This study explores replacing Transformers in Visual Language Models (VLMs) with Mamba, a recent structured state space model (SSM) that demonstrates promising performance in sequence modeling. We test models up to 3B parameters under controlled conditions, showing that Mamba-based VLMs outperforms Transformers-based VLMs in captioning, question answering, and reading comprehension. However, we find that Transformers achieve greater performance in visual grounding and the performance gap widens with scale. We explore two hypotheses to explain this phenomenon: 1) the effect of task-agnostic visual encoding on the updates of the hidden states, and 2) the difficulty in performing visual grounding from the perspective of in-context multimodal retrieval. Our results indicate that a task-aware encoding yields minimal performance gains on grounding, however, Transformers significantly outperform Mamba at in-context multimodal retrieval. Overall, Mamba shows promising performance on tasks where the correct output relies on a summary of the image but struggles when retrieval of explicit information from the context is required.

The Generalized Referring Expression Comprehension (GREC) task extends classic REC by generating image bounding boxes for objects referred to in natural language expressions, which may indicate zero, one, or multiple targets. This generalization enhances the practicality of REC models for diverse real-world applications. However, the presence of varying numbers of targets in samples makes GREC a more complex task, both in terms of training supervision and final prediction selection strategy. Addressing these challenges, we introduce RECANTFormer, a one-stage method for GREC that combines a decoder-free (encoder-only) transformer architecture with DETR-like Hungarian matching. Our approach consistently outperforms baselines by significant margins in three GREC datasets.

AI personal assistants deployed via robots or wearables require embodied understanding to collaborate with humans effectively. However, current Vision-Language Models (VLMs) primarily focus on third-person view videos, neglecting the richness of egocentric perceptual experience. To address this gap, we propose three key contributions. First, we introduce the Egocentric Video Understanding Dataset (EVUD) for training VLMs on video captioning and question answering tasks specific to egocentric videos. Second, we present , a 7B parameter VLM trained using parameter-efficient methods on EVUD. Finally, we evaluate ‘s capabilities on OpenEQA, a challenging benchmark for embodied video question answering. Our model achieves state-of-the-art performance, outperforming open-source models including strong Socratic models using GPT-4 as a planner by 3.6%.Additionally, we outperform Claude 3 and Gemini Pro Vision 1.0 and showcase competitive results compared to Gemini Pro 1.5 and GPT-4V, even surpassing the latter in spatial reasoning. This research paves the way for building efficient VLMs that can be deployed in robots or wearables, leveraging embodied video understanding to collaborate seamlessly with humans in everyday tasks, contributing to the advancement of next-generation Embodied AI.

Ambiguous Candidate Identification(ACI) in multimodal dialogue is the task of identifying all potential objects that a user’s utterance could be referring to in a visual scene, in cases where the reference cannot be uniquely determined. End-to-end models are the dominant approach for this task, but have limited real-world applicability due to unrealistic inference-time assumptions such as requiring predefined catalogues of items. Focusing on a more generalized and realistic ACI setup, we demonstrate that a modular approach, which first emphasizes language-only reasoning over dialogue context before performing vision-language fusion, significantly outperforms end-to-end trained baselines. To mitigate the lack of annotations for training the language-only module (student), we propose a pseudo-labelling strategy with a prompted Large Language Model (LLM) as the teacher.

An effective method for combining frozen large language models (LLM) and visual encoders involves a resampler module that creates a ‘visual prompt’ which is provided to the LLM, along with the textual prompt. While this approach has enabled impressive performance across many coarse-grained tasks like image captioning and visual question answering, more fine-grained tasks that require spatial understanding have not been thoroughly examined. In this paper, we use diagnostic classifiers to measure the extent to which the visual prompt produced by the resampler encodes spatial information. Our results show that this information is largely absent from the resampler output when kept frozen during training of the classifiers. However, when the resampler and classifier are trained jointly, we observe a significant performance boost. This shows that the compression achieved by the resamplers can in principle encode the requisite spatial information, but that more object-aware objectives are needed at the pretraining stage to facilitate this capability.

We have deployed an LLM-based spoken dialogue system in a real hospital. The ARI social robot embodies our system, which patients and their companions can have multi-party conversations with together. In order to enable this multi-party ability, multimodality is critical. Our system, therefore, receives speech and video as input, and generates both speech and gestures (arm, head, and eye movements). In this paper, we describe our complex setting and the architecture of our dialogue system. Each component is detailed, and a video of the full system is available with the appropriate components highlighted in real-time. Our system decides when it should take its turn, generates human-like clarification requests when the patient pauses mid-utterance, answers in-domain questions (grounding to the in-prompt knowledge), and responds appropriately to out-of-domain requests (like generating jokes or quizzes). This latter feature is particularly remarkable as real patients often utter unexpected sentences that could not be handled previously.

People understand and produce language incrementally on a word by word basis. This gives rise to many characteristic conversational phenomena including long mid-sentence pauses that are followed by incremental clarification requests (iCRs) intended to recover the rest of the truncated turn (see Fig. 1; (A), (B), (C)). The ability to generate iCRs is important in natural conversational AI systems, and crucial to their accessibility to users with memory impairment. In this paper, we collect, release and analyse SLUICE-CR: a large corpus of 3000 human produced iCRs. We then use this corpus to probe the incremental processing capability of a number of state of the art LLMs by evaluating the quality of the model’s generated iCRs in response to incomplete questions. Our evaluations show that the ability to generate contextually appropriate iCRs only emerges at larger LLM sizes, and only when prompted with example iCRs from our corpus. They also indicate that autoregressive LMs are, in principle, able to both understand and generate language incrementally.

This paper evaluates the extent to which current LLMs can capture task-oriented multi-party conversations (MPCs). We have recorded and transcribed 29 MPCs between patients, their companions, and a social robot in a hospital. We then annotated this corpus for multi-party goal-tracking and intent-slot recognition. People share goals, answer each other’s goals, and provide other people’s goals in MPCs - none of which occur in dyadic interactions. To understand user goals in MPCs, we compared three methods in zero-shot and few-shot settings: we fine-tuned T5, created pre-training tasks to train DialogLM using LED, and employed prompt engineering techniques with GPT-3.5-turbo, to determine which approach can complete this novel task with limited data. GPT-3.5-turbo significantly outperformed the others in a few-shot setting. The ‘reasoning’ style prompt, when given 7% of the corpus as example annotated conversations, was the best performing method. It correctly annotated 62.32% of the goal tracking MPCs, and 69.57% of the intent-slot recognition MPCs. A ‘story’ style prompt increased model hallucination, which could be detrimental if deployed in safety-critical settings. We conclude that multi-party conversations still challenge state-of-the-art LLMs.

We demonstrate an embodied conversational agent that can function as a receptionist and generate a mixture of open and closed-domain dialogue along with facial expressions, by using a large language model (LLM) to develop an engaging conversation. We deployed the system onto a Furhat robot, which is highly expressive and capable of using both verbal and nonverbal cues during interaction. The system was designed specifically for the National Robotarium to interact with visitors through natural conversations, providing them with information about the facilities, research, news, upcoming events, etc. The system utilises the state-of-the-art GPT-3.5 model to generate such information along with domain-general conversations and facial expressions based on prompt engineering.

Interactive and embodied tasks pose at least two fundamental challenges to existing Vision & Language (VL) models, including 1) grounding language in trajectories of actions and observations, and 2) referential disambiguation. To tackle these challenges, we propose an Embodied MultiModal Agent (EMMA): a unified encoder-decoder model that reasons over images and trajectories, and casts action prediction as multimodal text generation. By unifying all tasks as text generation, EMMA learns a language of actions which facilitates transfer across tasks. Different to previous modular approaches with independently trained components, we use a single multitask model where each task contributes to goal completion. EMMA performs on par with similar models on several VL benchmarks and sets a new state-of-the-art performance (36.81% success rate) on the Dialog-guided Task Completion (DTC), a benchmark to evaluate dialog-guided agents in the Alexa Arena.

SimpleMTOD is a simple language model which recasts several sub-tasks in multimodal task-oriented dialogues as sequence prediction tasks. SimpleMTOD is built on a large-scale transformer-based auto-regressive architecture, which has already proven to be successful in uni-modal task-oriented dialogues, and effectively leverages transfer learning from pretrained GPT-2. In-order to capture the semantics of visual scenes, we introduce both local and de-localized tokens for objects within a scene. De-localized tokens represent the type of an object rather than the specific object itself and so possess a consistent meaning across the dataset. SimpleMTOD achieves a state-of-the-art BLEU score (0.327) in the Response Generation sub-task of the SIMMC 2.0 test-std dataset while performing on par in other multimodal sub-tasks: Disambiguation, Coreference Resolution, and Dialog State Tracking. This is despite taking a minimalist approach for extracting visual (and non-visual) informa- tion. In addition the model does not rely on task-specific architectural changes such as classification heads.

Socially Assistive Robots (SARs) have the potential to play an increasingly important role in a variety of contexts including healthcare, but most existing systems have very limited interactive capabilities. We will demonstrate a robot receptionist that not only supports task-based and social dialogue via natural spoken conversation but is also capable of visually grounded dialogue; able to perceive and discuss the shared physical environment (e.g. helping users to locate personal belongings or objects of interest). Task-based dialogues include check-in, navigation and FAQs about facilities, alongside social features such as chit-chat, access to the latest news and a quiz game to play while waiting. We also show how visual context (objects and their spatial relations) can be combined with linguistic representations of dialogue context, to support visual dialogue and question answering. We will demonstrate the system on a humanoid ARI robot, which is being deployed in a hospital reception area.

We demonstrate EMMA, an embodied multimodal agent which has been developed for the Alexa Prize SimBot challenge. The agent acts within a 3D simulated environment for household tasks. EMMA is a unified and multimodal generative model aimed at solving embodied tasks. In contrast to previous work, our approach treats multiple multimodal tasks as a single multimodal conditional text generation problem, where a model learns to output text given both language and visual input. Furthermore, we showcase that a single generative agent can solve tasks with visual inputs of varying length, such as answering questions about static images, or executing actions given a sequence of previous frames and dialogue utterances. The demo system will allow users to interact conversationally with EMMA in embodied dialogues in different 3D environments from the TEACh dataset.

Visual Question Answering (VQA) systems are increasingly adept at a variety of tasks, and this technology can be used to assist blind and partially sighted people. To do this, the system’s responses must not only be accurate, but usable. It is also vital for assistive technologies to be designed with a focus on: (1) privacy, as the camera may capture a user’s mail, medication bottles, or other sensitive information; (2) transparency, so that the system’s behaviour can be explained and trusted by users; and (3) controllability, to tailor the system for a particular domain or user group. We have therefore extended a conversational VQA framework, called Aye-saac, with these objectives in mind. Specifically, we gave Aye-saac the ability to answer visual questions in the kitchen, a particularly challenging area for visually impaired people. Our system can now answer questions about quantity, positioning, and system confidence in regards to 299 kitchen objects. Questions about the spatial relations between these objects are particularly helpful to visually impaired people, and our system output more usable answers than other state of the art end-to-end VQA systems.

Guessing games are a prototypical instance of the “learning by interacting” paradigm. This work investigates how well an artificial agent can benefit from playing guessing games when later asked to perform on novel NLP downstream tasks such as Visual Question Answering (VQA). We propose two ways to exploit playing guessing games: 1) a supervised learning scenario in which the agent learns to mimic successful guessing games and 2) a novel way for an agent to play by itself, called Self-play via Iterated Experience Learning (SPIEL). We evaluate the ability of both procedures to generalise: an in-domain evaluation shows an increased accuracy (+7.79) compared with competitors on the evaluation suite CompGuessWhat?!; a transfer evaluation shows improved performance for VQA on the TDIUC dataset in terms of harmonic average accuracy (+5.31) thanks to more fine-grained object representations learned via SPIEL.

In visual guessing games, a Guesser has to identify a target object in a scene by asking questions to an Oracle. An effective strategy for the players is to learn conceptual representations of objects that are both discriminative and expressive enough to ask questions and guess correctly. However, as shown by Suglia et al. (2020), existing models fail to learn truly multi-modal representations, relying instead on gold category labels for objects in the scene both at training and inference time. This provides an unnatural performance advantage when categories at inference time match those at training time, and it causes models to fail in more realistic “zero-shot” scenarios where out-of-domain object categories are involved. To overcome this issue, we introduce a novel “imagination” module based on Regularized Auto-Encoders, that learns context-aware and category-aware latent embeddings without relying on category labels at inference time. Our imagination module outperforms state-of-the-art competitors by 8.26% gameplay accuracy in the CompGuessWhat?! zero-shot scenario (Suglia et al., 2020), and it improves the Oracle and Guesser accuracy by 2.08% and 12.86% in the GuessWhat?! benchmark, when no gold categories are available at inference time. The imagination module also boosts reasoning about object properties and attributes.

Approaches to Grounded Language Learning are commonly focused on a single task-based final performance measure which may not depend on desirable properties of the learned hidden representations, such as their ability to predict object attributes or generalize to unseen situations. To remedy this, we present GroLLA, an evaluation framework for Grounded Language Learning with Attributes based on three sub-tasks: 1) Goal-oriented evaluation; 2) Object attribute prediction evaluation; and 3) Zero-shot evaluation. We also propose a new dataset CompGuessWhat?! as an instance of this framework for evaluating the quality of learned neural representations, in particular with respect to attribute grounding. To this end, we extend the original GuessWhat?! dataset by including a semantic layer on top of the perceptual one. Specifically, we enrich the VisualGenome scene graphs associated with the GuessWhat?! images with several attributes from resources such as VISA and ImSitu. We then compare several hidden state representations from current state-of-the-art approaches to Grounded Language Learning. By using diagnostic classifiers, we show that current models’ learned representations are not expressive enough to encode object attributes (average F1 of 44.27). In addition, they do not learn strategies nor representations that are robust enough to perform well when novel scenes or objects are involved in gameplay (zero-shot best accuracy 50.06%).

We will demonstrate a deployed conversational AI system that acts as a host of a smart-building on a university campus. The system combines open-domain social conversation with task-based conversation regarding navigation in the building, live resource updates (e.g. available computers) and events in the building. We are able to demonstrate the system on several platforms: Google Home devices, Android phones, and a Furhat robot.

Goal-oriented dialogue systems are now being widely adopted in industry where it is of key importance to maintain a rapid prototyping cycle for new products and domains. Data-driven dialogue system development has to be adapted to meet this requirement — therefore, reducing the amount of data and annotations necessary for training such systems is a central research problem. In this paper, we present the Dialogue Knowledge Transfer Network (DiKTNet), a state-of-the-art approach to goal-oriented dialogue generation which only uses a few example dialogues (i.e. few-shot learning), none of which has to be annotated. We achieve this by performing a 2-stage training. Firstly, we perform unsupervised dialogue representation pre-training on a large source of goal-oriented dialogues in multiple domains, the MetaLWOz corpus. Secondly, at the transfer stage, we train DiKTNet using this representation together with 2 other textual knowledge sources with different levels of generality: ELMo encoder and the main dataset’s source domains. Our main dataset is the Stanford Multi-Domain dialogue corpus. We evaluate our model on it in terms of BLEU and Entity F1 scores, and show that our approach significantly and consistently improves upon a series of baseline models as well as over the previous state-of-the-art dialogue generation model, ZSDG. The improvement upon the latter — up to 10% in Entity F1 and the average of 3% in BLEU score — is achieved using only 10% equivalent of ZSDG’s in-domain training data.

Learning with minimal data is one of the key challenges in the development of practical, production-ready goal-oriented dialogue systems. In a real-world enterprise setting where dialogue systems are developed rapidly and are expected to work robustly for an ever-growing variety of domains, products, and scenarios, efficient learning from a limited number of examples becomes indispensable. In this paper, we introduce a technique to achieve state-of-the-art dialogue generation performance in a few-shot setup, without using any annotated data. We do this by leveraging background knowledge from a larger, more highly represented dialogue source — namely, the MetaLWOz dataset. We evaluate our model on the Stanford Multi-Domain Dialogue Dataset, consisting of human-human goal-oriented dialogues in in-car navigation, appointment scheduling, and weather information domains. We show that our few-shot approach achieves state-of-the art results on that dataset by consistently outperforming the previous best model in terms of BLEU and Entity F1 scores, while being more data-efficient than it by not requiring any data annotation.

We present a new neural architecture for wide-coverage Natural Language Understanding in Spoken Dialogue Systems. We develop a hierarchical multi-task architecture, which delivers a multi-layer representation of sentence meaning (i.e., Dialogue Acts and Frame-like structures). The architecture is a hierarchy of self-attention mechanisms and BiLSTM encoders followed by CRF tagging layers. We describe a variety of experiments, showing that our approach obtains promising results on a dataset annotated with Dialogue Acts and Frame Semantics. Moreover, we demonstrate its applicability to a different, publicly available NLU dataset annotated with domain-specific intents and corresponding semantic roles, providing overall performance higher than state-of-the-art tools such as RASA, Dialogflow, LUIS, and Watson. For example, we show an average 4.45% improvement in entity tagging F-score over Rasa, Dialogflow and LUIS.

The overall objective of ‘social’ dialogue systems is to support engaging, entertaining, and lengthy conversations on a wide variety of topics, including social chit-chat. Apart from raw dialogue data, user-provided ratings are the most common signal used to train such systems to produce engaging responses. In this paper we show that social dialogue systems can be trained effectively from raw unannotated data. Using a dataset of real conversations collected in the 2017 Alexa Prize challenge, we developed a neural ranker for selecting ‘good’ system responses to user utterances, i.e. responses which are likely to lead to long and engaging conversations. We show that (1) our neural ranker consistently outperforms several strong baselines when trained to optimise for user ratings; (2) when trained on larger amounts of data and only using conversation length as the objective, the ranker performs better than the one trained using ratings – ultimately reaching a Precision@1 of 0.87. This advance will make data collection for social conversational agents simpler and less expensive in the future.

We investigate an end-to-end method for automatically inducing task-based dialogue systems from small amounts of unannotated dialogue data. It combines an incremental semantic grammar - Dynamic Syntax and Type Theory with Records (DS-TTR) - with Reinforcement Learning (RL), where language generation and dialogue management are a joint decision problem. The systems thus produced are incremental: dialogues are processed word-by-word, shown previously to be essential in supporting natural, spontaneous dialogue. We hypothesised that the rich linguistic knowledge within the grammar should enable a combinatorially large number of dialogue variations to be processed, even when trained on very few dialogues. Our experiments show that our model can process 74% of the Facebook AI bAbI dataset even when trained on only 0.13% of the data (5 dialogues). It can in addition process 65% of bAbI+, a corpus we created by systematically adding incremental dialogue phenomena such as restarts and self-corrections to bAbI. We compare our model with a state-of-the-art retrieval model, MEMN2N. We find that, in terms of semantic accuracy, the MEMN2N model shows very poor robustness to the bAbI+ transformations even when trained on the full bAbI dataset.

In this paper we present a comparative evaluation of various negotiation strategies within an online version of the game “Settlers of Catan”. The comparison is based on human subjects playing games against artificial game-playing agents (‘bots’) which implement different negotiation dialogue strategies, using a chat dialogue interface to negotiate trades. Our results suggest that a negotiation strategy that uses persuasion, as well as a strategy that is trained from data using Deep Reinforcement Learning, both lead to an improved win rate against humans, compared to previous rule-based and supervised learning baseline dialogue negotiators.

We motivate and describe a new freely available human-human dialogue data set for interactive learning of visually grounded word meanings through ostensive definition by a tutor to a learner. The data has been collected using a novel, character-by-character variant of the DiET chat tool (Healey et al., 2003; anon.) with a novel task, where a Learner needs to learn invented visual attribute words (such as “burchak” for square) from a tutor. As such, the text-based interactions closely resemble face-to-face conversation and thus contain many of the linguistic phenomena encountered in natural, spontaneous dialogue. These include self- and other-correction, mid-sentence continuations, interruptions, turn overlaps, fillers, hedges and many kinds of ellipsis. We also present a generic n-gram framework for building user (i.e. tutor) simulations from this type of incremental dialogue data, which is freely available to researchers. We show that the simulations produce outputs that are similar to the original data (e.g. 78% turn match similarity). Finally, we train and evaluate a Reinforcement Learning dialogue control agent for learning visually grounded word meanings, trained from the BURCHAK corpus. The learned policy shows comparable performance to a rule-based system built previously.

We present an optimised multi-modal dialogue agent for interactive learning of visually grounded word meanings from a human tutor, trained on real human-human tutoring data. Within a life-long interactive learning period, the agent, trained using Reinforcement Learning (RL), must be able to handle natural conversations with human users, and achieve good learning performance (i.e. accuracy) while minimising human effort in the learning process. We train and evaluate this system in interaction with a simulated human tutor, which is built on the BURCHAK corpus – a Human-Human Dialogue dataset for the visual learning task. The results show that: 1) The learned policy can coherently interact with the simulated user to achieve the goal of the task (i.e. learning visual attributes of objects, e.g. colour and shape); and 2) it finds a better trade-off between classifier accuracy and tutoring costs than hand-crafted rule-based policies, including ones with dynamic policies.

Recognition of social signals, coming from human facial expressions or prosody of human speech, is a popular research topic in human-robot interaction studies. There is also a long line of research in the spoken dialogue community that investigates user satisfaction in relation to dialogue characteristics. However, very little research relates a combination of multimodal social signals and language features detected during spoken face-to-face human-robot interaction to the resulting user perception of a robot. In this paper we show how different emotional facial expressions of human users, in combination with prosodic characteristics of human speech and features of human-robot dialogue, correlate with users’ impressions of the robot after a conversation. We find that happiness in the user’s recognised facial expression strongly correlates with likeability of a robot, while dialogue-related features (such as number of human turns or number of sentences per robot utterance) correlate with perceiving a robot as intelligent. In addition, we show that the facial expression emotional features and prosody are better predictors of human ratings related to perceived robot likeability and anthropomorphism, while linguistic and non-linguistic features more often predict perceived robot intelligence and interpretability. As such, these characteristics may in future be used as an online reward signal for in-situ Reinforcement Learning-based adaptive human-robot dialogue systems.

We present VOILA: an optimised, multi-modal dialogue agent for interactive learning of visually grounded word meanings from a human user. VOILA is: (1) able to learn new visual categories interactively from users from scratch; (2) trained on real human-human dialogues in the same domain, and so is able to conduct natural spontaneous dialogue; (3) optimised to find the most effective trade-off between the accuracy of the visual categories it learns and the cost it incurs to users. VOILA is deployed on Furhat, a human-like, multi-modal robot head with back-projection of the face, and a graphical virtual character.

The ultimate goal when building dialogue systems is to satisfy the needs of real users, but quality assurance for dialogue strategies is a non-trivial problem. The applied evaluation metrics and resulting design principles are often obscure, emerge by trial-and-error, and are highly context dependent. This paper introduces data-driven methods for obtaining reliable objective functions for system design. In particular, we test whether an objective function obtained from Wizard-of-Oz (WOZ) data is a valid estimate of real users’ preferences. We test this in a test-retest comparison between the model obtained from the WOZ study and the models obtained when testing with real users. We can show that, despite a low fit to the initial data, the objective function obtained from WOZ data makes accurate predictions for automatic dialogue evaluation, and, when automatically optimising a policy using these predictions, the improvement over a strategy simply mimicking the data becomes clear from an error analysis.