the other robot can change the dialogue topic, which decreases the risk of generating discontinuous utterances. Even if a dialogue nearly breaks down, such multiple robots can present predefined natural interactions among themselves to repair it. Our experiments show that the inter-robot interaction effectively improves the establishment of conversational dialogue that helps users continue the dialogue.

The development of open-domain conversational systems remains difficult since user utterances are widely varied for such systems to respond appropriately. To address this is- sue, previous research has retrieved sentences from the web as system utterances by shallow sentence matching with user utterances. However, since the retrieved sentences include the inherent contexts of the document in which the sentences originally appeared, the retrieved sentences have the possibility of containing information that is irrelevant to user utter-ances. We propose combining two strongly related semantic units (phrase pairs with dependency relations) to create a system utterance. Here, the first semantic unit is the one found in the user utterance and the second semantic unit is the one that has a dependency relation with the first one in a large text cor- pus. This way, we can guarantee that the generated utterance is related to the input user utterance. Our experiments, which examine the appropriateness of response sentences, show that our proposed method significantly outperforms other retrieval and rule-based approaches.

Dialog systems that realize dialog control with reinforcement learning have recently been proposed. However, reinforcement learning has an open problem that it requires a reward function that is difficult to set appropriately. To set the appropriate reward function automatically, we propose preference-learning based inverse reinforcement learning (PIRL) that estimates a reward function from dialog sequences and their pairwisepreferences, which is calculated with annotated ratings to the sequences. Inverse reinforcement learning finds a reward function, with which a system generates similar sequences to the training ones. This indicates that current IRL supposes that the sequences are equally appropriate for a given task; thus, it cannot utilize the ratings. In contrast, our PIRL can utilize pairwise preferences of the ratings to estimate the reward function. We examine the advantages of PIRL through comparisons between competitive algorithms that have been widely used to realize the dialog control. Our experiments show that our PIRL outperforms the other algorithms and has a potential to be an evaluation simulator of dialog control.

This paper proposes a HMM-based user's information demand estimation model for autonomous informational assistance robots to avoid providing information prematurely. The model estimates the user's implicit information demands by predicting a user's next information request using user's head movements. Through a word-association quiz-dialog experiment, our model demonstrated superior prediction performance over the usual HMM-based classifier.