We present GLM-5V-Turbo, a step toward native foundation models for multimodal agents. As foundation models are increasingly deployed in real environments, agentic capability depends not only on language reasoning, but also on the ability to perceive, interpret, and act over heterogeneous contexts such as images, videos, webpages, documents, GUIs. GLM-5V-Turbo is built around this objective: multimodal perception is integrated as a core component of reasoning, planning, tool use, and execution, rather than as an auxiliary interface to a language model. This report summarizes the main improvements behind GLM-5V-Turbo across model design, multimodal training, reinforcement learning, toolchain expansion, and integration with agent frameworks. These developments lead to strong performance in multimodal coding, visual tool use, and framework-based agentic tasks, while preserving competitive text-only coding capability. More importantly, our development process offers practical insights for building multimodal agents, highlighting the central role of multimodal perception, hierarchical optimization, and reliable end-to-end verification.

Computer use agents (CUAs) are emerging as a powerful interface for automating complex digital workflows through visual perception and GUI execution. Online reinforcement learning with verifiable rewards (RLVR) has emerged as a key direction for scaling their capabilities. However, this paradigm is bottlenecked by verifiable data scarcity and online RL inefficiency. To break these barriers, we introduce ScaleCUA, a unified framework that scales online RL for CUAs via verifiable task synthesis and efficient training. At the data level, we design VeriGen, an end-to-end framework for generating verifiable RL tasks through iterative docker interactions and a multi-agent feedback loop. Scaled to 100+ concurrent agent workers via a shared docker interaction probe, this pipeline produces 24K+ verifiable tasks and nearly 3K high-quality RL tasks. To maximize sample efficiency, we propose Frontier Sampling, which tracks per-task capability and allocates rollouts to the current learning frontier. On the training side, we further design Visual Context Segmentation, a sliding window over recent visual context that balances rollout and training-engine pressure, yielding a 2.83x training speedup over step-wise decomposition. Together, ScaleCUA achieves 68.7

We introduce ComputerRL, a framework for autonomous desktop intelligence that enables agents to operate complex digital workspaces skillfully. ComputerRL features the API-GUI paradigm, which unifies programmatic API calls and direct GUI interaction to address the inherent mismatch between machine agents and human-centric desktop environments. Scaling end-to-end RL training is crucial for improvement and generalization across diverse desktop tasks; however, it remains challenging due to environmental inefficiency and instability during extended training. To support scalable and robust training, we develop a distributed RL infrastructure capable of orchestrating thousands of parallel virtual desktop environments to accelerate large-scale online RL. Furthermore, we propose Entropulse, a training strategy that alternates reinforcement learning with supervised fine-tuning, effectively mitigating entropy collapse during extended training runs. We employ ComputerRL on open models GLM-4-9B-0414 and GLM-4.1V-9B-Thinking, and evaluate them on the OSWorld benchmark. The GLM-ComputerRL-9B achieves a new state-of-the-art accuracy of 48.9%, demonstrating significant improvements for general agents in desktop automation. Our code and demos are available at https://computer-rl.vercel.app/.

Visual generative models have achieved remarkable progress in synthesizing photorealistic images and videos, yet aligning their outputs with human preferences across critical dimensions remains a persistent challenge. Though reinforcement learning from human feedback offers promise for preference alignment, existing reward models for visual generation face limitations, including black-box scoring without interpretability and potentially resultant unexpected biases. We present VisionReward, a general framework for learning human visual preferences in both image and video generation. Specifically, we employ a hierarchical visual assessment framework to capture fine-grained human preferences, and leverages linear weighting to enable interpretable preference learning. Furthermore, we propose a multi-dimensional consistent strategy when using VisionReward as a reward model during preference optimization for visual generation. Experiments show that VisionReward can significantly outperform existing image and video reward models on both machine metrics and human evaluation. Notably, VisionReward surpasses VideoScore by 17.2% in preference prediction accuracy, and text-to-video models with VisionReward achieve a 31.6% higher pairwise win rate compared to the same models using VideoScore.

Reinforcement learning (RL) for large language models is moving toward multi-turn agentic workloads, where rollout tasks repeatedly pause for external environments, resume with growing contexts, and finish at highly variable times. In this setting, RL training goodput, measured by training throughput, matters more than raw GPU occupancy: GPU waiting and repeated prefill recomputation are pure overhead. We present TideRL, a readiness-aware elastic RL system with Continuous Task Batching, Resource-Aware Ref-Actor Pipelining, and Elastic Resource Scaling. CTB preserves useful rollout state, RA^2P selects between decoupled streaming and colocated aggregation from the ready backlog and arrival interval, and ERS moves ranks between rollout and training using the same readiness signals. Across text-only and multi-modal agentic workloads, TideRL improves RL training goodput by up to 5.6× over synchronous baselines and over 33

Large language models (LLMs) conventionally represent text as sequences of discrete tokens, making long-context scaling largely a matter of processing more tokens more efficiently.We instead explore a complementary direction: increasing how much original context each token represents.To this end, we introduce Glyph, a framework that renders long texts into compact visual pages and processes them with a vision-language model (VLM), allowing a fixed context window to cover substantially more text.To make visual compression practical, Glyph combines continual pre-training on rendered long-text data, an LLM-driven genetic search to identify rendering configurations that balance compression and task performance, and post-training with supervised fine-tuning and reinforcement learning.Across multiple long-context benchmarks, Glyph achieves 3–4× token compression while maintaining performance comparable to strong text-only LLMs such as Qwen3-8B, with over 4× faster prefilling and decoding and 2× faster supervised fine-tuning.Under more aggressive compression, a VLM with a 128K context window can handle tasks that would otherwise require up to 1M input tokens.Our code and model are released at https://github.com/thu-coai/Glyph.

Generalist robot policies aim to map multimodal observations and linguistic task instructions to actions across diverse tasks. However, existing methods typically represent the future as a fixed, short video-action chunk. This short-term future captures local scene evolution for action execution, but it does not explicitly describe the stage-level future that specifies how a task should progress from its current stage to the next. We therefore distinguish two complementary futures for robot manipulation: a short-term physical future to capture local scene evolution and a stage-level semantic future to represent task progress. We introduce StageWAM, which augments a Motus-based World Action Model (WAM) with Stage-JEPA, a goal-conditioned Joint-Embedding Predictive Architecture (JEPA) predictor. Given the current observation and task instruction, Stage-JEPA uses a frozen V-JEPA2 encoder to extract the current-state representation and predicts the latent target of the next inferred stage. Across 50 RoboTwin 2.0 tasks in clean and randomized environments, StageWAM achieves 90.25

Large Language Models have shown remarkable potential as autonomous agents, but their effectiveness in knowledge-intensive tasks remains limited by passive knowledge utilization. We introduce KARL (Knowledge-Augmented Reinforcement Learning), a framework that enables LLM agents to dynamically explore structured knowledge sources through multi-turn interactions. Unlike existing retrieval-augmented approaches, KARL empowers agents to proactively decide when and what knowledge to acquire during task execution. Our framework incorporates online reinforcement learning with curiosity-driven reward shaping, explicitly incentivizing knowledge exploration while optimizing tool-use behaviors end-to-end. Extensive evaluation across six structured knowledge benchmarks demonstrates that KARL achieves state-of-the-art performance, with our Qwen2.5-14B-based agent significantly outperforming GPT-4o, Claude-4, and o4-mini on both knowledge graph and database tasks.Source code is available at https://github.com/THUDM/KARL.

Building general-purpose graphical user interface (GUI) agents has become increasingly promising with the progress in vision language models. However, developing effective mobile GUI agents with reinforcement learning (RL) remains challenging due to the heavy-tailed distribution of task difficulty and the inefficiency of large-scale environment sampling. We present an online agentic reinforcement learning framework MOBILERL to enhance GUI agents in mobile environments. Its core component is the Difficulty-Adaptive GRPO (ADAGRPO) algorithm. In ADAGRPO, we design difficulty-adaptive positive replay and failure curriculum filtering to adapt the model to different task difficulties. We introduce the shortest-path reward adjustment strategy to reshape rewards concerning the task length in multi-turn agentic tasks. Those strategies jointly stabilize RL training, improve sample efficiency, and generate strong performance across diverse mobile apps and tasks. We apply MOBILERL to two open models (Qwen2.5-VL-7B-Instruct and GLM-4.1V-9B-Base). The resultant MOBILERL-9B model achieves state-of-the-art results in terms of success rates on both AndroidWorld (80.2%) and Android-Lab (53.6%). The MOBILERL framework is open-sourced at https://anonymous.4open.science/r/MobileRL-iclr-4513.

Large language models (LLMs) have achieved remarkable proficiency in understanding and generating human natural languages, mainly owing to the “scaling law” that optimizes relationships among language modeling loss, model parameters, and pre-trained tokens. However, with the exhaustion of high-quality internet corpora and increasing computational demands, the sustainability of pre-training scaling needs to be addressed. This paper presents a comprehensive survey of post-training scaling, an emergent paradigm aiming to relieve the limitations of traditional pre-training by focusing on the alignment phase, which traditionally accounts for a minor fraction of the total training computation. Our survey categorizes post-training scaling into three key methodologies: Supervised Fine-tuning (SFT), Reinforcement Learning from Feedback (RLxF), and Test-time Compute (TTC). We provide an in-depth analysis of the motivation behind post-training scaling, the scalable variants of these methodologies, and a comparative discussion against traditional approaches. By examining the latest advancements, identifying promising application scenarios, and highlighting unresolved issues, we seek a coherent understanding and map future research trajectories in the landscape of post-training scaling for LLMs.

We present WebGLM, an enhanced Large Language Model (LLM)-based retrieval question-answering system based on the ChatGLM3-6B, offering significant improvements over previous systems. We aim to augment a pre-trained LLM with web search and reliable retrieval capabilities while being efficient for real-world deployments. Leveraging LLM’s in-context learning ability and a robust filter strategy, we create a high-quality training dataset and address the hallucination issue with a self-check mechanism. Our base model, ChatGLM3-6B, excels in extracting critical information and generating desired responses. We tackle the decline in retrieval effectiveness for complex queries with a keywording technique and incorporate more web content for references. We align with user preferences by training a human preference-aware scorer and employing DPO training for direct alignment. Extensive experiments, including human evaluations and the Turing test, demonstrate WebGLM’s superior performance against leading web-enhanced question-answering systems, significantly enhancing performance and efficiency. The code, demo, and data are at https://github.com/THUDM/WebGLM .

3D Multi-modal Large Language Models (MLLMs) still lag behind their 2D peers, largely because large-scale, high-quality 3D scene-dialogue datasets remain scarce. Prior efforts hinge on expensive human annotation and leave two key ambiguities unresolved: viewpoint ambiguity, where spatial language presumes unknown camera poses, and object referring ambiguity, where non-exclusive descriptions blur the line between targets and distractors. We therefore present a fully automated pipeline that converts raw 3D scans into unambiguous, high-quality dialogue data at a fraction of the previous cost. By synergizing rule-based constraints with 2D MLLMs and LLMs, the pipeline enables controllable, scalable generation without human intervention. The pipeline comprises four stages: (1) meta-annotation collection harvesting object-, frame-, and scene-level captions, (2) scene graph construction with relation correction to capture proximal object relations, (3) discriminative object referring that generates exclusive and compact descriptions, and (4) multi-task data generation synthesizing diverse dialogues. Our pipeline systematically mitigates inherent flaws in source datasets and produces the final Disc3D dataset, over 2 million samples in 25K hybrid 3D scenes, spanning scene, view, and object captioning, visual grounding, and five object-centric QA tasks. Extensive experiments demonstrate that training with Disc3D yields consistent, significant improvements on both public benchmarks and our multifaceted Disc3D-QA tasks. Code, data, and models will be publicly available.

Agentic workflows have become the dominant paradigm for building complex AI systems, orchestrating specialized components, such as planning, reasoning, action execution, and reflection, to tackle sophisticated real-world tasks. However, systematically analyzing and optimizing these workflows remains challenging due to intricate component interdependencies and the lack of principled attribution methods. In this work, we introduce ShapleyFlow, the first framework that employs cooperative game theory to analyze and optimize agentic workflows. By applying the Shapley value to evaluate all possible component configurations, ShapleyFlow enables fine-grained attribution of each component's contribution and facilitates the identification of task-specific optimal configurations. Through a constructed dataset evaluated across 7 scenarios, such as navigation, math and OS, we demonstrate 3 key contributions: (1) Theoretical Framework: a principled game-theoretic approach for the attribution of contributions in agentic workflows. (2) Optimal Workflow Discovery: ShapleyFlow identifies task-specific component configurations that consistently outperform workflows relying on a single LLM across all tested tasks. (3) Comprehensive Analysis: we construct and analyze over 1,500 tasks, providing actionable insights and design guidelines for optimizing workflows across multiple domains.

We introduce AIBrix, a cloud-native, open-source framework designed to optimize and simplify large-scale LLM deployment in cloud environments. Unlike traditional cloud-native stacks, AIBrix follows a co-design philosophy, ensuring every layer of the infrastructure is purpose-built for seamless integration with inference engines like vLLM. AIBrix introduces several key innovations to reduce inference costs and enhance performance including high-density LoRA management for dynamic adapter scheduling, LLM-specific autoscalers, and prefix-aware, load-aware routing. To further improve efficiency, AIBrix incorporates a distributed KV cache, boosting token reuse across nodes, leading to a 50 inference latency. AIBrix also supports unified AI runtime which streamlines model management while maintaining vendor-agnostic engine compatibility. For large-scale multi-node inference, AIBrix employs hybrid orchestration – leveraging Kubernetes for coarse-grained scheduling and Ray for fine-grained execution – to balance efficiency and flexibility. Additionally, an SLO-driven GPU optimizer dynamically adjusts resource allocations, optimizing heterogeneous serving to maximize cost efficiency while maintaining service guarantees. Finally, AIBrix enhances system reliability with AI accelerator diagnostic tools, enabling automated failure detection and mock-up testing to improve fault resilience. AIBrix is available at https://github.com/vllm-project/aibrix.

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.

Large Multimodal Models (LMMs) have ushered in a new era in artificial intelligence, merging capabilities in both language and vision to form highly capable \textbf{Visual Foundation Agents} that are postulated to excel across a myriad of tasks. However, existing benchmarks fail to sufficiently challenge or showcase the full potential of LMMs as visual foundation agents in complex, real-world environments. To address this gap, we introduce VisualAgentBench (VAB), a comprehensive and unified benchmark specifically designed to train and evaluate LMMs as visual foundation agents across diverse scenarios in one standard setting, including Embodied, Graphical User Interface, and Visual Design, with tasks formulated to probe the depth of LMMs' understanding and interaction capabilities. Through rigorous testing across 9 proprietary LMM APIs and 9 open models (18 in total), we demonstrate the considerable yet still developing visual agent capabilities of these models. Additionally, VAB explores the synthesizing of visual agent trajectory data through hybrid methods including Program-based Solvers, LMM Agent Bootstrapping, and Human Demonstrations, offering insights into obstacles, solutions, and trade-offs one may meet in developing open LMM agents. Our work not only aims to benchmark existing models but also provides an instrumental playground for future development into visual foundation agents. Code, train, and test data are available at \url{https://github.com/THUDM/VisualAgentBench}.

Autonomous agents have become increasingly important for interacting with the real world. Android agents, in particular, have been 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 https://github.com/THUDM/Android-Lab.

Augmenting large language models (LLMs) with browsing tools substantially improves their potential as deep search agents to solve complex, real-world tasks. Yet, open LLMs still perform poorly in such settings due to limited long-horizon reasoning capacity with browsing tools and the lack of sufficiently difficult supervised data. To address these challenges, we present DeepDive to advance deep search agents. First, we propose a strategy to automatically synthesize complex, difficult, and hard-to-find questions from open knowledge graphs. Second, we apply end-to-end multi-turn reinforcement learning (RL) to enhance LLMs' long-horizon reasoning with deep search. Experiments show that DeepDive-32B achieves a new open-source competitive result on BrowseComp, outperforming WebSailor, DeepSeek-R1-Browse, and Search-o1. We demonstrate that multi-turn RL training improves deep search ability and significantly contributes to the performance improvements across multiple benchmarks. We observe that DeepDive enables test-time scaling of tool calls and parallel sampling. All datasets, models, and code are publicly available at https://github.com/THUDM/DeepDive.

We present GLM-4.1V-Thinking, GLM-4.5V, and GLM-4.6V, a family of vision-language models (VLMs) designed to advance general-purpose multimodal understanding and reasoning. In this report, we share our key findings in the development of the reasoning-centric training framework. We first develop a capable vision foundation model with significant potential through large-scale pre-training, which arguably sets the upper bound for the final performance. We then propose Reinforcement Learning with Curriculum Sampling (RLCS) to unlock the full potential of the model, leading to comprehensive capability enhancement across a diverse range of tasks, including STEM problem solving, video understanding, content recognition, coding, grounding, GUI-based agents, and long document interpretation. In a comprehensive evaluation across 42 public benchmarks, GLM-4.5V achieves state-of-the-art performance on nearly all tasks among open-source models of similar size, and demonstrates competitive or even superior results compared to closed-source models such as Gemini-2.5-Flash on challenging tasks including Coding and GUI Agents. Meanwhile, the smaller GLM-4.1V-9B-Thinking remains highly competitive-achieving superior results to the much larger Qwen2.5-VL-72B on 29 benchmarks. We open-source both GLM-4.1V-9B-Thinking and GLM-4.5V. We further introduce the GLM-4.6V series, open-source multimodal models with native tool use and a 128K context window. A brief overview is available at https://z.ai/blog/glm-4.6v. Code, models and more information are released at https://github.com/zai-org/GLM-V.

Instruction-following is a fundamental capability of language models, requiring the model to recognize even the most subtle requirements in the instructions and accurately reflect them in its output.Such an ability is well-suited for and often optimized by preference learning.However, existing methods often directly sample multiple independent responses from the model when creating preference pairs.Such practice can introduce content variations irrelevant to whether the instruction is precisely followed (e.g., different expressions about the same semantic), interfering with the goal of teaching models to recognize the key differences that lead to improved instruction following.In light of this, we introduce SPaR, a self-play framework integrating tree-search self-refinement to yield valid and comparable preference pairs free from distractions.By playing against itself, an LLM employs a tree-search strategy to refine its previous responses with respect to the instruction while minimizing unnecessary variations.Our experiments show that a LLaMA3-8B model, trained over three iterations guided by SPaR, surpasses GPT-4-Turbo on the IFEval benchmark without losing general capabilities. Furthermore, SPaR demonstrates promising scalability, greatly enhancing models like GLM-4-9B and LLaMA3-70B.We also identify how inference scaling in tree search would impact model performance.Our code and data are publicly available at https://github.com/thu-coai/SPaR.