Large language models (LLMs) have recently gained much attention in building autonomous agents. However, the performance of current LLM-based web agents in long-horizon tasks is far from optimal, often yielding errors such as repeatedly buying a non-refundable flight ticket. By contrast, humans can avoid such an irreversible mistake, as we have an awareness of the potential outcomes (e.g., losing money) of our actions, also known as the "world model". Motivated by this, our study first starts with preliminary analyses, confirming the absence of world models in current LLMs (e.g., GPT-4o, Claude-3.5-Sonnet, etc.). Then, we present a World-model-augmented (WMA) web agent, which simulates the outcomes of its actions for better decision-making. To overcome the challenges in training LLMs as world models predicting next observations, such as repeated elements across observations and long HTML inputs, we propose a transition-focused observation abstraction, where the prediction objectives are free-form natural language descriptions exclusively highlighting important state differences between time steps. Experiments on WebArena and Mind2Web show that our world models improve agents' policy selection without training and demonstrate our agents' cost- and time-efficiency compared to recent tree-search-based agents.
LineRetriever: Planning-Aware Observation Reduction for Web Agents
INSA Lyon（里昂国立应用科学学院）；Esker（Esker公司）；ServiceNow Research（ServiceNow研究）；Mila AI Institute（Mila人工智能研究所）；McGill University（麦吉尔大学）；Université Claude Bernard Lyon 1（里昂第一大学）；LIRIS（LIRIS实验室）
Free University of Bozen-Bolzano（博尔扎诺自由大学）；Universidad de Chile（智利大学）
The massive semantic data sources linked in the Web of Data give new meaning to old features like navigation; introduce new challenges like semantic specification of Web fragments; and make it possible to specify actions relying on semantic data. In this paper we introduce a declarative language to face these challenges. Based on navigational features, it is designed to specify fragments of the Web of Data and actions to be performed based on these data. We implement it in a centralized fashion, and show its power and performance. Finally, we explore the same ideas in a distributed setting, sh…
University of Illinois at Chicago（伊利诺伊大学芝加哥分校）；Microsoft Research（微软研究院）
AI assistants can now carry out tasks for users by directly interacting with website UIs. Current semantic parsing and slot-filling techniques cannot flexibly adapt to many different websites without being constantly re-trained. We propose FLIN, a natural language interface for web navigation that maps user commands to concept-level actions (rather than low-level UI actions), thus being able to flexibly adapt to different websites and handle their transient nature. We frame this as a ranking problem: given a user command and a webpage, FLIN learns to score the most relevant navigation instruct…
When Users Change Their Mind: Evaluating Interruptible Agents in Long-Horizon Web Navigation
University of Illinois Chicago（伊利诺伊大学芝加哥分校）；McGill University（麦吉尔大学）；MBZUAI（穆罕默德·本·扎耶德人工智能大学）；University of California Santa Barbara（加州大学圣塔芭芭拉分校）；University of Southern California（南加州大学）
Training statistical dialog models in spoken dialog systems (SDS) requires large amounts of annotated data. The lack of scalable methods for data mining and annotation poses a significant hurdle for state-of-the-art statistical dialog managers. This paper presents an approach that directly leverage billions of web search and browse sessions to overcome this hurdle. The key insight is that task completion through web search and browse sessions is (a) predictable and (b) generalizes to spoken dialog task completion. The new method automatically mines behavioral search and browse patterns from we…
Georgia Institute of Technology（佐治亚理工学院）；Oregon State University（俄勒冈州立大学）；Google（谷歌）；Facebook AI Research (FAIR)（Facebook人工智能研究院（FAIR））
Following a navigation instruction such as 'Walk down the stairs and stop at the brown sofa' requires embodied AI agents to ground scene elements referenced via language (e.g. 'stairs') to visual content in the environment (pixels corresponding to 'stairs'). We ask the following question -- can we leverage abundant 'disembodied' web-scraped vision-and-language corpora (e.g. Conceptual Captions) to learn visual groundings (what do 'stairs' look like?) that improve performance on a relatively data-starved embodied perception task (Vision-and-Language Navigation)? Specifically, we develop VLN-BER…
Insight Centre for Data Analytics（数据 analytics 洞察中心）；National University of Ireland Galway（爱尔兰国立高威大学）；University of Kurdistan Hewlêr（埃尔比勒库尔德斯坦大学）；Koya University（科亚大学）
Machine translation has been a major motivation of development in natural language processing. Despite the burgeoning achievements in creating more efficient machine translation systems thanks to deep learning methods, parallel corpora have remained indispensable for progress in the field. In an attempt to create parallel corpora for the Kurdish language, in this paper, we describe our approach in retrieving potentially-alignable news articles from multi-language websites and manually align them across dialects and languages based on lexical similarity and transliteration of scripts. We presen…
Modern urbanization is demanding smarter technologies to improve a variety of applications in intelligent transportation systems to relieve the increasing amount of vehicular traffic congestion and incidents. Existing incident detection techniques are limited to the use of sensors in the transportation network and hang on human-inputs. Despite of its data abundance, social media is not well-exploited in such context. In this paper, we develop an automated traffic alert system based on Natural Language Processing (NLP) that filters this flood of information and extract important traffic-related…
