Boyuan Zheng

Xiang Yue, Yuansheng Ni, Kai Zhang et al. · microsoft-research, princeton

We introduce MMMU: a new benchmark designed to evaluate multimodal models on massive multi-discipline tasks demanding college-level subject knowledge and deliberate reasoning. MMMU includes 11.5K meticulously collected multimodal questions from college exams, quizzes, and textbooks, covering six core disciplines: Art & Design, Business, Science, Health & Medicine, Humanities & Social Science, and Tech & Engineering. These questions span 30 subjects and 183 subfields, comprising 30 highly heterogeneous image types, such as charts, diagrams, maps, tables, music sheets, and chemical structures. Unlike existing benchmarks, MMMU focuses on advanced perception and reasoning with domain-specific knowledge, challenging models to perform tasks akin to those faced by experts. The evaluation of 14 open-source LMMs as well as the proprietary GPT-4V(ision) and Gemini highlights the substantial challenges posed by MMMU. Even the advanced GPT-4V and Gemini Ultra only achieve accuracies of 56% and 59% respectively, indicating significant room for improvement. We believe MMMU will stimulate the community to build next-generation multimodal foundation models towards expert artificial general intelligence.

Xiang Deng, Yu Gu, Boyuan Zheng et al. · microsoft-research

We introduce Mind2Web, the first dataset for developing and evaluating generalist agents for the web that can follow language instructions to complete complex tasks on any website. Existing datasets for web agents either use simulated websites or only cover a limited set of websites and tasks, thus not suitable for generalist web agents. With over 2,000 open-ended tasks collected from 137 websites spanning 31 domains and crowdsourced action sequences for the tasks, Mind2Web provides three necessary ingredients for building generalist web agents: 1) diverse domains, websites, and tasks, 2) use of real-world websites instead of simulated and simplified ones, and 3) a broad spectrum of user interaction patterns. Based on Mind2Web, we conduct an initial exploration of using large language models (LLMs) for building generalist web agents. While the raw HTML of real-world websites are often too large to be fed to LLMs, we show that first filtering it with a small LM significantly improves the effectiveness and efficiency of LLMs. Our solution demonstrates a decent level of performance, even on websites or entire domains the model has never seen before, but there is still a substantial room to improve towards truly generalizable agents. We open-source our dataset, model implementation, and trained models (https://osu-nlp-group.github.io/Mind2Web) to facilitate further research on building a generalist agent for the web.

Weiwei Gu, Boyuan Zheng, Yunmo Chen et al. · microsoft-research

We present an empirical study on methods for span finding, the selection of consecutive tokens in text for some downstream tasks. We focus on approaches that can be employed in training end-to-end information extraction systems, and find there is no definitive solution without considering task properties, and provide our observations to help with future design choices: 1) a tagging approach often yields higher precision while span enumeration and boundary prediction provide higher recall; 2) span type information can benefit a boundary prediction approach; 3) additional contextualization does not help span finding in most cases.

Yukun Feng, Feng Li, Ziang Song et al.

The Transformer architecture has led to significant gains in machine translation. However, most studies focus on only sentence-level translation without considering the context dependency within documents, leading to the inadequacy of document-level coherence. Some recent research tried to mitigate this issue by introducing an additional context encoder or translating with multiple sentences or even the entire document. Such methods may lose the information on the target side or have an increasing computational complexity as documents get longer. To address such problems, we introduce a recurrent memory unit to the vanilla Transformer, which supports the information exchange between the sentence and previous context. The memory unit is recurrently updated by acquiring information from sentences, and passing the aggregated knowledge back to subsequent sentence states. We follow a two-stage training strategy, in which the model is first trained at the sentence level and then finetuned for document-level translation. We conduct experiments on three popular datasets for document-level machine translation and our model has an average improvement of 0.91 s-BLEU over the sentence-level baseline. We also achieve state-of-the-art results on TED and News, outperforming the previous work by 0.36 s-BLEU and 1.49 d-BLEU on average.

Boyuan Zheng, Patrick Xia, Mahsa Yarmohammadi et al.

Existing multiparty dialogue datasets for entity coreference resolution are nascent, and many challenges are still unaddressed. We create a large-scale dataset, Multilingual Multiparty Coref (MMC), for this task based on TV transcripts. Due to the availability of gold-quality subtitles in multiple languages, we propose reusing the annotations to create silver coreference resolution data in other languages (Chinese and Farsi) via annotation projection. On the gold (English) data, off-the-shelf models perform relatively poorly on MMC, suggesting that MMC has broader coverage of multiparty coreference than prior datasets. On the silver data, we find success both using it for data augmentation and training from scratch, which effectively simulates the zero-shot cross-lingual setting.

Boyuan Zheng, Boyu Gou, Jihyung Kil et al. · microsoft-research

The recent development on large multimodal models (LMMs), especially GPT-4V(ision) and Gemini, has been quickly expanding the capability boundaries of multimodal models beyond traditional tasks like image captioning and visual question answering. In this work, we explore the potential of LMMs like GPT-4V as a generalist web agent that can follow natural language instructions to complete tasks on any given website. We propose SEEACT, a generalist web agent that harnesses the power of LMMs for integrated visual understanding and acting on the web. We evaluate on the recent MIND2WEB benchmark. In addition to standard offline evaluation on cached websites, we enable a new online evaluation setting by developing a tool that allows running web agents on live websites. We show that GPT-4V presents a great potential for web agents -- it can successfully complete 51.1 of the tasks on live websites if we manually ground its textual plans into actions on the websites. This substantially outperforms text-only LLMs like GPT-4 or smaller models (FLAN-T5 and BLIP-2) specifically fine-tuned for web agents. However, grounding still remains a major challenge. Existing LMM grounding strategies like set-of-mark prompting turns out to be not effective for web agents, and the best grounding strategy we develop in this paper leverages both the HTML structure and visuals. Yet, there is still a substantial gap with oracle grounding, leaving ample room for further improvement. All code, data, and evaluation tools are available at https://github.com/OSU-NLP-Group/SeeAct.

Yiqiao Li, Jianlong Zhou, Boyuan Zheng et al.

With the rapid deployment of graph neural networks (GNNs) based techniques into a wide range of applications such as link prediction, node classification, and graph classification the explainability of GNNs has become an indispensable component for predictive and trustworthy decision-making. Thus, it is critical to explain why graph neural network (GNN) makes particular predictions for them to be believed in many applications. Some GNNs explainers have been proposed recently. However, they lack to generate accurate and real explanations. To mitigate these limitations, we propose GANExplainer, based on Generative Adversarial Network (GAN) architecture. GANExplainer is composed of a generator to create explanations and a discriminator to assist with the Generator development. We investigate the explanation accuracy of our models by comparing the performance of GANExplainer with other state-of-the-art methods. Our empirical results on synthetic datasets indicate that GANExplainer improves explanation accuracy by up to 35\% compared to its alternatives.

CocoaBench Team, Shibo Hao, Zhining Zhang et al.

LLM agents now perform strongly in software engineering, deep research, GUI automation, and various other applications, while recent agent scaffolds and models are increasingly integrating these capabilities into unified systems. Yet, most evaluations still test these capabilities in isolation, which leaves a gap for more diverse use cases that require agents to combine different capabilities. We introduce CocoaBench, a benchmark for unified digital agents built from human-designed, long-horizon tasks that require flexible composition of vision, search, and coding. Tasks are specified only by an instruction and an automatic evaluation function over the final output, enabling reliable and scalable evaluation across diverse agent infrastructures. We also present CocoaAgent, a lightweight shared scaffold for controlled comparison across model backbones. Experiments show that current agents remain far from reliable on CocoaBench, with the best evaluated system achieving only 45.1% success rate. Our analysis further points to substantial room for improvement in reasoning and planning, tool use and execution, and visual grounding.

Boyuan Zheng, Jianlong Zhou, Fang Chen

Imitation learning demonstrates remarkable performance in various domains. However, imitation learning is also constrained by many prerequisites. The research community has done intensive research to alleviate these constraints, such as adding the stochastic policy to avoid unseen states, eliminating the need for action labels, and learning from the suboptimal demonstrations. Inspired by the natural reproduction process, we proposed a method called GenIL that integrates the Genetic Algorithm with imitation learning. The involvement of the Genetic Algorithm improves the data efficiency by reproducing trajectories with various returns and assists the model in estimating more accurate and compact reward function parameters. We tested GenIL in both Atari and Mujoco domains, and the result shows that it successfully outperforms the previous extrapolation methods over extrapolation accuracy, robustness, and overall policy performance when input data is limited.

Boyuan Zheng, Jianlong Zhou, Chunjie Liu et al.

As one of the prevalent methods to achieve automation systems, Imitation Learning (IL) presents a promising performance in a wide range of domains. However, despite the considerable improvement in policy performance, the corresponding research on the explainability of IL models is still limited. Inspired by the recent approaches in explainable artificial intelligence methods, we proposed a model-agnostic explaining framework for IL models called R2RISE. R2RISE aims to explain the overall policy performance with respect to the frames in demonstrations. It iteratively retrains the black-box IL model from the randomized masked demonstrations and uses the conventional evaluation outcome environment returns as the coefficient to build an importance map. We also conducted experiments to investigate three major questions concerning frames' importance equality, the effectiveness of the importance map, and connections between importance maps from different IL models. The result shows that R2RISE successfully distinguishes important frames from the demonstrations.

Xinyuan Wang, Bowen Wang, Dunjie Lu et al. · cmu

Vision-language models have demonstrated impressive capabilities as computer-use agents (CUAs) capable of automating diverse computer tasks. As their commercial potential grows, critical details of the most capable CUA systems remain closed. As these agents will increasingly mediate digital interactions and execute consequential decisions on our behalf, the research community needs access to open CUA frameworks to study their capabilities, limitations, and risks. To bridge this gap, we propose OpenCUA, a comprehensive open-source framework for scaling CUA data and foundation models. Our framework consists of: (1) an annotation infrastructure that seamlessly captures human computer-use demonstrations; (2) AgentNet, the first large-scale computer-use task dataset spanning 3 operating systems and 200+ applications and websites; (3) a scalable pipeline that transforms demonstrations into state-action pairs with reflective long Chain-of-Thought reasoning that sustain robust performance gains as data scales. Our end-to-end agent models demonstrate strong performance across CUA benchmarks. In particular, OpenCUA-72B achieves an average success rate of 45.0% on OSWorld-Verified, establishing a new state-of-the-art (SOTA) among open-source models. Further analysis confirms that our approach generalizes well across domains and benefits significantly from increased test-time computation. We release our annotation tool, datasets, code, and models to build open foundations for further CUA research.

Ruiqian Nai, Boyuan Zheng, Junming Zhao et al.

Current approaches for humanoid whole-body manipulation, primarily relying on teleoperation or visual sim-to-real reinforcement learning, are hindered by hardware logistics and complex reward engineering. Consequently, demonstrated autonomous skills remain limited and are typically restricted to controlled environments. In this paper, we present the Humanoid Manipulation Interface (HuMI), a portable and efficient framework for learning diverse whole-body manipulation tasks across various environments. HuMI enables robot-free data collection by capturing rich whole-body motion using portable hardware. This data drives a hierarchical learning pipeline that translates human motions into dexterous and feasible humanoid skills. Extensive experiments across five whole-body tasks--including kneeling, squatting, tossing, walking, and bimanual manipulation--demonstrate that HuMI achieves a 3x increase in data collection efficiency compared to teleoperation and attains a 70% success rate in unseen environments.

Ruiyang Zhu, Yuehan He, Boyuan Zheng et al.

End-to-end autonomous driving systems powered by Vision-Language-Action (VLA) models achieve strong performance on common driving scenarios, yet remain brittle in rare but safety-critical long-tail situations such as active construction zones and complex yielding geometries. In this paper, we present a method that addresses the long-tail challenging scenes beyond data scaling and model training. We introduce CLAP (Contrastive Latent-space Prompt optimization), a location-aware adaptation framework that augments a frozen VLA driving model with per-roadblock soft prompts, optimized from crowdsourced data and retrieved on demand via Vehicle-to-Everything (V2X) communication. Our approach rests on two observations from VLAs' latent space: (i) at the VLA's hidden-state layer, scenarios from the same roadblock cluster tightly and occupy compact regions of the latent space; and (ii) within a single roadblock, long-tail and normal frames are heavily intermixed in the latent representation, making it difficult to improve one without disturbing the other. CLAP addresses this via a two-stage pipeline: supervised contrastive learning to discover a roadblock-specific hard-scene direction, followed by directionally regularized prompt optimization that selectively improves challenging frames while preserving normal frame performance. On the NAVSIM benchmark with various state-of-the-art VLA backbones, CLAP reduces challenging scenario planning error by 24% with no regression on normal frames, significantly improving planning performance.

Hanwen Wang, Weizhi Zhao, Xiangyu Wang et al.

Achieving human-level manipulation requires dexterous robotic hands capable of complex object interactions. Advancing such capabilities further demands standardized benchmarks for systematic evaluation. However, existing dexterous benchmarks lack tasks that reflect the unique manipulation capabilities of dexterous hands over parallel grippers, as well as comprehensive evaluation pipelines. In this paper, we present DexJoCo, a benchmark and toolkit for task-oriented dexterous manipulation, comprising 11 functionally grounded tasks that evaluate tool-use, bimanual coordination, long-horizon execution, and reasoning. We develop a low-cost data collection system and collect 1.1K trajectories across these tasks, with support for domain randomization to assess robustness. We benchmark modern models under diverse settings, including visual and dynamics randomization, multi-task training, and action-head adaptation. Through extensive empirical analysis, we identify several important insights and common limitations of current policies in dexterous manipulation, highlighting key challenges for future research in dexterous hand robot learning. Project page available at: https://dexjoco.github.io

Xu Zheng, Zihao Dongfang, Lutao Jiang et al.

Humans possess spatial reasoning abilities that enable them to understand spaces through multimodal observations, such as vision and sound. Large multimodal reasoning models extend these abilities by learning to perceive and reason, showing promising performance across diverse spatial tasks. However, systematic reviews and publicly available benchmarks for these models remain limited. In this survey, we provide a comprehensive review of multimodal spatial reasoning tasks with large models, categorizing recent progress in multimodal large language models (MLLMs) and introducing open benchmarks for evaluation. We begin by outlining general spatial reasoning, focusing on post-training techniques, explainability, and architecture. Beyond classical 2D tasks, we examine spatial relationship reasoning, scene and layout understanding, as well as visual question answering and grounding in 3D space. We also review advances in embodied AI, including vision-language navigation and action models. Additionally, we consider emerging modalities such as audio and egocentric video, which contribute to novel spatial understanding through new sensors. We believe this survey establishes a solid foundation and offers insights into the growing field of multimodal spatial reasoning. Updated information about this survey, codes and implementation of the open benchmarks can be found at https://github.com/zhengxuJosh/Awesome-Spatial-Reasoning.

Fei Teng, Buyin Deng, Boyuan Zheng et al.

Field of Parallax (FoP)}, a spatial field that distills the common features from different LF representations to provide flexible and consistent support for multi-task learning. FoP is built upon three core features--projection difference, adjacency divergence, and contextual consistency--which are essential for cross-task adaptability. To implement FoP, we design a two-step angular adapter: the first step captures angular-specific differences, while the second step consolidates contextual consistency to ensure robust representation. Leveraging the FoP-based representation, we introduce the LFX framework, the first to handle arbitrary LF representations seamlessly, unifying LF multi-task vision. We evaluated LFX across three different tasks, achieving new state-of-the-art results, compared with previous task-specific architectures: 84.74% in mIoU for semantic segmentation on UrbanLF, 0.84% in AP for object detection on PKU, and 0.030 in MAE and 0.026 in MAE for salient object detection on Duftv2 and PKU, respectively. The source code will be made publicly available at https://github.com/warriordby/LFX.

Boyuan Zheng, Xiaoyu Yang, Yu-Ping Ruan et al.

This paper introduces the SemEval-2021 shared task 4: Reading Comprehension of Abstract Meaning (ReCAM). This shared task is designed to help evaluate the ability of machines in representing and understanding abstract concepts. Given a passage and the corresponding question, a participating system is expected to choose the correct answer from five candidates of abstract concepts in a cloze-style machine reading comprehension setup. Based on two typical definitions of abstractness, i.e., the imperceptibility and nonspecificity, our task provides three subtasks to evaluate the participating models. Specifically, Subtask 1 aims to evaluate how well a system can model concepts that cannot be directly perceived in the physical world. Subtask 2 focuses on models' ability in comprehending nonspecific concepts located high in a hypernym hierarchy given the context of a passage. Subtask 3 aims to provide some insights into models' generalizability over the two types of abstractness. During the SemEval-2021 official evaluation period, we received 23 submissions to Subtask 1 and 28 to Subtask 2. The participating teams additionally made 29 submissions to Subtask 3. The leaderboard and competition website can be found at https://competitions.codalab.org/competitions/26153. The data and baseline code are available at https://github.com/boyuanzheng010/SemEval2021-Reading-Comprehension-of-Abstract-Meaning.

Yu Gu, Kai Zhang, Yuting Ning et al. · microsoft-research

Language agents based on large language models (LLMs) have demonstrated great promise in automating web-based tasks. Recent work has shown that incorporating advanced planning algorithms, e.g., tree search, is advantageous over reactive planning for web agents. However, unlike simulated sandbox environments, real-world environments such as the web are rife with irreversible actions. This undermines the feasibility of backtracking, a cornerstone of (tree) search. Overly relying on test-time search also hurts efficiency. We advocate model-based planning for web agents that employs a world model to simulate and deliberate over the outcome of each candidate action before committing to one. We systematically explore this paradigm by (1) Proposing a model-based planning framework, WebDreamer, which employs LLMs to serve as both world models and value functions; (2) Training specialized LLMs as world models with a scalable data synthesis pipeline. Empirical results demonstrate that WebDreamer achieves substantial performance improvements over reactive baselines. It is competitive, while being 4-5 times more efficient, with tree search in sandbox environments (VisualWebArena) and also works effectively on real-world websites (Online-Mind2Web and Mind2Web-Live). Furthermore, our trained world model, Dreamer-7B, performs comparable to GPT-4o, highlighting the potential of specialized world models for efficient and effective planning in complex web environments.

Lingfeng Shen, Weiting Tan, Sihao Chen et al.

As the influence of large language models (LLMs) spans across global communities, their safety challenges in multilingual settings become paramount for alignment research. This paper examines the variations in safety challenges faced by LLMs across different languages and discusses approaches to alleviating such concerns. By comparing how state-of-the-art LLMs respond to the same set of malicious prompts written in higher- vs. lower-resource languages, we observe that (1) LLMs tend to generate unsafe responses much more often when a malicious prompt is written in a lower-resource language, and (2) LLMs tend to generate more irrelevant responses to malicious prompts in lower-resource languages. To understand where the discrepancy can be attributed, we study the effect of instruction tuning with reinforcement learning from human feedback (RLHF) or supervised finetuning (SFT) on the HH-RLHF dataset. Surprisingly, while training with high-resource languages improves model alignment, training in lower-resource languages yields minimal improvement. This suggests that the bottleneck of cross-lingual alignment is rooted in the pretraining stage. Our findings highlight the challenges in cross-lingual LLM safety, and we hope they inform future research in this direction.

Boyuan Zheng, Michael Y. Fatemi, Xiaolong Jin et al. · microsoft-research

To survive and thrive in complex environments, humans have evolved sophisticated self-improvement mechanisms through environment exploration, hierarchical abstraction of experiences into reuseable skills, and collaborative construction of an ever-growing skill repertoire. Despite recent advancements, autonomous web agents still lack crucial self-improvement capabilities, struggling with procedural knowledge abstraction, refining skills, and skill composition. In this work, we introduce SkillWeaver, a skill-centric framework enabling agents to self-improve by autonomously synthesizing reusable skills as APIs. Given a new website, the agent autonomously discovers skills, executes them for practice, and distills practice experiences into robust APIs. Iterative exploration continually expands a library of lightweight, plug-and-play APIs, significantly enhancing the agent's capabilities. Experiments on WebArena and real-world websites demonstrate the efficacy of SkillWeaver, achieving relative success rate improvements of 31.8% and 39.8%, respectively. Additionally, APIs synthesized by strong agents substantially enhance weaker agents through transferable skills, yielding improvements of up to 54.3% on WebArena. These results demonstrate the effectiveness of honing diverse website interactions into APIs, which can be seamlessly shared among various web agents.

Lingbo Mo, Zeyi Liao, Boyuan Zheng et al. · microsoft-research

Language agents powered by large language models (LLMs) have seen exploding development. Their capability of using language as a vehicle for thought and communication lends an incredible level of flexibility and versatility. People have quickly capitalized on this capability to connect LLMs to a wide range of external components and environments: databases, tools, the Internet, robotic embodiment, etc. Many believe an unprecedentedly powerful automation technology is emerging. However, new automation technologies come with new safety risks, especially for intricate systems like language agents. There is a surprisingly large gap between the speed and scale of their development and deployment and our understanding of their safety risks. Are we building a house of cards? In this position paper, we present the first systematic effort in mapping adversarial attacks against language agents. We first present a unified conceptual framework for agents with three major components: Perception, Brain, and Action. Under this framework, we present a comprehensive discussion and propose 12 potential attack scenarios against different components of an agent, covering different attack strategies (e.g., input manipulation, adversarial demonstrations, jailbreaking, backdoors). We also draw connections to successful attack strategies previously applied to LLMs. We emphasize the urgency to gain a thorough understanding of language agent risks before their widespread deployment.

Jihyung Kil, Chan Hee Song, Boyuan Zheng et al. · microsoft-research

Automatic web navigation aims to build a web agent that can follow language instructions to execute complex and diverse tasks on real-world websites. Existing work primarily takes HTML documents as input, which define the contents and action spaces (i.e., actionable elements and operations) of webpages. Nevertheless, HTML documents may not provide a clear task-related context for each element, making it hard to select the right (sequence of) actions. In this paper, we propose to contextualize HTML elements through their "dual views" in webpage screenshots: each HTML element has its corresponding bounding box and visual content in the screenshot. We build upon the insight -- web developers tend to arrange task-related elements nearby on webpages to enhance user experiences -- and propose to contextualize each element with its neighbor elements, using both textual and visual features. The resulting representations of HTML elements are more informative for the agent to take action. We validate our method on the recently released Mind2Web dataset, which features diverse navigation domains and tasks on real-world websites. Our method consistently outperforms the baseline in all the scenarios, including cross-task, cross-website, and cross-domain ones.

Tanmay Gupta, Piper Wolters, Zixian Ma et al.

Web agents--autonomous systems that navigate and execute tasks on the web on behalf of users--have the potential to transform how people interact with the digital world. However, the most capable web agents today rely on proprietary models with undisclosed training data and recipes, limiting scientific understanding, reproducibility, and community-driven progress. We believe agents for the open web should be built in the open. To this end, we introduce (1) MolmoWebMix, a large and diverse mixture of browser task demonstrations and web-GUI perception data and (2) MolmoWeb, a family of fully open multimodal web agents. Specifically, MolmoWebMix combines over 100K synthetic task trajectories from multiple complementary generation pipelines with 30K+ human demonstrations, atomic web-skill trajectories, and GUI perception data, including referring expression grounding and screenshot question answering. MolmoWeb agents operate as instruction-conditioned visual-language action policies: given a task instruction and a webpage screenshot, they predict the next browser action, requiring no access to HTML, accessibility trees, or specialized APIs. Available in 4B and 8B size, on browser-use benchmarks like WebVoyager, Online-Mind2Web, and DeepShop, MolmoWeb agents achieve state-of-the-art results outperforming similar scale open-weight-only models such as Fara-7B, UI-Tars-1.5-7B, and Holo1-7B. MolmoWeb-8B also surpasses set-of-marks (SoM) agents built on much larger closed frontier models like GPT-4o. We further demonstrate consistent gains through test-time scaling via parallel rollouts with best-of-N selection, achieving 94.7% and 60.5% pass@4 (compared to 78.2% and 35.3% pass@1) on WebVoyager and Online-Mind2Web respectively. We will release model checkpoints, training data, code, and a unified evaluation harness to enable reproducibility and accelerate open research on web agents.

Tong Zhang, Boyuan Zheng, Ruiqian Nai et al.

The human body demonstrates exceptional motor capabilities-such as standing steadily on one foot or performing a high kick with the leg raised over 1.5 meters-both requiring precise balance control. While recent research on humanoid control has leveraged reinforcement learning to track human motions for skill acquisition, applying this paradigm to balance-intensive tasks remains challenging. In this work, we identify three key obstacles: instability from reference motion errors, learning difficulties due to morphological mismatch, and the sim-to-real gap caused by sensor noise and unmodeled dynamics. To address these challenges, we propose HuB (Humanoid Balance), a unified framework that integrates reference motion refinement, balance-aware policy learning, and sim-to-real robustness training, with each component targeting a specific challenge. We validate our approach on the Unitree G1 humanoid robot across challenging quasi-static balance tasks, including extreme single-legged poses such as Swallow Balance and Bruce Lee's Kick. Our policy remains stable even under strong physical disturbances-such as a forceful soccer strike-while baseline methods consistently fail to complete these tasks. Project website: https://hub-robot.github.io

Boyu Gou, Zanming Huang, Yuting Ning et al. · microsoft-research

Agentic search such as Deep Research systems-where agents autonomously browse the web, synthesize information, and return comprehensive citation-backed answers-represents a major shift in how users interact with web-scale information. While promising greater efficiency and cognitive offloading, the growing complexity and open-endedness of agentic search have outpaced existing evaluation benchmarks and methodologies, which largely assume short search horizons and static answers. In this paper, we introduce Mind2Web 2, a benchmark of 130 realistic, high-quality, and long-horizon tasks that require real-time web browsing and extensive information synthesis, constructed with over 1000 hours of human labor. To address the challenge of evaluating time-varying and complex answers, we propose a novel Agent-as-a-Judge framework. Our method constructs task-specific judge agents based on a tree-structured rubric design to automatically assess both answer correctness and source attribution. We conduct a comprehensive evaluation of ten frontier agentic search systems and human performance, along with a detailed error analysis to draw insights for future development. The best-performing system, OpenAI Deep Research, can already achieve 50-70% of human performance while spending half the time, highlighting its great potential. Altogether, Mind2Web 2 provides a rigorous foundation for developing and benchmarking the next generation of agentic search systems.

Boyuan Zheng, Shouyi Lu, Renbo Huang et al.

We introduce R2LDM, an innovative approach for generating dense and accurate 4D radar point clouds, guided by corresponding LiDAR point clouds. Instead of utilizing range images or bird's eye view (BEV) images, we represent both LiDAR and 4D radar point clouds using voxel features, which more effectively capture 3D shape information. Subsequently, we propose the Latent Voxel Diffusion Model (LVDM), which performs the diffusion process in the latent space. Additionally, a novel Latent Point Cloud Reconstruction (LPCR) module is utilized to reconstruct point clouds from high-dimensional latent voxel features. As a result, R2LDM effectively generates LiDAR-like point clouds from paired raw radar data. We evaluate our approach on two different datasets, and the experimental results demonstrate that our model achieves 6- to 10-fold densification of radar point clouds, outperforming state-of-the-art baselines in 4D radar point cloud super-resolution. Furthermore, the enhanced radar point clouds generated by our method significantly improve downstream tasks, achieving up to 31.7% improvement in point cloud registration recall rate and 24.9% improvement in object detection accuracy.

Boyuan Zheng, Zeyi Liao, Scott Salisbury et al. · microsoft-research

The rapid development of autonomous web agents powered by Large Language Models (LLMs), while greatly elevating efficiency, exposes the frontier risk of taking unintended or harmful actions. This situation underscores an urgent need for effective safety measures, akin to access controls for human users. To address this critical challenge, we introduce WebGuard, the first comprehensive dataset designed to support the assessment of web agent action risks and facilitate the development of guardrails for real-world online environments. In doing so, WebGuard specifically focuses on predicting the outcome of state-changing actions and contains 4,939 human-annotated actions from 193 websites across 22 diverse domains, including often-overlooked long-tail websites. These actions are categorized using a novel three-tier risk schema: SAFE, LOW, and HIGH. The dataset includes designated training and test splits to support evaluation under diverse generalization settings. Our initial evaluations reveal a concerning deficiency: even frontier LLMs achieve less than 60% accuracy in predicting action outcomes and less than 60% recall in lagging HIGH-risk actions, highlighting the risks of deploying current-generation agents without dedicated safeguards. We therefore investigate fine-tuning specialized guardrail models using WebGuard. We conduct comprehensive evaluations across multiple generalization settings and find that a fine-tuned Qwen2.5VL-7B model yields a substantial improvement in performance, boosting accuracy from 37% to 80% and HIGH-risk action recall from 20% to 76%. Despite these improvements, the performance still falls short of the reliability required for high-stakes deployment, where guardrails must approach near-perfect accuracy and recall.

Minqing Huang, Shouyi Lu, Boyuan Zheng et al.

4D radar super-resolution, which aims to reconstruct sparse and noisy point clouds into dense and geometrically consistent representations, is a foundational problem in autonomous perception. However, existing methods often suffer from high training cost or rely on complex diffusion-based sampling, resulting in high inference latency and poor generalization, making it difficult to balance accuracy and efficiency. To address these limitations, we propose MSDNet, a multi-stage distillation framework that efficiently transfers dense LiDAR priors to 4D radar features to achieve both high reconstruction quality and computational efficiency. The first stage performs reconstruction-guided feature distillation, aligning and densifying the student's features through feature reconstruction. In the second stage, we propose diffusion-guided feature distillation, which treats the stage-one distilled features as a noisy version of the teacher's representations and refines them via a lightweight diffusion network. Furthermore, we introduce a noise adapter that adaptively aligns the noise level of the feature with a predefined diffusion timestep, enabling a more precise denoising. Extensive experiments on the VoD and in-house datasets demonstrate that MSDNet achieves both high-fidelity reconstruction and low-latency inference in the task of 4D radar point cloud super-resolution, and consistently improves performance on downstream tasks. The code will be publicly available upon publication.

Xiao Tang, Guirong Zhuo, Cong Wang et al.

3D reconstruction and novel view synthesis are critical for validating autonomous driving systems and training advanced perception models. Recent self-supervised methods have gained significant attention due to their cost-effectiveness and enhanced generalization in scenarios where annotated bounding boxes are unavailable. However, existing approaches, which often rely on frequency-domain decoupling or optical flow, struggle to accurately reconstruct dynamic objects due to imprecise motion estimation and weak temporal consistency, resulting in incomplete or distorted representations of dynamic scene elements. To address these challenges, we propose 4DRadar-GS, a 4D Radar-augmented self-supervised 3D reconstruction framework tailored for dynamic driving scenes. Specifically, we first present a 4D Radar-assisted Gaussian initialization scheme that leverages 4D Radar's velocity and spatial information to segment dynamic objects and recover monocular depth scale, generating accurate Gaussian point representations. In addition, we propose a Velocity-guided PointTrack (VGPT) model, which is jointly trained with the reconstruction pipeline under scene flow supervision, to track fine-grained dynamic trajectories and construct temporally consistent representations. Evaluated on the OmniHD-Scenes dataset, 4DRadar-GS achieves state-of-the-art performance in dynamic driving scene 3D reconstruction.

Xinran Zhao, Boyuan Zheng, Chenglei Si et al.

This paper revisits Ramon Llull's Ars combinatoria - a medieval framework for generating knowledge through symbolic recombination - as a conceptual foundation for building a modern Llull's thinking machine for research ideation. Our approach defines three compositional axes: Theme (e.g., efficiency, adaptivity), Domain (e.g., question answering, machine translation), and Method (e.g., adversarial training, linear attention). These elements represent high-level abstractions common in scientific work - motivations, problem settings, and technical approaches - and serve as building blocks for LLM-driven exploration. We mine elements from human experts or conference papers and show that prompting LLMs with curated combinations produces research ideas that are diverse, relevant, and grounded in current literature. This modern thinking machine offers a lightweight, interpretable tool for augmenting scientific creativity and suggests a path toward collaborative ideation between humans and AI.

Boyuan Zheng, Victor W. Chu, Zhidong Li et al.

Climate change has intensified the frequency and severity of extreme weather events, presenting unprecedented challenges to the agricultural industry worldwide. In this investigation, we focus on kiwifruit farming in New Zealand. We propose to examine the impacts of climate-induced extreme events, specifically frost, drought, extreme rainfall, and heatwave, on kiwifruit harvest yields. These four events were selected due to their significant impacts on crop productivity and their prevalence as recorded by climate monitoring institutions in the country. We employed Isolation Forest, an unsupervised anomaly detection method, to analyse climate history and recorded extreme events, alongside with kiwifruit yields. Our analysis reveals considerable variability in how different types of extreme event affect kiwifruit yields underscoring notable discrepancies between climatic extremes and individual farm's yield outcomes. Additionally, our study highlights critical limitations of current anomaly detection approaches, particularly in accurately identifying events such as frost. These findings emphasise the need for integrating supplementary features like farm management strategies with climate adaptation practices. Our further investigation will employ ensemble methods that consolidate nearby farms' yield data and regional climate station features to reduce variance, thereby enhancing the accuracy and reliability of extreme event detection and the formulation of response strategies.

Boyuan Zheng, Victor W. Chu

Climate extremes present escalating risks to agriculture intensifying the need for reliable multi-hazard early warning systems (EWS). The situation is evolving due to climate change and hence such systems should have the intelligent to continue to learn from recent climate behaviours. However, traditional single-hazard forecasting methods fall short in capturing complex interactions among concurrent climatic events. To address this deficiency, in this paper, we combine sequential deep learning models and advanced Explainable Artificial Intelligence (XAI) techniques to introduce a multi-hazard forecasting framework for agriculture. In our experiments, we utilize meteorological data from four prominent agricultural regions in the United States (between 2010 and 2023) to validate the predictive accuracy of our framework on multiple severe event types, which are extreme cold, floods, frost, hail, heatwaves, and heavy rainfall, with tailored models for each area. The framework uniquely integrates attention mechanisms with TimeSHAP (a recurrent XAI explainer for time series) to provide comprehensive temporal explanations revealing not only which climatic features are influential but precisely when their impacts occur. Our results demonstrate strong predictive accuracy, particularly with the BiLSTM architecture, and highlight the system's capacity to inform nuanced, proactive risk management strategies. This research significantly advances the explainability and applicability of multi-hazard EWS, fostering interdisciplinary trust and effective decision-making process for climate risk management in the agricultural industry.

Boyuan Zheng, Zerui Fang, Zhe Xu et al.

While AI for Science (AI4S) serves as an analytical tool in the current research paradigm, it doesn't solve its core inefficiency. We propose "Agent for Science" (Agent4S)-the use of LLM-driven agents to automate the entire research workflow-as the true Fifth Scientific Paradigm. This paper introduces a five-level classification for Agent4S, outlining a clear roadmap from simple task automation to fully autonomous, collaborative "AI Scientists." This framework defines the next revolutionary step in scientific discovery.

Hai Zhang, Boyuan Zheng, Tianying Ji et al.

Offline meta reinforcement learning (OMRL) has emerged as a promising approach for interaction avoidance and strong generalization performance by leveraging pre-collected data and meta-learning techniques. Previous context-based approaches predominantly rely on the intuition that alternating optimization between the context encoder and the policy can lead to performance improvements, as long as the context encoder follows the principle of maximizing the mutual information between the task variable $M$ and its latent representation $Z$ ($I(Z;M)$) while the policy adopts the standard offline reinforcement learning (RL) algorithms conditioning on the learned task representation.Despite promising results, the theoretical justification of performance improvements for such intuition remains underexplored.Inspired by the return discrepancy scheme in the model-based RL field, we find that the previous optimization framework can be linked with the general RL objective of maximizing the expected return, thereby explaining performance improvements. Furthermore, after scrutinizing this optimization framework, we observe that the condition for monotonic performance improvements does not consider the variation of the task representation. When these variations are considered, the previously established condition may no longer be sufficient to ensure monotonicity, thereby impairing the optimization process.We name this issue task representation shift and theoretically prove that the monotonic performance improvements can be guaranteed with appropriate context encoder updates.Our work opens up a new avenue for OMRL, leading to a better understanding between the task representation and performance improvements.

Lingfeng Shen, Weiting Tan, Boyuan Zheng et al.

With growing capabilities of large language models, prompting them has become the dominant way to access them. This has motivated the development of strategies for automatically selecting effective language prompts. In this paper, we introduce prompt flatness, a new metric to quantify the expected utility of a language prompt. This metric is inspired by flatness regularization in statistical learning that quantifies the robustness of the model towards its parameter perturbations. We provide theoretical foundations for this metric and its relationship with other prompt selection metrics, providing a comprehensive understanding of existing methods. Empirically, we show that combining prompt flatness with existing metrics improves both performance and sample efficiency. Our metric outperforms the previous prompt selection metrics with an average increase of 5% in accuracy and 10% in Pearson correlation across 6 classification benchmarks.

Cunxiang Wang, Boyuan Zheng, Yuchen Niu et al.

To quantitatively and intuitively explore the generalization ability of pre-trained language models (PLMs), we have designed several tasks of arithmetic and logical reasoning. We both analyse how well PLMs generalize when the test data is in the same distribution as the train data and when it is different, for the latter analysis, we have also designed a cross-distribution test set other than the in-distribution test set. We conduct experiments on one of the most advanced and publicly released generative PLM - BART. Our research finds that the PLMs can easily generalize when the distribution is the same, however, it is still difficult for them to generalize out of the distribution.

Boyuan Zheng, Sunny Verma, Jianlong Zhou et al.

Imitation learning aims to extract knowledge from human experts' demonstrations or artificially created agents in order to replicate their behaviors. Its success has been demonstrated in areas such as video games, autonomous driving, robotic simulations and object manipulation. However, this replicating process could be problematic, such as the performance is highly dependent on the demonstration quality, and most trained agents are limited to perform well in task-specific environments. In this survey, we provide a systematic review on imitation learning. We first introduce the background knowledge from development history and preliminaries, followed by presenting different taxonomies within Imitation Learning and key milestones of the field. We then detail challenges in learning strategies and present research opportunities with learning policy from suboptimal demonstration, voice instructions and other associated optimization schemes.