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.

Return Augmented Decision Transformer for Off-Dynamics Reinforcement Learning

We study offline off-dynamics reinforcement learning (RL) to utilize data from an easily accessible source domain to enhance policy learning in a target domain with limited data. Our approach centers on return-conditioned supervised learning (RCSL), particularly focusing on the decision transformer (DT), which can predict actions conditioned on desired return guidance and complete trajectory history. Previous works tackle the dynamics shift problem by augmenting the reward in the trajectory from the source domain to match the optimal trajectory in the target domain. However, this strategy can not be directly applicable in RCSL owing to (1) the unique form of the RCSL policy class, which explicitly depends on the return, and (2) the absence of a straightforward representation of the optimal trajectory distribution. We propose the Return Augmented Decision Transformer (RADT) method, where we augment the return in the source domain by aligning its distribution with that in the target domain. We provide the theoretical analysis demonstrating that the RCSL policy learned from RADT achieves the same level of suboptimality as would be obtained without a dynamics shift. We introduce two practical implementations RADT-DARA and RADT-MV respectively. Extensive experiments conducted on D4RL datasets reveal that our methods generally outperform dynamic programming based methods in off-dynamics RL scenarios.

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.

Personality Analysis from Online Short Video Platforms with Multi-domain Adaptation

Personality analysis from online short videos has gained prominence due to its applications in personalized recommendation systems, sentiment analysis, and human-computer interaction. Traditional assessment methods, such as questionnaires based on the Big Five Personality Framework, are limited by self-report biases and are impractical for large-scale or real-time analysis. Leveraging the rich, multi-modal data present in short videos offers a promising alternative for more accurate personality inference. However, integrating these diverse and asynchronous modalities poses significant challenges, particularly in aligning time-varying data and ensuring models generalize well to new domains with limited labeled data. In this paper, we propose a novel multi-modal personality analysis framework that addresses these challenges by synchronizing and integrating features from multiple modalities and enhancing model generalization through domain adaptation. We introduce a timestamp-based modality alignment mechanism that synchronizes data based on spoken word timestamps, ensuring accurate correspondence across modalities and facilitating effective feature integration. To capture temporal dependencies and inter-modal interactions, we employ Bidirectional Long Short-Term Memory networks and self-attention mechanisms, allowing the model to focus on the most informative features for personality prediction. Furthermore, we develop a gradient-based domain adaptation method that transfers knowledge from multiple source domains to improve performance in target domains with scarce labeled data. Extensive experiments on real-world datasets demonstrate that our framework significantly outperforms existing methods in personality prediction tasks, highlighting its effectiveness in capturing complex behavioral cues and robustness in adapting to new domains.

FastAttention: Extend FlashAttention2 to NPUs and Low-resource GPUs

FlashAttention series has been widely applied in the inference of large language models (LLMs). However, FlashAttention series only supports the high-level GPU architectures, e.g., Ampere and Hopper. At present, FlashAttention series is not easily transferrable to NPUs and low-resource GPUs. Moreover, FlashAttention series is inefficient for multi- NPUs or GPUs inference scenarios. In this work, we propose FastAttention which pioneers the adaptation of FlashAttention series for NPUs and low-resource GPUs to boost LLM inference efficiency. Specifically, we take Ascend NPUs and Volta-based GPUs as representatives for designing our FastAttention. We migrate FlashAttention series to Ascend NPUs by proposing a novel two-level tiling strategy for runtime speedup, tiling-mask strategy for memory saving and the tiling-AllReduce strategy for reducing communication overhead, respectively. Besides, we adapt FlashAttention for Volta-based GPUs by redesigning the operands layout in shared memory and introducing a simple yet effective CPU-GPU cooperative strategy for efficient memory utilization. On Ascend NPUs, our FastAttention can achieve a 10.7$\times$ speedup compared to the standard attention implementation. Llama-7B within FastAttention reaches up to 5.16$\times$ higher throughput than within the standard attention. On Volta architecture GPUs, FastAttention yields 1.43$\times$ speedup compared to its equivalents in \texttt{xformers}. Pangu-38B within FastAttention brings 1.46$\times$ end-to-end speedup using FasterTransformer. Coupled with the propose CPU-GPU cooperative strategy, FastAttention supports a maximal input length of 256K on 8 V100 GPUs. All the codes will be made available soon.

Graph Neural Patching for Cold-Start Recommendations

The cold start problem in recommender systems remains a critical challenge. Current solutions often train hybrid models on auxiliary data for both cold and warm users/items, potentially degrading the experience for the latter. This drawback limits their viability in practical scenarios where the satisfaction of existing warm users/items is paramount. Although graph neural networks (GNNs) excel at warm recommendations by effective collaborative signal modeling, they haven't been effectively leveraged for the cold-start issue within a user-item graph, which is largely due to the lack of initial connections for cold user/item entities. Addressing this requires a GNN adept at cold-start recommendations without sacrificing performance for existing ones. To this end, we introduce Graph Neural Patching for Cold-Start Recommendations (GNP), a customized GNN framework with dual functionalities: GWarmer for modeling collaborative signal on existing warm users/items and Patching Networks for simulating and enhancing GWarmer's performance on cold-start recommendations. Extensive experiments on three benchmark datasets confirm GNP's superiority in recommending both warm and cold users/items.

SecCodePLT: A Unified Platform for Evaluating the Security of Code GenAI

Existing works have established multiple benchmarks to highlight the security risks associated with Code GenAI. These risks are primarily reflected in two areas: a model potential to generate insecure code (insecure coding) and its utility in cyberattacks (cyberattack helpfulness). While these benchmarks have made significant strides, there remain opportunities for further improvement. For instance, many current benchmarks tend to focus more on a model ability to provide attack suggestions rather than its capacity to generate executable attacks. Additionally, most benchmarks rely heavily on static evaluation metrics, which may not be as precise as dynamic metrics such as passing test cases. Conversely, expert-verified benchmarks, while offering high-quality data, often operate at a smaller scale. To address these gaps, we develop SecCodePLT, a unified and comprehensive evaluation platform for code GenAIs' risks. For insecure code, we introduce a new methodology for data creation that combines experts with automatic generation. Our methodology ensures the data quality while enabling large-scale generation. We also associate samples with test cases to conduct code-related dynamic evaluation. For cyberattack helpfulness, we set up a real environment and construct samples to prompt a model to generate actual attacks, along with dynamic metrics in our environment. We conduct extensive experiments and show that SecCodePLT outperforms the state-of-the-art (SOTA) benchmark CyberSecEval in security relevance. Furthermore, it better identifies the security risks of SOTA models in insecure coding and cyberattack helpfulness. Finally, we apply SecCodePLT to the SOTA code agent, Cursor, and, for the first time, identify non-trivial security risks in this advanced coding agent.

DQFormer: Towards Unified LiDAR Panoptic Segmentation with Decoupled Queries

LiDAR panoptic segmentation, which jointly performs instance and semantic segmentation for things and stuff classes, plays a fundamental role in LiDAR perception tasks. While most existing methods explicitly separate these two segmentation tasks and utilize different branches (i.e., semantic and instance branches), some recent methods have embraced the query-based paradigm to unify LiDAR panoptic segmentation. However, the distinct spatial distribution and inherent characteristics of objects(things) and their surroundings(stuff) in 3D scenes lead to challenges, including the mutual competition of things/stuff and the ambiguity of classification/segmentation. In this paper, we propose decoupling things/stuff queries according to their intrinsic properties for individual decoding and disentangling classification/segmentation to mitigate ambiguity. To this end, we propose a novel framework dubbed DQFormer to implement semantic and instance segmentation in a unified workflow. Specifically, we design a decoupled query generator to propose informative queries with semantics by localizing things/stuff positions and fusing multi-level BEV embeddings. Moreover, a query-oriented mask decoder is introduced to decode corresponding segmentation masks by performing masked cross-attention between queries and mask embeddings. Finally, the decoded masks are combined with the semantics of the queries to produce panoptic results. Extensive experiments on nuScenes and SemanticKITTI datasets demonstrate the superiority of our DQFormer framework.

Driving in the Occupancy World: Vision-Centric 4D Occupancy Forecasting and Planning via World Models for Autonomous Driving

World models envision potential future states based on various ego actions. They embed extensive knowledge about the driving environment, facilitating safe and scalable autonomous driving. Most existing methods primarily focus on either data generation or the pretraining paradigms of world models. Unlike the aforementioned prior works, we propose Drive-OccWorld, which adapts a vision-centric 4D forecasting world model to end-to-end planning for autonomous driving. Specifically, we first introduce a semantic and motion-conditional normalization in the memory module, which accumulates semantic and dynamic information from historical BEV embeddings. These BEV features are then conveyed to the world decoder for future occupancy and flow forecasting, considering both geometry and spatiotemporal modeling. Additionally, we propose injecting flexible action conditions, such as velocity, steering angle, trajectory, and commands, into the world model to enable controllable generation and facilitate a broader range of downstream applications. Furthermore, we explore integrating the generative capabilities of the 4D world model with end-to-end planning, enabling continuous forecasting of future states and the selection of optimal trajectories using an occupancy-based cost function. Extensive experiments on the nuScenes dataset demonstrate that our method can generate plausible and controllable 4D occupancy, opening new avenues for driving world generation and end-to-end planning.

MalLight: Influence-Aware Coordinated Traffic Signal Control for Traffic Signal Malfunctions

Urban traffic is subject to disruptions that cause extended waiting time and safety issues at signalized intersections. While numerous studies have addressed the issue of intelligent traffic systems in the context of various disturbances, traffic signal malfunction, a common real-world occurrence with significant repercussions, has received comparatively limited attention. The primary objective of this research is to mitigate the adverse effects of traffic signal malfunction, such as traffic congestion and collision, by optimizing the control of neighboring functioning signals. To achieve this goal, this paper presents a novel traffic signal control framework (MalLight), which leverages an Influence-aware State Aggregation Module (ISAM) and an Influence-aware Reward Aggregation Module (IRAM) to achieve coordinated control of surrounding traffic signals. To the best of our knowledge, this study pioneers the application of a Reinforcement Learning(RL)-based approach to address the challenges posed by traffic signal malfunction. Empirical investigations conducted on real-world datasets substantiate the superior performance of our proposed methodology over conventional and deep learning-based alternatives in the presence of signal malfunction, with reduction of throughput alleviated by as much as 48.6$\%$.