WebRL: Training LLM Web Agents via Self-Evolving Online Curriculum Reinforcement Learning

Large language models (LLMs) have shown remarkable potential as autonomous agents, particularly in web-based tasks. However, existing LLM web agents heavily rely on expensive proprietary LLM APIs, while open LLMs lack the necessary decision-making capabilities. This paper introduces WebRL, a self-evolving online curriculum reinforcement learning framework designed to train high-performance web agents using open LLMs. WebRL addresses three key challenges in building LLM web agents, including the scarcity of training tasks, sparse feedback signals, and policy distribution drift in online learning. Specifically, WebRL incorporates 1) a self-evolving curriculum that generates new tasks from unsuccessful attempts, 2) a robust outcome-supervised reward model (ORM), and 3) adaptive reinforcement learning strategies to ensure consistent improvements. We apply WebRL to transform open Llama-3.1 and GLM-4 models into proficient web agents. On WebArena-Lite, WebRL improves the success rate of Llama-3.1-8B from 4.8% to 42.4%, and from 6.1% to 43% for GLM-4-9B. These open models significantly surpass the performance of GPT-4-Turbo (17.6%) and GPT-4o (13.9%) and outperform previous state-of-the-art web agents trained on open LLMs (AutoWebGLM, 18.2%). Our findings demonstrate WebRL's effectiveness in bridging the gap between open and proprietary LLM-based web agents, paving the way for more accessible and powerful autonomous web interaction systems.

AndroidLab: Training and Systematic Benchmarking of Android Autonomous Agents

Autonomous agents have become increasingly important for interacting with the real world. Android agents, in particular, have been recently a frequently-mentioned interaction method. However, existing studies for training and evaluating Android agents lack systematic research on both open-source and closed-source models. In this work, we propose AndroidLab as a systematic Android agent framework. It includes an operation environment with different modalities, action space, and a reproducible benchmark. It supports both large language models (LLMs) and multimodal models (LMMs) in the same action space. AndroidLab benchmark includes predefined Android virtual devices and 138 tasks across nine apps built on these devices. By using the AndroidLab environment, we develop an Android Instruction dataset and train six open-source LLMs and LMMs, lifting the average success rates from 4.59\% to 21.50\% for LLMs and from 1.93\% to 13.28\% for LMMs. AndroidLab is open-sourced and publicly available at \url{https://github.com/THUDM/Android-Lab}.

AutoGLM: Autonomous Foundation Agents for GUIs

We present AutoGLM, a new series in the ChatGLM family, designed to serve as foundation agents for autonomous control of digital devices through Graphical User Interfaces (GUIs). While foundation models excel at acquiring human knowledge, they often struggle with decision-making in dynamic real-world environments, limiting their progress toward artificial general intelligence. This limitation underscores the importance of developing foundation agents capable of learning through autonomous environmental interactions by reinforcing existing models. Focusing on Web Browser and Phone as representative GUI scenarios, we have developed AutoGLM as a practical foundation agent system for real-world GUI interactions. Our approach integrates a comprehensive suite of techniques and infrastructures to create deployable agent systems suitable for user delivery. Through this development, we have derived two key insights: First, the design of an appropriate "intermediate interface" for GUI control is crucial, enabling the separation of planning and grounding behaviors, which require distinct optimization for flexibility and accuracy respectively. Second, we have developed a novel progressive training framework that enables self-evolving online curriculum reinforcement learning for AutoGLM. Our evaluations demonstrate AutoGLM's effectiveness across multiple domains. For web browsing, AutoGLM achieves a 55.2% success rate on VAB-WebArena-Lite (improving to 59.1% with a second attempt) and 96.2% on OpenTable evaluation tasks. In Android device control, AutoGLM attains a 36.2% success rate on AndroidLab (VAB-Mobile) and 89.7% on common tasks in popular Chinese APPs.

A Survey of AI-Generated Video Evaluation

The growing capabilities of AI in generating video content have brought forward significant challenges in effectively evaluating these videos. Unlike static images or text, video content involves complex spatial and temporal dynamics which may require a more comprehensive and systematic evaluation of its contents in aspects like video presentation quality, semantic information delivery, alignment with human intentions, and the virtual-reality consistency with our physical world. This survey identifies the emerging field of AI-Generated Video Evaluation (AIGVE), highlighting the importance of assessing how well AI-generated videos align with human perception and meet specific instructions. We provide a structured analysis of existing methodologies that could be potentially used to evaluate AI-generated videos. By outlining the strengths and gaps in current approaches, we advocate for the development of more robust and nuanced evaluation frameworks that can handle the complexities of video content, which include not only the conventional metric-based evaluations, but also the current human-involved evaluations, and the future model-centered evaluations. This survey aims to establish a foundational knowledge base for both researchers from academia and practitioners from the industry, facilitating the future advancement of evaluation methods for AI-generated video content.

A Statistical Machine Learning Approach for Adapting Reduced-Order Models using Projected Gaussian Process

The Proper Orthogonal Decomposition (POD) computes the optimal basis modes that span a low-dimensional subspace where the Reduced-Order Models (ROMs) reside. Because a governing equation is often parameterized by a set of parameters, challenges immediately arise when one would like to investigate how systems behave differently over the parameter space (in design, control, uncertainty quantification and real-time operations). In this case, the POD basis needs to be updated so as to adapt ROM that accurately captures the variation of a system's behavior over its parameter space. This paper proposes a Projected Gaussian Process (pGP) and formulate the problem of adapting POD basis as a supervised statistical learning problem, for which the goal is to learn a mapping from the parameter space to the Grassmann Manifold that contains the optimal vector subspaces. A mapping is firstly found between the Euclidean space and the horizontal space of an orthogonal matrix that spans a reference subspace in the Grassmann Manifold. Then, a second mapping from the horizontal space to the Grassmann Manifold is established through the Exponential/Logarithm maps between the manifold and its tangent space. Finally, given a new parameter, the conditional distribution of a vector can be found in the Euclidean space using the Gaussian Process (GP) regression, and such a distribution is projected to the Grassmann Manifold that yields the optimal subspace for the new parameter. The proposed statistical learning approach allows us to optimally estimate model parameters given data (i.e., the prediction/interpolation becomes problem-specific), and quantify the uncertainty associated with the prediction. Numerical examples are presented to demonstrate the advantages of the proposed pGP for adapting POD basis against parameter changes.

DA-Code: Agent Data Science Code Generation Benchmark for Large Language Models

We introduce DA-Code, a code generation benchmark specifically designed to assess LLMs on agent-based data science tasks. This benchmark features three core elements: First, the tasks within DA-Code are inherently challenging, setting them apart from traditional code generation tasks and demanding advanced coding skills in grounding and planning. Second, examples in DA-Code are all based on real and diverse data, covering a wide range of complex data wrangling and analytics tasks. Third, to solve the tasks, the models must utilize complex data science programming languages, to perform intricate data processing and derive the answers. We set up the benchmark in a controllable and executable environment that aligns with real-world data analysis scenarios and is scalable. The annotators meticulously design the evaluation suite to ensure the accuracy and robustness of the evaluation. We develop the DA-Agent baseline. Experiments show that although the baseline performs better than other existing frameworks, using the current best LLMs achieves only 30.5% accuracy, leaving ample room for improvement. We release our benchmark at [https://da-code-bench.github.io](https://da-code-bench.github.io).

YouTube Video Analytics for Patient Engagement: Evidence from Colonoscopy Preparation Videos

Videos can be an effective way to deliver contextualized, just-in-time medical information for patient education. However, video analysis, from topic identification and retrieval to extraction and analysis of medical information and understandability from a patient perspective are extremely challenging tasks. This study demonstrates a data analysis pipeline that utilizes methods to retrieve medical information from YouTube videos on preparing for a colonoscopy exam, a much maligned and disliked procedure that patients find challenging to get adequately prepared for. We first use the YouTube Data API to collect metadata of desired videos on select search keywords and use Google Video Intelligence API to analyze texts, frames and objects data. Then we annotate the YouTube video materials on medical information, video understandability and overall recommendation. We develop a bidirectional long short-term memory (BiLSTM) model to identify medical terms in videos and build three classifiers to group videos based on the levels of encoded medical information and video understandability, and whether the videos are recommended or not. Our study provides healthcare stakeholders with guidelines and a scalable approach for generating new educational video content to enhance management of a vast number of health conditions.

LogicGame: Benchmarking Rule-Based Reasoning Abilities of Large Language Models

Large Language Models (LLMs) have demonstrated notable capabilities across various tasks, showcasing complex problem-solving abilities. Understanding and executing complex rules, along with multi-step planning, are fundamental to logical reasoning and critical for practical LLM agents and decision-making systems. However, evaluating LLMs as effective rule-based executors and planners remains underexplored. In this paper, we introduce LogicGame, a novel benchmark designed to evaluate the comprehensive rule understanding, execution, and planning capabilities of LLMs. Unlike traditional benchmarks, LogicGame provides diverse games that contain a series of rules with an initial state, requiring models to comprehend and apply predefined regulations to solve problems. We create simulated scenarios in which models execute or plan operations to achieve specific outcomes. These game scenarios are specifically designed to distinguish logical reasoning from mere knowledge by relying exclusively on predefined rules. This separation allows for a pure assessment of rule-based reasoning capabilities. The evaluation considers not only final outcomes but also intermediate steps, providing a comprehensive assessment of model performance. Moreover, these intermediate steps are deterministic and can be automatically verified. LogicGame defines game scenarios with varying difficulty levels, from simple rule applications to complex reasoning chains, in order to offer a precise evaluation of model performance on rule understanding and multi-step execution. Utilizing LogicGame, we test various LLMs and identify notable shortcomings in their rule-based logical reasoning abilities.

FMRFT: Fusion Mamba and DETR for Query Time Sequence Intersection Fish Tracking

Growth, abnormal behavior, and diseases of fish can be early detected by monitoring fish tracking through the method of image processing, which is of great significance for factory aquaculture. However, underwater reflections and some reasons with fish, such as the high similarity , rapid swimming caused by stimuli and multi-object occlusion bring challenges to multi-target tracking of fish. To address these challenges, this paper establishes a complex multi-scene sturgeon tracking dataset and proposes a real-time end-to-end fish tracking model, FMRFT. In this model, the Mamba In Mamba (MIM) architecture with low memory consumption is introduced into the tracking algorithm to realize multi-frame video timing memory and fast feature extraction, which improves the efficiency of correlation analysis for contiguous frames in multi-fish video. Additionally, the superior feature interaction and a priori frame processing capabilities of RT-DETR are leveraged to provide an effective tracking algorithm. By incorporating the QTSI query interaction processing module, the model effectively handles occluded objects and redundant tracking frames, resulting in more accurate and stable fish tracking. Trained and tested on the dataset, the model achieves an IDF1 score of 90.3% and a MOTA accuracy of 94.3%. Experimental results demonstrate that the proposed FMRFT model effectively addresses the challenges of high similarity and mutual occlusion in fish populations, enabling accurate tracking in factory farming environments.

Fishers Harvest Parallel Unlearning in Inherited Model Networks

Unlearning in various learning frameworks remains challenging, with the continuous growth and updates of models exhibiting complex inheritance relationships. This paper presents a novel unlearning framework, which enables fully parallel unlearning among models exhibiting inheritance. A key enabler is the new Unified Model Inheritance Graph (UMIG), which captures the inheritance using a Directed Acyclic Graph (DAG).Central to our framework is the new Fisher Inheritance Unlearning (FIUn) algorithm, which utilizes the Fisher Information Matrix (FIM) from initial unlearning models to pinpoint impacted parameters in inherited models. By employing FIM, the FIUn method breaks the sequential dependencies among the models, facilitating simultaneous unlearning and reducing computational overhead. We further design to merge disparate FIMs into a single matrix, synchronizing updates across inherited models. Experiments confirm the effectiveness of our unlearning framework. For single-class tasks, it achieves complete unlearning with 0\% accuracy for unlearned labels while maintaining 94.53\% accuracy for retained labels on average. For multi-class tasks, the accuracy is 1.07\% for unlearned labels and 84.77\% for retained labels on average. Our framework accelerates unlearning by 99\% compared to alternative methods.