论文检索

Applications · Language, Speech and Dialog

Xing Han Lù, Zdeněk Kasner, Siva Reddy

We propose the problem of conversational web navigation, where a digital agent controls a web browser and follows user instructions to solve real-world tasks in a multi-turn dialogue fashion. To support this problem, we introduce WEBLINX - a large-scale benchmark of 100K interactions across 2300 expert demonstrations of conversational web navigation. Our benchmark covers a broad range of patterns on over 150 real-world websites and can be used to train and evaluate agents in diverse scenarios. Due to the magnitude of information present, Large Language Models (LLMs) cannot process entire web pages in real-time. To solve this bottleneck, we design a retrieval-inspired model that efficiently prunes HTML pages by ranking relevant elements. We use the selected elements, along with screenshots and action history, to assess a variety of models for their ability to replicate human behavior when navigating the web. Our experiments span from small text-only to proprietary multimodal LLMs. We find that smaller finetuned decoders surpass the best zero-shot LLMs (including GPT-4V), but also larger finetuned multimodal models which were explicitly pretrained on screenshots. However, all finetuned models struggle to generalize to unseen websites. Our findings highlight the need for large multimodal models that can generalize to novel settings.

Shusheng Xu, Wei Fu, Jiaxuan Gao, Wenjie Ye, Weilin Liu, Zhiyu Mei, Guangju Wang, Chao Yu, Yi Wu

Reinforcement Learning from Human Feedback (RLHF) is currently the most widely used method to align large language models (LLMs) with human preferences. Existing RLHF methods can be roughly categorized as either reward-based or reward-free. Novel applications such as ChatGPT and Claude leverage reward-based methods that first learn a reward model and apply actor-critic algorithms, such as Proximal Policy Optimization (PPO). However, in academic benchmarks, state-of-the-art results are often achieved via reward-free methods, such as Direct Preference Optimization (DPO). Is DPO truly superior to PPO? Why does PPO perform poorly on these benchmarks? In this paper, we first conduct both theoretical and empirical studies on the algorithmic properties of DPO and show that DPO may have fundamental limitations. Moreover, we also comprehensively examine PPO and reveal the key factors for the best performances of PPO in fine-tuning LLMs. Finally, we benchmark DPO and PPO across a collection of RLHF testbeds, ranging from dialogue to code generation. Experiment results demonstrate that PPO is able to surpass other alignment methods in all cases and achieve state-of-the-art results in challenging code competitions.

Applications · Language, Speech and Dialog

Hanlei Zhang, Xin Wang, Hua Xu, Qianrui Zhou, Kai Gao, Jianhua Su, jinyue Zhao, Wenrui Li, Yanting Chen

Multimodal intent recognition poses significant challenges, requiring the incorporation of non-verbal modalities from real-world contexts to enhance the comprehension of human intentions. However, most existing multimodal intent benchmark datasets are limited in scale and suffer from difficulties in handling out-of-scope samples that arise in multi-turn conversational interactions. In this paper, we introduce MIntRec2.0, a large-scale benchmark dataset for multimodal intent recognition in multi-party conversations. It contains 1,245 high-quality dialogues with 15,040 samples, each annotated within a new intent taxonomy of 30 fine-grained classes, across text, video, and audio modalities. In addition to more than 9,300 in-scope samples, it also includes over 5,700 out-of-scope samples appearing in multi-turn contexts, which naturally occur in real-world open scenarios, enhancing its practical applicability. Furthermore, we provide comprehensive information on the speakers in each utterance, enriching its utility for multi-party conversational research. We establish a general framework supporting the organization of single-turn and multi-turn dialogue data, modality feature extraction, multimodal fusion, as well as in-scope classification and out-of-scope detection. Evaluation benchmarks are built using classic multimodal fusion methods, ChatGPT, and human evaluators. While existing methods incorporating nonverbal information yield improvements, effectively leveraging context information and detecting out-of-scope samples remains a substantial challenge. Notably, powerful large language models exhibit a significant performance gap compared to humans, highlighting the limitations of machine learning methods in the advanced cognitive intent understanding task. We believe that MIntRec2.0 will serve as a valuable resource, providing a pioneering foundation for research in human-machine conversational interactions, and significantly facilitating related applications.The full dataset and codes are available for use at https://github.com/thuiar/MIntRec2.0.

Applications · Language, Speech and Dialog

Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, zili wang, Steven Yau, Zijuan Lin 等

Recently, remarkable progress has been made on automated problem solving through societies of agents based on large language models (LLMs). Previous LLM-based multi-agent systems can already solve simple dialogue tasks. More complex tasks, however, face challenges through logic inconsistencies due to cascading hallucinations caused by naively chaining LLMs. Here we introduce MetaGPT, an innovative meta-programming framework incorporating efficient human workflows into LLM-based multi-agent collaborations. MetaGPT encodes Standardized Operating Procedures (SOPs) into prompt sequences for more streamlined workflows, thus allowing agents with human-like domain expertise to verify intermediate results and reduce errors. MetaGPT utilizes an assembly line paradigm to assign diverse roles to various agents, efficiently breaking down complex tasks into subtasks involving many agents working together. On collaborative software engineering benchmarks, MetaGPT generates more coherent solutions than previous chat-based multi-agent systems.

Deep Learning · Large Language Models

Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, zili wang, Steven Yau, Zijuan Lin 等

Recently, remarkable progress has been made on automated problem solving through societies of agents based on large language models (LLMs). Previous LLM-based multi-agent systems can already solve simple dialogue tasks. More complex tasks, however, face challenges through logic inconsistencies due to cascading hallucinations caused by naively chaining LLMs. Here we introduce MetaGPT, an innovative meta-programming framework incorporating efficient human workflows into LLM-based multi-agent collaborations. MetaGPT encodes Standardized Operating Procedures (SOPs) into prompt sequences for more streamlined workflows, thus allowing agents with human-like domain expertise to verify intermediate results and reduce errors. MetaGPT utilizes an assembly line paradigm to assign diverse roles to various agents, efficiently breaking down complex tasks into subtasks involving many agents working together. On collaborative software engineering benchmarks, MetaGPT generates more coherent solutions than previous chat-based multi-agent systems.

Miscellaneous Aspects of Machine Learning

Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang 等

This paper introduces InternVid, a large-scale video-centric multimodal dataset that enables learning powerful and transferable video-text representations for multimodal understanding and generation. InternVid contains over 7 million videos lasting nearly 760K hours, yielding 234M video clips accompanied by detailed descriptions of total 4.1B words. Our core contribution is to develop a scalable approach to autonomously build a high-quality video-text dataset with large language models (LLM), thereby showcasing its efficacy in learning video-language representation at scale. Specifically, we utilize a multi-scale approach to generate video-related descriptions. Furthermore, we introduce ViCLIP, a video-text representation learning model based on ViT-L. Learned on InternVid via contrastive learning, this model demonstrates leading zero-shot action recognition and competitive video retrieval performance. Beyond basic video understanding tasks like recognition and retrieval, our dataset and model have broad applications. They are particularly beneficial for generating interleaved video-text data for learning a video-centric dialogue system, advancing video-to-text and text-to-video generation research. These proposed resources provide a tool for researchers and practitioners interested in multimodal video understanding and generation.

Deep Learning · Large Language Models

Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, Mike Lewis

Deploying Large Language Models (LLMs) in streaming applications such as multi-round dialogue, where long interactions are expected, is urgently needed but poses two major challenges.Firstly, during the decoding stage, caching previous tokens' Key and Value states (KV) consumes extensive memory.Secondly, popular LLMs cannot generalize to longer texts than the training sequence length.Window attention, where only the most recent KVs are cached, is a natural approach --- but we show that it fails when the text length surpasses the cache size.We observe an interesting phenomenon, namely attention sink, that keeping the KV of initial tokens will largely recover the performance of window attention. In this paper, we first demonstrate that the emergence of attention sink is due to the strong attention scores towards initial tokens as a ``sink'' even if they are not semantically important.Based on the above analysis, we introduce StreamingLLM, an efficient framework that enables LLMs trained with a finite length attention window to generalize to infinite sequence length without any fine-tuning.We show that StreamingLLM can enable Llama-2, MPT, Falcon, and Pythia to perform stable and efficient language modeling with up to 4 million tokens and more.In addition, we discover that adding a placeholder token as a dedicated attention sink during pre-training can further improve streaming deployment. In streaming settings, StreamingLLM outperforms the sliding window recomputation baseline by up to 22.2$\times$ speedup.Code and datasets are provided at https://github.com/mit-han-lab/streaming-llm.

Hongcheng Guo, Jian Yang, Jiaheng Liu, Liqun Yang, Linzheng Chai, Jiaqi Bai, Junran Peng, Xiaorong Hu, Chao Chen, Dongfeng Zhang 等

With the rapid advancement of IT operations, managing and analyzing large data volumes efficiently for practical applications has become increasingly critical. Natural Language Processing (NLP) techniques have demonstrated remarkable capabilities in various tasks, including named entity recognition, machine translation, and dialogue systems. Recently, Large Language Models (LLMs) have achieved significant improvements across various domain-specific areas. However, there is a noticeable gap in the development of specialized Large Language Models (LLMs) tailored for IT operations. In this paper, we introduce the OWL, a large language model trained on our constructed Owl-Instruct with a wide range of IT-related information. Specifically, limited by the maximum input length, we propose the \textbf{H}omogeneous \textbf{M}arkov \textbf{C}ontext \textbf{E}xtension method (HMCE). The mixture-of-adapter strategy is leveraged to improve the parameter-efficient tuning across different domains or tasks.Further, we evaluate the performance of OWL on the Owl-Bench established by us and open IT-related benchmarks. OWL demonstrates superior performance results on IT tasks, which outperforms existing models by significant margins. Moreover, we hope that the findings of our work will provide more insights to revolutionize the techniques of IT operations with specialized LLMs.

Deep Learning · Large Language Models

David Vazquez, Stefania Raimondo, Christopher Pal, Pau Rodriguez, Amine El hattami, Issam Laradji

Text-based dialogues are now widely used to solve real-world problems. In cases where solution strategies are already known, they can sometimes be codified into workflows and used to guide humans or artificial agents through the task of helping clients. We introduce a new problem formulation that we call Workflow Discovery (WD) in which we are interested in the situation where a formal workflow may not yet exist. Still, we wish to discover the set of actions that have been taken to resolve a particular problem. We also examine a sequence-to-sequence (Seq2Seq) approach for this novel task. We present experiments where we extract workflows from dialogues in the Action-Based Conversations Dataset (ABCD). Since the ABCD dialogues follow known workflows to guide agents, we can evaluate our ability to extract such workflows using ground truth sequences of actions. We propose and evaluate an approach that conditions models on the set of possible actions, and we show that using this strategy, we can improve WD performance. Our conditioning approach also improves zero-shot and few-shot WD performance when transferring learned models to unseen domains within and across datasets. Further, on ABCD a modified variant of our Seq2Seq method achieves state-of-the-art performance on related but different problems of Action State Tracking (AST) and Cascading Dialogue Success (CDS) across many evaluation metrics.

Joey Hong, Anca Dragan, Sergey Levine

Offline reinforcement learning (RL) can in principle synthesize more optimal behavior from a dataset consisting only of suboptimal trials. One way that this can happen is by "stitching" together the best parts of otherwise suboptimal trajectories that overlap on similar states, to create new behaviors where each individual state is in-distribution, but the overall returns are higher. However, in many interesting and complex applications, such as autonomous navigation and dialogue systems, the state is partially observed. Even worse, the state representation is unknown or not easy to define. In such cases, policies and value functions are often conditioned on observation histories instead of states. In these cases, it is not clear if the same kind of "stitching" is feasible at the level of observation histories, since two different trajectories would always have different histories, and thus "similar states" that might lead to effective stitching cannot be leveraged. Theoretically, we show that standard offline RL algorithms conditioned on observation histories suffer from poor sample complexity, in accordance with the above intuition. We then identify sufficient conditions under which offline RL can still be efficient -- intuitively, it needs to learn a compact representation of history comprising only features relevant for action selection. We introduce a bisimulation loss that captures the extent to which this happens, and propose that offline RL can explicitly optimize this loss to aid worst-case sample complexity. Empirically, we show that across a variety of tasks either our proposed loss improves performance, or the value of this loss is already minimized as a consequence of standard offline RL, indicating that it correlates well with good performance.

Deep Learning · Large Language Models

Hao Liu, Carmelo Sferrazza, Pieter Abbeel

Learning from human preferences is important for language models to match human needs and to align with human and social values. Prior works have achieved remarkable successes by learning from human feedback to understand and follow instructions. Nonetheless, these methods are either founded on hand-picked model generations that are favored by human annotators, rendering them inefficient in terms of data utilization and challenging to apply in general, or they depend on reinforcement learning, which often suffers from imperfect reward functions and relies on extremely challenging optimizations. In this work, we propose a novel technique, Chain of Hindsight, that is easy to optimize and can learn from any form of feedback, regardless of its polarity. Our idea is inspired by how humans learn from extensive feedback presented in the form of languages. We convert all types of feedback into sequences of sentences, which are then used to fine-tune the model, allowing us to take advantage of the language comprehension capabilities of language models.We condition the model on a sequence of model generations paired with feedback. By doing so, the model is trained to generate outputs based on feedback, while learning to identify and correct negative attributes or errors. Applying our method to large language models, we observed that Chain of Hindsight significantly surpasses previous methods in aligning language models with human preferences. We report significant improvements on summarization and dialogue benchmarks, with our approach markedly preferred in human evaluations.

Applications · Language, Speech and Dialog

Yang Deng, Wenxuan Zhang, Wai Lam, See-Kiong Ng, Tat-Seng Chua

Proactive dialogues serve as a practical yet challenging dialogue problem in the era of large language models (LLMs), where the dialogue policy planning is the key to improving the proactivity of LLMs. Most existing studies enable the dialogue policy planning of LLMs using various prompting schemes or iteratively enhance this capability in handling the given case with verbal AI feedback. However, these approaches are either bounded by the policy planning capability of the frozen LLMs or hard to be transferred to new cases. In this work, we introduce a new dialogue policy planning paradigm to strategize LLMs for proactive dialogue problems with a tunable language model plug-in as a plug-and-play dialogue policy planner, named PPDPP. Specifically, we develop a novel training framework to facilitate supervised fine-tuning over available human-annotated data as well as reinforcement learning from goal-oriented AI feedback with dynamic interaction data collected by the LLM-based self-play simulation. In this manner, the LLM-powered dialogue agent can not only be generalized to different cases after the training, but also be applicable to different applications by just substituting the learned plug-in. In addition, we propose to evaluate the policy planning capability of dialogue systems under the interactive setting. Experimental results demonstrate that PPDPP consistently and substantially outperforms existing approaches on three different proactive dialogue applications, including negotiation, emotional support, and tutoring dialogues.

Ioana Croitoru, Simion-Vlad Bogolin, Samuel Albanie, Yang Liu, Zhaowen Wang, Seunghyun Yoon, Franck Dernoncourt, Hailin Jin, Trung Bui

Tutorial videos play an increasingly important role in professional development and self-directed education. For users to realise the full benefits of this medium, tutorial videos must be efficiently searchable. In this work, we focus on the task of moment detection, in which the goal is to localise the temporal window where a given event occurs within a given tutorial video. Prior work on moment detection has focused primarily on short videos (typically on videos shorter than three minutes). However, many tutorial videos are substantially longer (stretching to hours in duration), presenting significant challenges for existing moment detection approaches. To study this problem, we propose the first dataset of untrimmed, long-form tutorial videos for the task of Moment Detection called the Behance Moment Detection (BMD) dataset. BMD videos have an average duration of over one hour and are characterised by slowly evolving visual content and wide-ranging dialogue. To meet the unique challenges of this dataset, we propose a new framework, LongMoment-DETR, and demonstrate that it outperforms strong baselines. Additionally, we introduce a variation of the dataset that contains YouTube Chapter annotations and show that the features obtained by our framework can be successfully used to boost the performance on the task of chapter detection. Code and data can be found at https://github.com/ioanacroi/longmoment-detr.

Seungju Han, Jack Hessel, Nouha Dziri, Yejin Choi, Youngjae Yu

Visual information is central to conversation: body gestures and physical behaviour, for example, contribute to meaning that transcends words alone. To date, however, most neural conversational models are limited to just text. We introduce CHAMPAGNE, a generative model of conversations that can account for visual contexts. To train CHAMPAGNE, we collect and release YTD-18M, a large-scale corpus of 18M video-based dialogues. YTD-18M is constructed from web videos: crucial to our data collection pipeline is a pretrained language model that converts error-prone automatic transcripts to a cleaner dialogue format while maintaining meaning. Human evaluation reveals that YTD-18M is more sensible and specific than prior resources (MMDialog, 1M dialogues), while maintaining visual-groundedness. Experiments demonstrate that 1) CHAMPAGNE learns to conduct conversation from YTD-18M; and 2) when fine-tuned, it achieves state-of-the-art results on four vision-language tasks focused on real-world conversations. We release data, models, and code.

Kaiqu Liang, Samuel Albanie

To date, the majority of video retrieval systems have been optimized for a "single-shot" scenario in which the user submits a query in isolation, ignoring previous interactions with the system. Recently, there has been renewed interest in interactive systems to enhance retrieval, but existing approaches are complex and deliver limited gains in performance. In this work, we revisit this topic and propose several simple yet effective baselines for interactive video retrieval via question-answering. We employ a VideoQA model to simulate user interactions and show that this enables the productive study of the interactive retrieval task without access to ground truth dialogue data. Experiments on MSR-VTT, MSVD, and AVSD show that our framework using question-based interaction significantly improves the performance of text-based video retrieval systems. Code is available at https://github.com/kevinliang888/IVR-QA-baselines.

Yibo Cui, Liang Xie, Yakun Zhang, Meishan Zhang, Ye Yan, Erwei Yin

Cross-modal alignment is one key challenge for Vision-and-Language Navigation (VLN). Most existing studies concentrate on mapping the global instruction or single sub-instruction to the corresponding trajectory. However, another critical problem of achieving fine-grained alignment at the entity level is seldom considered. To address this problem, we propose a novel Grounded Entity-Landmark Adaptive (GELA) pre-training paradigm for VLN tasks. To achieve the adaptive pre-training paradigm, we first introduce grounded entity-landmark human annotations into the Room-to-Room (R2R) dataset, named GEL-R2R. Additionally, we adopt three grounded entity-landmark adaptive pre-training objectives: 1) entity phrase prediction, 2) landmark bounding box prediction, and 3) entity-landmark semantic alignment, which explicitly supervise the learning of fine-grained cross-modal alignment between entity phrases and environment landmarks. Finally, we validate our model on two downstream benchmarks: VLN with descriptive instructions (R2R) and dialogue instructions (CVDN). The comprehensive experiments show that our GELA model achieves state-of-the-art results on both tasks, demonstrating its effectiveness and generalizability.

Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, Andrew Zisserman

Audio Description (AD) is the task of generating descriptions of visual content, at suitable time intervals, for the benefit of visually impaired audiences. For movies, this presents notable challenges -- AD must occur only during existing pauses in dialogue, should refer to characters by name, and ought to aid understanding of the storyline as a whole. To this end, we develop a new model for automatically generating movie AD, given CLIP visual features of the frames, the cast list, and the temporal locations of the speech; addressing all three of the `who', `when', and `what' questions: (i) who -- we introduce a character bank consisting of the character's name, the actor that played the part, and a CLIP feature of their face, for the principal cast of each movie, and demonstrate how this can be used to improve naming in the generated AD; (ii) when -- we investigate several models for determining whether an AD should be generated for a time interval or not, based on the visual content of the interval and its neighbours; and (iii) what -- we implement a new vision-language model for this task, that can ingest the proposals from the character bank, whilst conditioning on the visual features using cross-attention, and demonstrate how this improves over previous architectures for AD text generation in an apples-to-apples comparison.

Qiya Song, Renwei Dian, Bin Sun 0001, Jie Xie 0002, Shutao Li 0001

Understanding and elucidating human behavior across diverse scenarios represents a pivotal research challenge in pursuing seamless human-computer interaction. However, previous research on multi-participant dialogues has mostly relied on proprietary datasets, which are not standardized and openly accessible. To propel advancements in this domain, the MultiMediate'23 Challenge presents two sub-challenges: Eye contact detection and Next speaker prediction, aiming to foster a comprehensive understanding of multi-participant behavior. To tackle these challenges, we propose a multi-scale conformer fusion network (MSCFN) for enhancing the perception of multi-participant group behaviors. The conformer block combines the strengths of transformers and convolution networks to facilitate the establishment of global and local contextual relationships between sequences. Then the output features from all Conformer blocks are concatenated to fusion multi-scale representations. Our proposed method was evaluated using the officially provided dataset, and it achieves the best and second best performance in next speaker prediction and gaze detection tasks of MultiMediate'23, respectively.

Wenzhe Du, Haoyang Su, Cam-Tu Nguyen, Jian Sun

Multimodal Conversational Recommendation aims to find appropriate products based on a multi-turn dialogue, where user requests and products can be presented in both visual and textual modalities. While previous studies have focused on understanding user preferences from conversational contexts, the task of product modeling has been relatively unexplored. This study targets to fill this gap and demonstrates that information from multiple product views and cross-view interactions are essential for recommendation, along with dialog information. To this end, a product image is first encoded using a gated multi-view image encoder, and representations for the global and local views are obtained. On the textual side, two views are considered: the structure view (product attributes) and the sequence view (product description/reviews). Two forms of inter-modal interactions for product representation are then modeled: interactions between the global image view and the textual structure view, and interactions between the local image view and the textual sequence view. Furthermore, the representation is enhanced to attend to the latest user request in the dialog context, resulting in query-aware product representation. The experimental results indicate that our method, named Enteract, achieves state-of-the-art performance on two well-known datasets (MMD and SIMMC).

Yayue Deng, Jinlong Xue, Fengping Wang, Yingming Gao, Ya Li 0001

Conversational Speech Synthesis (CSS) aims to produce speech appropriate for oral communication. However, the complexity of context dependency modeling poses significant challenges in the field of CSS, especially the mutual psychological influence between interlocutors. Previous studies have verified that prior commonsense knowledge helps machines understand subtle psychological information (e.g., feelings and intentions) in spontaneous oral dialogues. Therefore, to enhance context understanding and improve the naturalness of synthesized speech, we propose a novel conversational speech synthesis system (CMCU-CSS) that incorporates the Commonsense-based Multi-modal Context Understanding (CMCU) module to model the dynamic emotional interaction among interlocutors. Specifically, we first utilize three implicit states (intent state, internal state and external state) in CMCU to model the context dependency between inter/intra speakers with the help of commonsense knowledge. Furthermore, we infer emotion vectors from the fusion of these implicit states and multi-modal features to enhance the emotion discriminability of synthesized speech. This is the first attempt to combine commonsense knowledge with conversational speech synthesis, and its effect in terms of emotion discriminability of synthetic speech is evaluated by emotion recognition in conversation task. The results of subjective and objective evaluations demonstrate that the CMCU-CSS model achieves more natural speech with context-appropriate emotion and is equipped with the best emotion discriminability, surpassing that of other conversational speech synthesis models.