Xiaohan Zhang

Zhuoyi Yang, Jiayan Teng, Wendi Zheng et al. · tsinghua

We present CogVideoX, a large-scale text-to-video generation model based on diffusion transformer, which can generate 10-second continuous videos aligned with text prompt, with a frame rate of 16 fps and resolution of 768 * 1360 pixels. Previous video generation models often had limited movement and short durations, and is difficult to generate videos with coherent narratives based on text. We propose several designs to address these issues. First, we propose a 3D Variational Autoencoder (VAE) to compress videos along both spatial and temporal dimensions, to improve both compression rate and video fidelity. Second, to improve the text-video alignment, we propose an expert transformer with the expert adaptive LayerNorm to facilitate the deep fusion between the two modalities. Third, by employing a progressive training and multi-resolution frame pack technique, CogVideoX is adept at producing coherent, long-duration, different shape videos characterized by significant motions. In addition, we develop an effective text-video data processing pipeline that includes various data preprocessing strategies and a video captioning method, greatly contributing to the generation quality and semantic alignment. Results show that CogVideoX demonstrates state-of-the-art performance across both multiple machine metrics and human evaluations. The model weight of both 3D Causal VAE, Video caption model and CogVideoX are publicly available at https://github.com/THUDM/CogVideo.

Jifan Yu, Xiaozhi Wang, Shangqing Tu et al. · tsinghua

The unprecedented performance of large language models (LLMs) necessitates improvements in evaluations. Rather than merely exploring the breadth of LLM abilities, we believe meticulous and thoughtful designs are essential to thorough, unbiased, and applicable evaluations. Given the importance of world knowledge to LLMs, we construct a Knowledge-oriented LLM Assessment benchmark (KoLA), in which we carefully design three crucial factors: (1) For \textbf{ability modeling}, we mimic human cognition to form a four-level taxonomy of knowledge-related abilities, covering $19$ tasks. (2) For \textbf{data}, to ensure fair comparisons, we use both Wikipedia, a corpus prevalently pre-trained by LLMs, along with continuously collected emerging corpora, aiming to evaluate the capacity to handle unseen data and evolving knowledge. (3) For \textbf{evaluation criteria}, we adopt a contrastive system, including overall standard scores for better numerical comparability across tasks and models and a unique self-contrast metric for automatically evaluating knowledge-creating ability. We evaluate $28$ open-source and commercial LLMs and obtain some intriguing findings. The KoLA dataset and open-participation leaderboard are publicly released at https://kola.xlore.cn and will be continuously updated to provide references for developing LLMs and knowledge-related systems.

Xiao Liu, Xuanyu Lei, Shengyuan Wang et al. · tsinghua

Alignment has become a critical step for instruction-tuned Large Language Models (LLMs) to become helpful assistants. However, the effective evaluation of alignment for emerging Chinese LLMs is still largely unexplored. To fill in this gap, we introduce AlignBench, a comprehensive multi-dimensional benchmark for evaluating LLMs' alignment in Chinese. We design a human-in-the-loop data curation pipeline, containing eight main categories, 683 real-scenario rooted queries and corresponding human verified references. To ensure the correctness of references, each knowledge-intensive query is accompanied with evidences collected from reliable web sources (including URLs and quotations) by our annotators. For automatic evaluation, our benchmark employs a rule-calibrated multi-dimensional LLM-as-Judge~\cite{zheng2023judging} approach with Chain-of-Thought to generate explanations and final ratings, ensuring high reliability and interpretability. All evaluation code, data, and LLM generations are available at \url{https://github.com/THUDM/AlignBench}. Since its release, AlignBench has been adopted by top (Chinese) LLMs for evaluating their alignment capabilities in Chinese, including ChatGLM, Qwen, DeepSeek, Yi, Baichuan, and Abab.

Jing Zhang, Xiaokang Zhang, Daniel Zhang-Li et al. · tsinghua

We present GLM-Dialog, a large-scale language model (LLM) with 10B parameters capable of knowledge-grounded conversation in Chinese using a search engine to access the Internet knowledge. GLM-Dialog offers a series of applicable techniques for exploiting various external knowledge including both helpful and noisy knowledge, enabling the creation of robust knowledge-grounded dialogue LLMs with limited proper datasets. To evaluate the GLM-Dialog more fairly, we also propose a novel evaluation method to allow humans to converse with multiple deployed bots simultaneously and compare their performance implicitly instead of explicitly rating using multidimensional metrics.Comprehensive evaluations from automatic to human perspective demonstrate the advantages of GLM-Dialog comparing with existing open source Chinese dialogue models. We release both the model checkpoint and source code, and also deploy it as a WeChat application to interact with users. We offer our evaluation platform online in an effort to prompt the development of open source models and reliable dialogue evaluation systems. The additional easy-to-use toolkit that consists of short text entity linking, query generation, and helpful knowledge classification is also released to enable diverse applications. All the source code is available on Github.

Bo Liu, Yuqian Jiang, Xiaohan Zhang et al.

Large language models (LLMs) have demonstrated remarkable zero-shot generalization abilities: state-of-the-art chatbots can provide plausible answers to many common questions that arise in daily life. However, so far, LLMs cannot reliably solve long-horizon planning problems. By contrast, classical planners, once a problem is given in a formatted way, can use efficient search algorithms to quickly identify correct, or even optimal, plans. In an effort to get the best of both worlds, this paper introduces LLM+P, the first framework that incorporates the strengths of classical planners into LLMs. LLM+P takes in a natural language description of a planning problem, then returns a correct (or optimal) plan for solving that problem in natural language. LLM+P does so by first converting the language description into a file written in the planning domain definition language (PDDL), then leveraging classical planners to quickly find a solution, and then translating the found solution back into natural language. Along with LLM+P, we define a diverse set of different benchmark problems taken from common planning scenarios. Via a comprehensive set of experiments on these benchmark problems, we find that LLM+P is able to provide optimal solutions for most problems, while LLMs fail to provide even feasible plans for most problems.\footnote{The code and results are publicly available at https://github.com/Cranial-XIX/llm-pddl.git.

GLM-5 Team, Aohan Zeng, Xin Lv et al. · tsinghua

We present GLM-5, a next-generation foundation model designed to transition the paradigm of vibe coding to agentic engineering. Building upon the agentic, reasoning, and coding (ARC) capabilities of its predecessor, GLM-5 adopts DSA to significantly reduce training and inference costs while maintaining long-context fidelity. To advance model alignment and autonomy, we implement a new asynchronous reinforcement learning infrastructure that drastically improves post-training efficiency by decoupling generation from training. Furthermore, we propose novel asynchronous agent RL algorithms that further improve RL quality, enabling the model to learn from complex, long-horizon interactions more effectively. Through these innovations, GLM-5 achieves state-of-the-art performance on major open benchmarks. Most critically, GLM-5 demonstrates unprecedented capability in real-world coding tasks, surpassing previous baselines in handling end-to-end software engineering challenges. Code, models, and more information are available at https://github.com/zai-org/GLM-5.

Wenyi Hong, Weihan Wang, Ming Ding et al.

Beginning with VisualGLM and CogVLM, we are continuously exploring VLMs in pursuit of enhanced vision-language fusion, efficient higher-resolution architecture, and broader modalities and applications. Here we propose the CogVLM2 family, a new generation of visual language models for image and video understanding including CogVLM2, CogVLM2-Video and GLM-4V. As an image understanding model, CogVLM2 inherits the visual expert architecture with improved training recipes in both pre-training and post-training stages, supporting input resolution up to $1344 \times 1344$ pixels. As a video understanding model, CogVLM2-Video integrates multi-frame input with timestamps and proposes automated temporal grounding data construction. Notably, CogVLM2 family has achieved state-of-the-art results on benchmarks like MMBench, MM-Vet, TextVQA, MVBench and VCGBench. All models are open-sourced in https://github.com/THUDM/CogVLM2 and https://github.com/THUDM/GLM-4, contributing to the advancement of the field.

Xiaohan Zhang, Songlin Dong, Jinjie Chen et al.

In this paper, we focus on a new and challenging decentralized machine learning paradigm in which there are continuous inflows of data to be addressed and the data are stored in multiple repositories. We initiate the study of data decentralized class-incremental learning (DCIL) by making the following contributions. Firstly, we formulate the DCIL problem and develop the experimental protocol. Secondly, we introduce a paradigm to create a basic decentralized counterpart of typical (centralized) class-incremental learning approaches, and as a result, establish a benchmark for the DCIL study. Thirdly, we further propose a Decentralized Composite knowledge Incremental Distillation framework (DCID) to transfer knowledge from historical models and multiple local sites to the general model continually. DCID consists of three main components namely local class-incremental learning, collaborated knowledge distillation among local models, and aggregated knowledge distillation from local models to the general one. We comprehensively investigate our DCID framework by using different implementations of the three components. Extensive experimental results demonstrate the effectiveness of our DCID framework. The codes of the baseline methods and the proposed DCIL will be released at https://github.com/zxxxxh/DCIL.

Priyam Parashar, Vidhi Jain, Xiaohan Zhang et al. · cmu

Despite great strides in language-guided manipulation, existing work has been constrained to table-top settings. Table-tops allow for perfect and consistent camera angles, properties are that do not hold in mobile manipulation. Task plans that involve moving around the environment must be robust to egocentric views and changes in the plane and angle of grasp. A further challenge is ensuring this is all true while still being able to learn skills efficiently from limited data. We propose Spatial-Language Attention Policies (SLAP) as a solution. SLAP uses three-dimensional tokens as the input representation to train a single multi-task, language-conditioned action prediction policy. Our method shows an 80% success rate in the real world across eight tasks with a single model, and a 47.5% success rate when unseen clutter and unseen object configurations are introduced, even with only a handful of examples per task. This represents an improvement of 30% over prior work (20% given unseen distractors and configurations). We see a 4x improvement over baseline in mobile manipulation setting. In addition, we show how SLAPs robustness allows us to execute Task Plans from open-vocabulary instructions using a large language model for multi-step mobile manipulation. For videos, see the website: https://robotslap.github.io

Yutao Shi, Xiaohan Zhang, Xiangjing Zhang et al.

The Model Context Protocol (MCP) has emerged as a critical standard empowering Large Language Models (LLMs) to utilize external tools. In this ecosystem, LLMs rely on natural language descriptions provided by MCP servers to select and execute functions. This interaction implicitly assumes that tool descriptions faithfully reflect their underlying implementations, while this assumption is not mandatorily verified in practice. As a result, MCP deployments may suffer from a problem named Description-Code Inconsistency (DCI), where a tool's description of its capabilities and security boundaries is not consistent with what the code actually does. In this paper, we present a comprehensive study of DCI in real-world MCP servers. We formally define the problem and propose a comprehensive taxonomy spanning functionality inconsistencies and undeclared side effects. Guided by this taxonomy, we develop DCIChecker, an automated framework that combines structure-aware static analysis with the Direct-Reverse-Arbitration prompting method to cross-validate tool descriptions against actual code implementations. We apply this framework to a large-scale dataset comprising 19,200 description-code pairs extracted from 2,214 real-world MCP servers. Our measurement reveals that DCI is widespread, with 9.93% of these pairs exhibiting inconsistencies. We further demonstrate that DCI creates a critical defense blind spot, facilitating varied risks from operational failures to stealthy malicious behaviors. Finally, we propose mitigation strategies to enforce semantic consistency and enhance the reliability of the emerging agentic ecosystem.

Xiaohan Zhang, Bharat Lal Bhatnagar, Vladimir Guzov et al.

Humans interact with an object in many different ways by making contact at different locations, creating a highly complex motion space that can be difficult to learn, particularly when synthesizing such human interactions in a controllable manner. Existing works on synthesizing human scene interaction focus on the high-level control of action but do not consider the fine-grained control of motion. In this work, we study the problem of synthesizing scene interactions conditioned on different contact positions on the object. As a testbed to investigate this new problem, we focus on human-chair interaction as one of the most common actions which exhibit large variability in terms of contacts. We propose a novel synthesis framework COUCH that plans ahead the motion by predicting contact-aware control signals of the hands, which are then used to synthesize contact-conditioned interactions. Furthermore, we contribute a large human-chair interaction dataset with clean annotations, the COUCH Dataset. Our method shows significant quantitative and qualitative improvements over existing methods for human-object interactions. More importantly, our method enables control of the motion through user-specified or automatically predicted contacts.

Yong Xiao, Xiaohan Zhang, Guangming Shi et al.

Real-time machine learning has recently attracted significant interest due to its potential to support instantaneous learning, adaptation, and decision making in a wide range of application domains, including self-driving vehicles, intelligent transportation, and industry automation. We investigate real-time ML in a federated edge intelligence (FEI) system, an edge computing system that implements federated learning (FL) solutions based on data samples collected and uploaded from decentralized data networks. FEI systems often exhibit heterogenous communication and computational resource distribution, as well as non-i.i.d. data samples, resulting in long model training time and inefficient resource utilization. Motivated by this fact, we propose a time-sensitive federated learning (TS-FL) framework to minimize the overall run-time for collaboratively training a shared ML model. Training acceleration solutions for both TS-FL with synchronous coordination (TS-FL-SC) and asynchronous coordination (TS-FL-ASC) are investigated. To address straggler effect in TS-FL-SC, we develop an analytical solution to characterize the impact of selecting different subsets of edge servers on the overall model training time. A server dropping-based solution is proposed to allow slow-performance edge servers to be removed from participating in model training if their impact on the resulting model accuracy is limited. A joint optimization algorithm is proposed to minimize the overall time consumption of model training by selecting participating edge servers, local epoch number. We develop an analytical expression to characterize the impact of staleness effect of asynchronous coordination and straggler effect of FL on the time consumption of TS-FL-ASC. Experimental results show that TS-FL-SC and TS-FL-ASC can provide up to 63% and 28% of reduction, in the overall model training time, respectively.

Xuancheng Zhu, Yang Yue, Shuaibing Wan et al.

Large language agents are increasingly used for social simulation, yet it remains unclear whether they can sustain coherent behavior in structured organizations, where goals must propagate through hierarchy, tasks depend on prior execution, and artifacts accumulate over long horizons. We formulate long-horizon organizational simulation as a memory-centered coordination problem and introduce TaskWeave, a hierarchical agentic framework that maintains planning states through a Formulate-Partition-Diagnose-Align cycle and grounds execution through dependency-aware trace memory. We evaluate TaskWeave in a year-long IT company simulation and compare it with other multi-agent frameworks on organizational coherence, execution grounding, and downstream enterprise NLP utility. Experiments show that TaskWeave supports coherent and long-horizon organizational dynamics while producing grounded artifacts and adapting to external environments. These findings suggest that structured simulation memory is a key mechanism for building reliable LLM-based organizational simulators.

Tommaso Giannantonio, Anna Alperovich, Piercosimo Semeraro et al.

Surgery for gliomas (intrinsic brain tumors), especially when low-grade, is challenging due to the infiltrative nature of the lesion. Currently, no real-time, intra-operative, label-free and wide-field tool is available to assist and guide the surgeon to find the relevant demarcations for these tumors. While marker-based methods exist for the high-grade glioma case, there is no convenient solution available for the low-grade case; thus, marker-free optical techniques represent an attractive option. Although RGB imaging is a standard tool in surgical microscopes, it does not contain sufficient information for tissue differentiation. We leverage the richer information from hyperspectral imaging (HSI), acquired with a snapscan camera in the 468-787 nm range, coupled to a surgical microscope, to build a deep-learning-based diagnostic tool for cancer resection with potential for intra-operative guidance. However, the main limitation of the HSI snapscan camera is the image acquisition time, limiting its widespread deployment in the operation theater. Here, we investigate the effect of HSI channel reduction and pre-selection to scope the design space for the development of cheaper and faster sensors. Neural networks are used to identify the most important spectral channels for tumor tissue differentiation, optimizing the trade-off between the number of channels and precision to enable real-time intra-surgical application. We evaluate the performance of our method on a clinical dataset that was acquired during surgery on five patients. By demonstrating the possibility to efficiently detect low-grade glioma, these results can lead to better cancer resection demarcations, potentially improving treatment effectiveness and patient outcome.

Thomas Lew, Sumeet Singh, Mario Prats et al.

We propose a framework to enable multipurpose assistive mobile robots to autonomously wipe tables to clean spills and crumbs. This problem is challenging, as it requires planning wiping actions while reasoning over uncertain latent dynamics of crumbs and spills captured via high-dimensional visual observations. Simultaneously, we must guarantee constraints satisfaction to enable safe deployment in unstructured cluttered environments. To tackle this problem, we first propose a stochastic differential equation to model crumbs and spill dynamics and absorption with a robot wiper. Using this model, we train a vision-based policy for planning wiping actions in simulation using reinforcement learning (RL). To enable zero-shot sim-to-real deployment, we dovetail the RL policy with a whole-body trajectory optimization framework to compute base and arm joint trajectories that execute the desired wiping motions while guaranteeing constraints satisfaction. We extensively validate our approach in simulation and on hardware. Video: https://youtu.be/inORKP4F3EI

Jinfeng Zhou, Zhuang Chen, Dazhen Wan et al.

In this paper, we present CharacterGLM, a series of models built upon ChatGLM, with model sizes ranging from 6B to 66B parameters. Our CharacterGLM is designed for generating Character-based Dialogues (CharacterDial), which aims to equip a conversational AI system with character customization for satisfying people's inherent social desires and emotional needs. On top of CharacterGLM, we can customize various AI characters or social agents by configuring their attributes (identities, interests, viewpoints, experiences, achievements, social relationships, etc.) and behaviors (linguistic features, emotional expressions, interaction patterns, etc.). Our model outperforms most mainstream close-source large langauge models, including the GPT series, especially in terms of consistency, human-likeness, and engagement according to manual evaluations. We will release our 6B version of CharacterGLM and a subset of training data to facilitate further research development in the direction of character-based dialogue generation.

Shuaiqi Duan, Yadong Xue, Weihan Wang et al. · tsinghua

GLM-OCR is an efficient 0.9B-parameter compact multimodal model designed for real-world document understanding. It combines a 0.4B-parameter CogViT visual encoder with a 0.5B-parameter GLM language decoder, achieving a strong balance between computational efficiency and recognition performance. To address the inefficiency of standard autoregressive decoding in deterministic OCR tasks, GLM-OCR introduces a Multi-Token Prediction (MTP) mechanism that predicts multiple tokens per step, significantly improving decoding throughput while keeping memory overhead low through shared parameters. At the system level, a two-stage pipeline is adopted: PP-DocLayout-V3 first performs layout analysis, followed by parallel region-level recognition. Extensive evaluations on public benchmarks and industrial scenarios show that GLM-OCR achieves competitive or state-of-the-art performance in document parsing, text and formula transcription, table structure recovery, and key information extraction. Its compact architecture and structured generation make it suitable for both resource-constrained edge deployment and large-scale production systems.

Yifei Wang, Tianlin Li, Xiaohan Zhang et al.

Inference optimization is a vital technique for deploying LLMs at scale. Compilation is the most widely adopted optimization technique for LLMs. While it assumes semantic equivalence between the original and compiled graphs, we first uncover its numerical side effects can be maliciously exploited to implant stealthy backdoors in LLMs. We propose a unified optimization-triggered attack framework comprising two complementary strategies. Without any modification to the compiler or hardware, one strategy flips predictions for specific inputs only when the model is compiled, while the other uses a universal trigger that remains dormant under uncompiled execution but hijacks arbitrary inputs once compilation optimization is applied. Both attacks bypass standard safety evaluations run without compilation. We empirically demonstrate that these optimization-triggered backdoors achieve attack success rates averaging 90% across four mainstream open-source LLMs and four tasks, while clean accuracy is preserved at nearly 100% under all settings. Our findings reveal a novel attack surface at the intersection of optimization and security in the LLM deployment pipeline, and we investigate practical defenses to mitigate this threat.

Xiaohan Zhang, Xingyu Li, Waqas Sultani et al.

Cross-view geo-localization aims to estimate the location of a query ground image by matching it to a reference geo-tagged aerial images database. As an extremely challenging task, its difficulties root in the drastic view changes and different capturing time between two views. Despite these difficulties, recent works achieve outstanding progress on cross-view geo-localization benchmarks. However, existing methods still suffer from poor performance on the cross-area benchmarks, in which the training and testing data are captured from two different regions. We attribute this deficiency to the lack of ability to extract the spatial configuration of visual feature layouts and models' overfitting on low-level details from the training set. In this paper, we propose GeoDTR which explicitly disentangles geometric information from raw features and learns the spatial correlations among visual features from aerial and ground pairs with a novel geometric layout extractor module. This module generates a set of geometric layout descriptors, modulating the raw features and producing high-quality latent representations. In addition, we elaborate on two categories of data augmentations, (i) Layout simulation, which varies the spatial configuration while keeping the low-level details intact. (ii) Semantic augmentation, which alters the low-level details and encourages the model to capture spatial configurations. These augmentations help to improve the performance of the cross-view geo-localization models, especially on the cross-area benchmarks. Moreover, we propose a counterfactual-based learning process to benefit the geometric layout extractor in exploring spatial information. Extensive experiments show that GeoDTR not only achieves state-of-the-art results but also significantly boosts the performance on same-area and cross-area benchmarks.

Kun Yuan, Xinmeng Huang, Yiming Chen et al.

Decentralized optimization is effective to save communication in large-scale machine learning. Although numerous algorithms have been proposed with theoretical guarantees and empirical successes, the performance limits in decentralized optimization, especially the influence of network topology and its associated weight matrix on the optimal convergence rate, have not been fully understood. While (Lu and Sa, 2021) have recently provided an optimal rate for non-convex stochastic decentralized optimization with weight matrices defined over linear graphs, the optimal rate with general weight matrices remains unclear. This paper revisits non-convex stochastic decentralized optimization and establishes an optimal convergence rate with general weight matrices. In addition, we also establish the optimal rate when non-convex loss functions further satisfy the Polyak-Lojasiewicz (PL) condition. Following existing lines of analysis in literature cannot achieve these results. Instead, we leverage the Ring-Lattice graph to admit general weight matrices while maintaining the optimal relation between the graph diameter and weight matrix connectivity. Lastly, we develop a new decentralized algorithm to nearly attain the above two optimal rates under additional mild conditions.

Han Liu, Jian Zhang, Cen Zhang et al.

Static analysis tools have evolved over time to assist in detecting bugs. However, the excessive false warnings can impede developers' productivity and confidence in the tools. Previous research efforts have explored learning-based approaches to identify bug warnings. Nevertheless, their coarse granularity, focusing on either long-term warnings or function-level alerts, is insensitive to individual bugs. Also, they rely on manually crafted features or solely on source code semantics, which is inadequate for effective learning. In this paper, we propose DeepFWI, a learning-based approach that identifies bug-sensitive warnings at a fine-grained granularity. Specifically, we design a novel LSTM-based model that captures multi-modal semantics of source code and warnings from automated static analysis tools (ASATs) and highlights their correlations with cross-attention. To tackle the data scarcity of training and evaluation, we collected a large-scale dataset of 280,273 warnings. We conducted extensive experiments on the dataset to evaluate DeepFWI. The experimental results demonstrate the effectiveness of our approach, with an F1-score 67.06% for confirming true warnings in a finer-grained manner, significantly outperforming all baselines. Additionally, to validate the practicality of DeepFWI from the perspective of developers, we applied DeepFWI to four popular open-source projects. Our approach filtered out the vast majority of warnings, while still successfully surfacing 25 true bug-related warnings that were confirmed through manual analysis.

Mayar Lotfy Mostafa, Anna Alperovich, Tommaso Giannantonio et al.

Segmenting the boundary between tumor and healthy tissue during surgical cancer resection poses a significant challenge. In recent years, Hyperspectral Imaging (HSI) combined with Machine Learning (ML) has emerged as a promising solution. However, due to the extensive information contained within the spectral domain, most ML approaches primarily classify individual HSI (super-)pixels, or tiles, without taking into account their spatial context. In this paper, we propose an improved methodology that leverages the spatial context of tiles for more robust and smoother segmentation. To address the irregular shapes of tiles, we utilize Graph Neural Networks (GNNs) to propagate context information across neighboring regions. The features for each tile within the graph are extracted using a Convolutional Neural Network (CNN), which is trained simultaneously with the subsequent GNN. Moreover, we incorporate local image quality metrics into the loss function to enhance the training procedure's robustness against low-quality regions in the training images. We demonstrate the superiority of our proposed method using a clinical ex vivo dataset consisting of 51 HSI images from 30 patients. Despite the limited dataset, the GNN-based model significantly outperforms context-agnostic approaches, accurately distinguishing between healthy and tumor tissues, even in images from previously unseen patients. Furthermore, we show that our carefully designed loss function, accounting for local image quality, results in additional improvements. Our findings demonstrate that context-aware GNN algorithms can robustly find tumor demarcations on HSI images, ultimately contributing to better surgery success and patient outcome.

Xiaohan Zhang, Waqas Sultani, Safwan Wshah

Cross-view geo-localization aims to estimate the GPS location of a query ground-view image by matching it to images from a reference database of geo-tagged aerial images. To address this challenging problem, recent approaches use panoramic ground-view images to increase the range of visibility. Although appealing, panoramic images are not readily available compared to the videos of limited Field-Of-View (FOV) images. In this paper, we present the first cross-view geo-localization method that works on a sequence of limited FOV images. Our model is trained end-to-end to capture the temporal structure that lies within the frames using the attention-based temporal feature aggregation module. To robustly tackle different sequences length and GPS noises during inference, we propose to use a sequential dropout scheme to simulate variant length sequences. To evaluate the proposed approach in realistic settings, we present a new large-scale dataset containing ground-view sequences along with the corresponding aerial-view images. Extensive experiments and comparisons demonstrate the superiority of the proposed approach compared to several competitive baselines.

Xiaohan Zhang, Xingyu Li, Waqas Sultani et al.

Cross-View Geo-Localization (CVGL) estimates the location of a ground image by matching it to a geo-tagged aerial image in a database. Recent works achieve outstanding progress on CVGL benchmarks. However, existing methods still suffer from poor performance in cross-area evaluation, in which the training and testing data are captured from completely distinct areas. We attribute this deficiency to the lack of ability to extract the geometric layout of visual features and models' overfitting to low-level details. Our preliminary work introduced a Geometric Layout Extractor (GLE) to capture the geometric layout from input features. However, the previous GLE does not fully exploit information in the input feature. In this work, we propose GeoDTR+ with an enhanced GLE module that better models the correlations among visual features. To fully explore the LS techniques from our preliminary work, we further propose Contrastive Hard Samples Generation (CHSG) to facilitate model training. Extensive experiments show that GeoDTR+ achieves state-of-the-art (SOTA) results in cross-area evaluation on CVUSA, CVACT, and VIGOR by a large margin ($16.44\%$, $22.71\%$, and $13.66\%$ without polar transformation) while keeping the same-area performance comparable to existing SOTA. Moreover, we provide detailed analyses of GeoDTR+. Our code will be available at https://gitlab.com/vail-uvm/geodtr plus.

Ahmad Arrabi, Xiaohan Zhang, Waqas Sultani et al.

Aerial imagery analysis is critical for many research fields. However, obtaining frequent high-quality aerial images is not always accessible due to its high effort and cost requirements. One solution is to use the Ground-to-Aerial (G2A) technique to synthesize aerial images from easily collectible ground images. However, G2A is rarely studied, because of its challenges, including but not limited to, the drastic view changes, occlusion, and range of visibility. In this paper, we present a novel Geometric Preserving Ground-to-Aerial (G2A) image synthesis (GPG2A) model that can generate realistic aerial images from ground images. GPG2A consists of two stages. The first stage predicts the Bird's Eye View (BEV) segmentation (referred to as the BEV layout map) from the ground image. The second stage synthesizes the aerial image from the predicted BEV layout map and text descriptions of the ground image. To train our model, we present a new multi-modal cross-view dataset, namely VIGORv2 which is built upon VIGOR with newly collected aerial images, maps, and text descriptions. Our extensive experiments illustrate that GPG2A synthesizes better geometry-preserved aerial images than existing models. We also present two applications, data augmentation for cross-view geo-localization and sketch-based region search, to further verify the effectiveness of our GPG2A. The code and data will be publicly available.

Tengfei Liu, Yang Shi, Xuanyu Zhu et al.

Audio-visual generation is rapidly advancing from short clips to minute-long content, while existing evaluation protocols remain largely confined to short-form settings. Existing benchmarks primarily focus on 5--10 second text-conditioned generation and rarely support unified evaluation across text, image, and video conditioning modalities. Moreover, they provide limited insight into how identity consistency, narrative coherence, and audio-visual alignment degrade over extended temporal horizons. To bridge this gap, we introduce LongAV-Compass, a systematic benchmark for minute-long audio-visual generation. LongAV-Compass contains 284 curated test cases spanning text-to-audio-video (T2AV), image-to-audio-video (I2AV), and video-to-audio-video (V2AV), organized by application scenario and generation complexity. The benchmark combines taxonomy-guided benchmark construction with a unified evaluation framework that integrates MLLM-assisted assessment with complementary perceptual and multimodal metrics, including DINO-v2, ArcFace, CLIP, and ImageBind. The framework evaluates more than 20 fine-grained dimensions covering within-segment quality, cross-segment consistency, global narrative coherence, semantic alignment, and audio-visual synchronization. Through experiments on 11 representative models together with human-alignment validation, LongAV-Compass provides a diagnostic testbed for analyzing the limitations of current systems in sustaining coherent, semantically aligned, and temporally consistent minute-scale audio-visual generation across diverse input modalities.

Yan Ding, Xiaohan Zhang, Saeid Amiri et al.

Automated task planning algorithms have been developed to help robots complete complex tasks that require multiple actions. Most of those algorithms have been developed for "closed worlds" assuming complete world knowledge is provided. However, the real world is generally open, and the robots frequently encounter unforeseen situations that can potentially break the planner's completeness. This paper introduces a novel algorithm (COWP) for open-world task planning and situation handling that dynamically augments the robot's action knowledge with task-oriented common sense. In particular, common sense is extracted from Large Language Models based on the current task at hand and robot skills. For systematic evaluations, we collected a dataset that includes 561 execution-time situations in a dining domain, where each situation corresponds to a state instance of a robot being potentially unable to complete a task using a solution that normally works. Experimental results show that our approach significantly outperforms competitive baselines from the literature in the success rate of service tasks. Additionally, we have demonstrated COWP using a mobile manipulator. The project website is available at: https://cowplanning.github.io/, where a more detailed version can also be found. This version has been accepted for publication in Autonomous Robots.

Xiaokai Bai, Jiahao Cheng, Songkai Wang et al.

4D radar measurements offer an affordable and weather-robust solution for 3D perception. However, the inherent sparsity and noise of radar point clouds present significant challenges for accurate 3D object detection, underscoring the need for effective and robust point clouds densification. Despite recent progress, existing densification methods often fail to address the extreme sparsity of 4D radar point clouds and exhibit limited robustness when processing scenes with a small number of points. In this paper, we propose SD4R, a novel framework that transforms sparse radar point clouds into dense representations. SD4R begins by utilizing a foreground point generator (FPG) to mitigate noise propagation and produce densified point clouds. Subsequently, a logit-query encoder (LQE) enhances conventional pillarization, resulting in robust feature representations. Through these innovations, our SD4R demonstrates strong capability in both noise reduction and foreground point densification. Extensive experiments conducted on the publicly available View-of-Delft dataset demonstrate that SD4R achieves state-of-the-art performance. Source code is available at https://github.com/lancelot0805/SD4R.

Xiaokai Bai, Lianqing Zheng, Si-Yuan Cao et al.

4D millimeter-wave radar has emerged as a promising sensing modality for autonomous driving due to its robustness and affordability. However, its sparse and weak geometric cues make reliable instance activation difficult, limiting the effectiveness of existing radar-camera fusion paradigms. BEV-level fusion offers global scene understanding but suffers from weak instance focus, while perspective-level fusion captures instance details but lacks holistic context. To address these limitations, we propose SIFormer, a scene-instance aware transformer for 3D object detection using 4D radar and camera. SIFormer first suppresses background noise during view transformation through segmentation- and depth-guided localization. It then introduces a cross-view activation mechanism that injects 2D instance cues into BEV space, enabling reliable instance awareness under weak radar geometry. Finally, a transformer-based fusion module aggregates complementary image semantics and radar geometry for robust perception. As a result, with the aim of enhancing instance awareness, SIFormer bridges the gap between the two paradigms, combining their complementary strengths to address inherent sparse nature of radar and improve detection accuracy. Experiments demonstrate that SIFormer achieves state-of-the-art performance on View-of-Delft, TJ4DRadSet and NuScenes datasets. Source code is available at github.com/shawnnnkb/SIFormer.

Hanyu Lai, Xiao Liu, Iat Long Iong et al. · tsinghua

Large language models (LLMs) have fueled many intelligent web agents, but most existing ones perform far from satisfying in real-world web navigation tasks due to three factors: (1) the complexity of HTML text data (2) versatility of actions on webpages, and (3) task difficulty due to the open-domain nature of the web. In light of these challenges, we develop the open AutoWebGLM based on ChatGLM3-6B. AutoWebGLM can serve as a powerful automated web navigation agent that outperform GPT-4. Inspired by human browsing patterns, we first design an HTML simplification algorithm to represent webpages with vital information preserved succinctly. We then employ a hybrid human-AI method to build web browsing data for curriculum training. Finally, we bootstrap the model by reinforcement learning and rejection sampling to further facilitate webpage comprehension, browser operations, and efficient task decomposition by itself. For comprehensive evaluation, we establish a bilingual benchmark -- AutoWebBench -- for real-world web navigation tasks. We evaluate AutoWebGLM across diverse web navigation benchmarks, demonstrating its potential to tackle challenging tasks in real environments. Related code, model, and data are released at \url{https://github.com/THUDM/AutoWebGLM}.

Xiaohan Zhang, Tian Gao, Mingyue Cheng et al.

Time series forecasting plays a critical role in high-stakes domains such as energy, healthcare, and climate. Although recent advances have improved accuracy, most approaches still treat forecasting as a static one-time mapping task, lacking the interaction, reasoning, and adaptability of human experts. This gap limits their usefulness in complex real-world environments. To address this, we propose AlphaCast, a human wisdom-large language model (LLM) intelligence co-reasoning framework that redefines forecasting as an interactive process. The key idea is to enable step-by-step collaboration between human wisdom and LLM intelligence to jointly prepare, generate, and verify forecasts. The framework consists of two stages: (1) automated prediction preparation, where AlphaCast builds a multi-source cognitive foundation comprising a feature set that captures key statistics and time patterns, a domain knowledge base distilled from corpora and historical series, a contextual repository that stores rich information for each time window, and a case base that retrieves optimal strategies via pattern clustering and matching; and (2) generative reasoning and reflective optimization, where AlphaCast integrates statistical temporal features, prior knowledge, contextual information, and forecasting strategies, triggering a meta-reasoning loop for continuous self-correction and strategy refinement. Extensive experiments on short- and long-term datasets show that AlphaCast consistently outperforms state-of-the-art baselines in predictive accuracy. Code is available at this repository: https://github.com/SkyeGT/AlphaCast_Official .

GLM-4. 5 Team, Aohan Zeng, Xin Lv et al.

We present GLM-4.5, an open-source Mixture-of-Experts (MoE) large language model with 355B total parameters and 32B activated parameters, featuring a hybrid reasoning method that supports both thinking and direct response modes. Through multi-stage training on 23T tokens and comprehensive post-training with expert model iteration and reinforcement learning, GLM-4.5 achieves strong performance across agentic, reasoning, and coding (ARC) tasks, scoring 70.1% on TAU-Bench, 91.0% on AIME 24, and 64.2% on SWE-bench Verified. With much fewer parameters than several competitors, GLM-4.5 ranks 3rd overall among all evaluated models and 2nd on agentic benchmarks. We release both GLM-4.5 (355B parameters) and a compact version, GLM-4.5-Air (106B parameters), to advance research in reasoning and agentic AI systems. Code, models, and more information are available at https://github.com/zai-org/GLM-4.5.

Yuqi Chen, Xiaohan Zhang, Ahmad Arrabi et al.

Natural-language Guided Cross-view Geo-localization (NGCG) aims to retrieve geo-tagged satellite imagery using textual descriptions of ground scenes. While recent NGCG methods commonly rely on CLIP-style dual-encoder architectures, they often suffer from weak cross-modal generalization and require complex architectural designs. In contrast, Multimodal Large Language Models (MLLMs) offer powerful semantic reasoning capabilities but are not directly optimized for retrieval tasks. In this work, we present a simple yet effective framework to adapt MLLMs for NGCG via parameter-efficient finetuning. Our approach optimizes latent representations within the MLLM while preserving its pretrained multimodal knowledge, enabling strong cross-modal alignment without redesigning model architectures. Through systematic analysis of diverse variables, from model backbone to feature aggregation, we provide practical and generalizable insights for leveraging MLLMs in NGCG. Our method achieves SOTA on GeoText-1652 with a 12.2% improvement in Text-to-Image Recall@1 and secures top performance in 5 out of 12 subtasks on CVG-Text, all while surpassing baselines with far fewer trainable parameters. These results position MLLMs as a robust foundation for semantic cross-view retrieval and pave the way for MLLM-based NGCG to be adopted as a scalable, powerful alternative to traditional dual-encoder designs. Project page and code are available at https://yuqichen888.github.io/NGCG-MLLMs-web/.

Di Fan, Mingyang Liu, Xiaohan Zhang et al.

A novel human emotion recognition method based on automatically selected Galvanic Skin Response (GSR) signal features and SVM is proposed in this paper. GSR signals were acquired by e-Health Sensor Platform V2.0. Then, the data is de-noised by wavelet function and normalized to get rid of the individual difference. 30 features are extracted from the normalized data, however, directly using of these features will lead to a low recognition rate. In order to gain the optimized features, a covariance based feature selection is employed in our method. Finally, a SVM with input of the optimized features is utilized to achieve the human emotion recognition. The experimental results indicate that the proposed method leads to good human emotion recognition, and the recognition accuracy is more than 66.67%.

Jiazheng Xu, Yu Huang, Jiale Cheng et al. · tsinghua

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. All code and datasets are provided at https://github.com/THUDM/VisionReward.

Yifan Xu, Xiao Liu, Xinghan Liu et al. · tsinghua

Large language models (LLMs) have shown excellent mastering of human language, but still struggle in real-world applications that require mathematical problem-solving. While many strategies and datasets to enhance LLMs' mathematics are developed, it remains a challenge to simultaneously maintain and improve both language and mathematical capabilities in deployed LLM systems.In this work, we tailor the Self-Critique pipeline, which addresses the challenge in the feedback learning stage of LLM alignment. We first train a general Math-Critique model from the LLM itself to provide feedback signals. Then, we sequentially employ rejective fine-tuning and direct preference optimization over the LLM's own generations for data collection. Based on ChatGLM3-32B, we conduct a series of experiments on both academic and our newly created challenging dataset, MathUserEval. Results show that our pipeline significantly enhances the LLM's mathematical problem-solving while still improving its language ability, outperforming LLMs that could be two times larger. Related techniques have been deployed to ChatGLM\footnote{\url{https://chatglm.cn}}, an online serving LLM. Related evaluation dataset and scripts are released at \url{https://github.com/THUDM/ChatGLM-Math}.

Yannan He, Garvita Tiwari, Xiaohan Zhang et al.

We introduce MoLingo, a text-to-motion (T2M) model that generates realistic, lifelike human motion by denoising in a continuous latent space. Recent works perform latent space diffusion, either on the whole latent at once or auto-regressively over multiple latents. In this paper, we study how to make diffusion on continuous motion latents work best. We focus on two questions: (1) how to build a semantically aligned latent space so diffusion becomes more effective, and (2) how to best inject text conditioning so the motion follows the description closely. We propose a semantic-aligned motion encoder trained with frame-level text labels so that latents with similar text meaning stay close, which makes the latent space more diffusion-friendly. We also compare single-token conditioning with a multi-token cross-attention scheme and find that cross-attention gives better motion realism and text-motion alignment. With semantically aligned latents, auto-regressive generation, and cross-attention text conditioning, our model sets a new state of the art in human motion generation on standard metrics and in a user study. We will release our code and models for further research and downstream usage.

GLM-V Team, Wenyi Hong, Wenmeng Yu et al.

We present GLM-4.1V-Thinking and GLM-4.5V, 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. Code, models and more information are released at https://github.com/zai-org/GLM-V.

Huijun Zhou, Xiaohan Zhang, Haozhe Zhang et al.

The Model Context Protocol (MCP) is emerging as a common interface connecting large language models (LLMs) with external services. Remote deployments are becoming increasingly important as agents connect to user-linked online services, such as social, productivity, and financial services. In such deployments, the authentication boundary between MCP clients and remote servers becomes security-critical, yet remains underexplored. We present the first measurement study of authentication security in real-world remote MCP servers. We identify 7,973 live remote MCP servers, finding that 40.55% expose tools without authentication. Among authenticated servers, OAuth is the dominant authorization mechanism for reaching remote services, and OAuth deployments in the MCP ecosystem commonly exhibit three characteristics: open client environments, dynamic client registration, and delegated authorization. These characteristics distinguish MCP deployments from traditional OAuth and introduce new attack surfaces. Guided by this observation, we derive a taxonomy of authentication flaws comprising three MCP-specific categories and conventional OAuth misconfigurations, for a total of four categories and nine concrete flaw types. To evaluate these flaws at scale, we implement a semi-automated detection framework that combines passive traffic inspection with active dynamic probing. Applying it to 119 testable real-world OAuth-enabled MCP servers, we find that each server exhibits at least one flaw, with a total of 325 flaws identified, among which dynamic client registration flaws affect 96.6% of tested servers. Many of these flaws can lead to sensitive information leakage and account takeover. Through responsible disclosure, we obtained 9 CVE IDs. Our findings expose pervasive authentication weaknesses in the MCP ecosystem and underscore the urgent need for hardened OAuth-based remote deployments.

Jingyuan Sun, Xiaohan Zhang, Marie-Francine Moens

To understand the algorithm that supports the human brain's language representation, previous research has attempted to predict neural responses to linguistic stimuli using embeddings generated by artificial neural networks (ANNs), a process known as neural encoding. However, most of these studies have focused on probing neural representations of Germanic languages, such as English, with unsupervised ANNs. In this paper, we propose to bridge the gap between human brain and supervised ANN representations of the Chinese language. Specifically, we investigate how task tuning influences a pretained Transformer for neural encoding and which tasks lead to the best encoding performances. We generate supervised representations on eight Natural Language Understanding (NLU) tasks using prompt-tuning, a technique that is seldom explored in neural encoding for language. We demonstrate that prompt-tuning yields representations that better predict neural responses to Chinese stimuli than traditional fine-tuning on four tasks. Furthermore, we discover that tasks that require a fine-grained processing of concepts and entities lead to representations that are most predictive of brain activation patterns. Additionally, we reveal that the proportion of tuned parameters highly influences the neural encoding performance of fine-tuned models. Overall, our experimental findings could help us better understand the relationship between supervised artificial and brain language representations.

Qianrui Zhou, Hua Xu, Hao Li et al. · tsinghua

Multimodal intent recognition aims to leverage diverse modalities such as expressions, body movements and tone of speech to comprehend user's intent, constituting a critical task for understanding human language and behavior in real-world multimodal scenarios. Nevertheless, the majority of existing methods ignore potential correlations among different modalities and own limitations in effectively learning semantic features from nonverbal modalities. In this paper, we introduce a token-level contrastive learning method with modality-aware prompting (TCL-MAP) to address the above challenges. To establish an optimal multimodal semantic environment for text modality, we develop a modality-aware prompting module (MAP), which effectively aligns and fuses features from text, video and audio modalities with similarity-based modality alignment and cross-modality attention mechanism. Based on the modality-aware prompt and ground truth labels, the proposed token-level contrastive learning framework (TCL) constructs augmented samples and employs NT-Xent loss on the label token. Specifically, TCL capitalizes on the optimal textual semantic insights derived from intent labels to guide the learning processes of other modalities in return. Extensive experiments show that our method achieves remarkable improvements compared to state-of-the-art methods. Additionally, ablation analyses demonstrate the superiority of the modality-aware prompt over the handcrafted prompt, which holds substantial significance for multimodal prompt learning. The codes are released at https://github.com/thuiar/TCL-MAP.

Ayesh Abu Lehyeh, Xiaohan Zhang, Ahmad Arrabi et al.

Accurate and fast localization is vital for safe autonomous navigation in GPS-denied areas. Fine-Grained Cross-View Geolocalization (FG-CVG) aims to estimate the precise 2-Degree-of-Freedom (2-DoF) location of a ground image relative to a satellite image. However, current methods force a difficult trade-off, with high-accuracy models being slow for real-time use. In this paper, we introduce GeoFlow, a new approach that offers a lightweight and highly efficient framework that breaks this accuracy-speed trade-off. Our technique learns a direct probabilistic mapping, predicting the displacement (in distance and direction) required to correct any given location hypothesis. This is complemented by our novel inference algorithm, Iterative Refinement Sampling (IRS). Instead of trusting a single prediction, IRS refines a population of hypotheses, allowing them to iteratively 'flow' from random starting points to a robust, converged consensus. Even its iterative nature, this approach offers flexible inference-time scaling, allowing a direct trade-off between performance and computation without any re-training. Experiments on the KITTI and VIGOR datasets show that GeoFlow achieves state-of-the-art efficiency, running at real-time speeds of 29 FPS while maintaining competitive localization accuracy. This work opens a new path for the development of practical real-time geolocalization systems.

Omkar Patil, Ondrej Biza, Thomas Weng et al.

What happens when a pretrained generative robot policy is provided a constant initial noise as input, rather than repeatedly sampling it from a Gaussian? We demonstrate that the performance of a pretrained, frozen diffusion or flow matching policy can be improved with respect to a downstream reward by swapping the sampling of initial noise from the prior distribution (typically isotropic Gaussian) with a well-chosen, constant initial noise input -- a golden ticket. We propose a search method to find golden tickets using Monte-Carlo policy evaluation that keeps the pretrained policy frozen, does not train any new networks, and is applicable to all diffusion/flow matching policies (and therefore many VLAs). Our approach to policy improvement makes no assumptions beyond being able to inject initial noise into the policy and calculate (sparse) task rewards of episode rollouts, making it deployable with no additional infrastructure or models. Our method improves the performance of policies in 38 out of 43 tasks across simulated and real-world robot manipulation benchmarks, with relative improvements in success rate by up to 58% for some simulated tasks, and 60% within 50 search episodes for real-world tasks. We also show unique benefits of golden tickets for multi-task settings: the diversity of behaviors from different tickets naturally defines a Pareto frontier for balancing different objectives (e.g., speed, success rates); in VLAs, we find that a golden ticket optimized for one task can also boost performance in other related tasks. We release a codebase with pretrained policies and golden tickets for simulation benchmarks using VLAs, diffusion policies, and flow matching policies.

Lucen Zhong, Zhengxiao Du, Xiaohan Zhang et al.

Enhancing large language models (LLMs) with real-time APIs can help generate more accurate and up-to-date responses. However, evaluating the function calling abilities of LLMs in real-world scenarios remains under-explored due to the complexity of data collection and evaluation. In this work, we introduce ComplexFuncBench, a benchmark for complex function calling across five real-world scenarios. Compared to existing benchmarks, ComplexFuncBench encompasses multi-step and constrained function calling, which requires long-parameter filing, parameter value reasoning, and 128k long context. Additionally, we propose an automatic framework, ComplexEval, for quantitatively evaluating complex function calling tasks. Through comprehensive experiments, we demonstrate the deficiencies of state-of-the-art LLMs in function calling and suggest future directions for optimizing these capabilities. The data and code are available at \url{https://github.com/THUDM/ComplexFuncBench}.

Shudan Zhang, Hanlin Zhao, Xiao Liu et al. · tsinghua

Large language models (LLMs) have manifested strong ability to generate codes for productive activities. However, current benchmarks for code synthesis, such as HumanEval, MBPP, and DS-1000, are predominantly oriented towards introductory tasks on algorithm and data science, insufficiently satisfying challenging requirements prevalent in real-world coding. To fill this gap, we propose NaturalCodeBench (NCB), a challenging code benchmark designed to mirror the complexity and variety of scenarios in real coding tasks. NCB comprises 402 high-quality problems in Python and Java, meticulously selected from natural user queries from online coding services, covering 6 different domains. Noting the extraordinary difficulty in creating testing cases for real-world queries, we also introduce a semi-automated pipeline to enhance the efficiency of test case construction. Comparing with manual solutions, it achieves an efficiency increase of more than 4 times. Our systematic experiments on 39 LLMs find that performance gaps on NCB between models with close HumanEval scores could still be significant, indicating a lack of focus on practical code synthesis scenarios or over-specified optimization on HumanEval. On the other hand, even the best-performing GPT-4 is still far from satisfying on NCB. The evaluation toolkit and development set are available at https://github.com/THUDM/NaturalCodeBench.

Yuanmin Huang, Wenxuan Li, Mi Zhang et al.

Deep neural networks have recently achieved notable progress in 3D point cloud recognition, yet their vulnerability to adversarial perturbations poses critical security challenges in practical deployments. Conventional defense mechanisms struggle to address the evolving landscape of multifaceted attack patterns. Through systematic analysis of existing defenses, we identify that their unsatisfactory performance primarily originates from an entangled feature space, where adversarial attacks can be performed easily. To this end, we present 3D-ANC, a novel approach that capitalizes on the Neural Collapse (NC) mechanism to orchestrate discriminative feature learning. In particular, NC depicts where last-layer features and classifier weights jointly evolve into a simplex equiangular tight frame (ETF) arrangement, establishing maximally separable class prototypes. However, leveraging this advantage in 3D recognition confronts two substantial challenges: (1) prevalent class imbalance in point cloud datasets, and (2) complex geometric similarities between object categories. To tackle these obstacles, our solution combines an ETF-aligned classification module with an adaptive training framework consisting of representation-balanced learning (RBL) and dynamic feature direction loss (FDL). 3D-ANC seamlessly empowers existing models to develop disentangled feature spaces despite the complexity in 3D data distribution. Comprehensive evaluations state that 3D-ANC significantly improves the robustness of models with various structures on two datasets. For instance, DGCNN's classification accuracy is elevated from 27.2% to 80.9% on ModelNet40 -- a 53.7% absolute gain that surpasses leading baselines by 34.0%.

Ran Gong, Xiaohan Zhang, Jinghuan Shang et al.

Generalist robot learning remains constrained by data: large-scale, diverse, and high-quality interaction data are expensive to collect in the real world. While simulation has become a promising way for scaling up data collection, the related tasks, including simulation task design, task-aware scene generation, expert demonstration synthesis, and sim-to-real transfer, still demand substantial human effort. We present AnyTask, an automated framework that pairs massively parallel GPU simulation with foundation models to design diverse manipulation tasks and synthesize robot data. We introduce three AnyTask agents for generating expert demonstrations aiming to solve as many tasks as possible: 1) ViPR, a novel task and motion planning agent with VLM-in-the-loop Parallel Refinement; 2) ViPR-Eureka, a reinforcement learning agent with generated dense rewards and LLM-guided contact sampling; 3) ViPR-RL, a hybrid planning and learning approach that jointly produces high-quality demonstrations with only sparse rewards. We train behavior cloning policies on generated data, validate them in simulation, and deploy them directly on real robot hardware. The policies generalize to novel object poses, achieving 44% average success across a suite of real-world pick-and-place, drawer opening, contact-rich pushing, and long-horizon manipulation tasks. Our project website is at https://anytask.rai-inst.com .

Xiaohan Zhang, Sebastian Starke, Vladimir Guzov et al.

Synthesizing natural human motion that adapts to complex environments while allowing creative control remains a fundamental challenge in motion synthesis. Existing models often fall short, either by assuming flat terrain or lacking the ability to control motion semantics through text. To address these limitations, we introduce SCENIC, a diffusion model designed to generate human motion that adapts to dynamic terrains within virtual scenes while enabling semantic control through natural language. The key technical challenge lies in simultaneously reasoning about complex scene geometry while maintaining text control. This requires understanding both high-level navigation goals and fine-grained environmental constraints. The model must ensure physical plausibility and precise navigation across varied terrain, while also preserving user-specified text control, such as ``carefully stepping over obstacles" or ``walking upstairs like a zombie." Our solution introduces a hierarchical scene reasoning approach. At its core is a novel scene-dependent, goal-centric canonicalization that handles high-level goal constraint, and is complemented by an ego-centric distance field that captures local geometric details. This dual representation enables our model to generate physically plausible motion across diverse 3D scenes. By implementing frame-wise text alignment, our system achieves seamless transitions between different motion styles while maintaining scene constraints. Experiments demonstrate our novel diffusion model generates arbitrarily long human motions that both adapt to complex scenes with varying terrain surfaces and respond to textual prompts. Additionally, we show SCENIC can generalize to four real-scene datasets. Our code, dataset, and models will be released at \url{https://virtualhumans.mpi-inf.mpg.de/scenic/}.

Zifan Xu, Ran Gong, Maria Vittoria Minniti et al.

Learning generalizable and robust behavior cloning policies requires large volumes of high-quality robotics data. While human demonstrations (e.g., through teleoperation) serve as the standard source for expert behaviors, acquiring such data at scale in the real world is prohibitively expensive. This paper introduces ExpertGen, a framework that automates expert policy learning in simulation to enable scalable sim-to-real transfer. ExpertGen first initializes a behavior prior using a diffusion policy trained on imperfect demonstrations, which may be synthesized by large language models or provided by humans. Reinforcement learning is then used to steer this prior toward high task success by optimizing the diffusion model's initial noise while keep original policy frozen. By keeping the pretrained diffusion policy frozen, ExpertGen regularizes exploration to remain within safe, human-like behavior manifolds, while also enabling effective learning with only sparse rewards. Empirical evaluations on challenging manipulation benchmarks demonstrate that ExpertGen reliably produces high-quality expert policies with no reward engineering. On industrial assembly tasks, ExpertGen achieves a 90.5% overall success rate, while on long-horizon manipulation tasks it attains 85% overall success, outperforming all baseline methods. The resulting policies exhibit dexterous control and remain robust across diverse initial configurations and failure states. To validate sim-to-real transfer, the learned state-based expert policies are further distilled into visuomotor policies via DAgger and successfully deployed on real robotic hardware.

Yiming Pan, Chengwei Hu, Xuancheng Huang et al.

Large language models (LLMs) have demonstrated strong potential in agentic tasks, particularly in slide generation. However, slide generation poses a fundamental challenge: the generation process is text-centric, whereas its quality is governed by visual aesthetics. This modality gap leads current models to frequently produce slides with aesthetically suboptimal layouts. Existing solutions typically rely either on heavy visual reflection, which incurs high inference cost yet yields limited gains; or on fine-tuning with large-scale datasets, which still provides weak and indirect aesthetic supervision. In contrast, the explicit use of aesthetic principles as supervision remains unexplored. In this work, we present AeSlides, a reinforcement learning framework with verifiable rewards for Aesthetic layout supervision in Slide generation. We introduce a suite of meticulously designed verifiable metrics to quantify slide layout quality, capturing key layout issues in an accurate, efficient, and low-cost manner. Leveraging these verifiable metrics, we develop a GRPO-based reinforcement learning method that directly optimizes slide generation models for aesthetically coherent layouts. With only 5K training prompts on GLM-4.7-Flash, AeSlides improves aspect ratio compliance from 36% to 85%, while reducing whitespace by 44%, element collisions by 43%, and visual imbalance by 28%. Human evaluation further shows a substantial improvement in overall quality, increasing scores from 3.31 to 3.56 (+7.6%), outperforming both model-based reward optimization and reflection-based agentic approaches, and even edging out Claude-Sonnet-4.5. These results demonstrate that such a verifiable aesthetic paradigm provides an efficient and scalable approach to aligning slide generation with human aesthetic preferences. Our repository is available at https://github.com/ympan0508/aeslides.