Ce Zheng

Qingxiu Dong, Lei Li, Damai Dai et al. · cmu, pku

With the increasing capabilities of large language models (LLMs), in-context learning (ICL) has emerged as a new paradigm for natural language processing (NLP), where LLMs make predictions based on contexts augmented with a few examples. It has been a significant trend to explore ICL to evaluate and extrapolate the ability of LLMs. In this paper, we aim to survey and summarize the progress and challenges of ICL. We first present a formal definition of ICL and clarify its correlation to related studies. Then, we organize and discuss advanced techniques, including training strategies, prompt designing strategies, and related analysis. Additionally, we explore various ICL application scenarios, such as data engineering and knowledge updating. Finally, we address the challenges of ICL and suggest potential directions for further research. We hope that our work can encourage more research on uncovering how ICL works and improving ICL.

Qitao Zhao, Ce Zheng, Mengyuan Liu et al.

Recently, transformer-based methods have gained significant success in sequential 2D-to-3D lifting human pose estimation. As a pioneering work, PoseFormer captures spatial relations of human joints in each video frame and human dynamics across frames with cascaded transformer layers and has achieved impressive performance. However, in real scenarios, the performance of PoseFormer and its follow-ups is limited by two factors: (a) The length of the input joint sequence; (b) The quality of 2D joint detection. Existing methods typically apply self-attention to all frames of the input sequence, causing a huge computational burden when the frame number is increased to obtain advanced estimation accuracy, and they are not robust to noise naturally brought by the limited capability of 2D joint detectors. In this paper, we propose PoseFormerV2, which exploits a compact representation of lengthy skeleton sequences in the frequency domain to efficiently scale up the receptive field and boost robustness to noisy 2D joint detection. With minimum modifications to PoseFormer, the proposed method effectively fuses features both in the time domain and frequency domain, enjoying a better speed-accuracy trade-off than its precursor. Extensive experiments on two benchmark datasets (i.e., Human3.6M and MPI-INF-3DHP) demonstrate that the proposed approach significantly outperforms the original PoseFormer and other transformer-based variants. Code is released at \url{https://github.com/QitaoZhao/PoseFormerV2}.

Ce Zheng, Matias Mendieta, Chen Chen

Facial expression recognition (FER) is an important task in computer vision, having practical applications in areas such as human-computer interaction, education, healthcare, and online monitoring. In this challenging FER task, there are three key issues especially prevalent: inter-class similarity, intra-class discrepancy, and scale sensitivity. While existing works typically address some of these issues, none have fully addressed all three challenges in a unified framework. In this paper, we propose a two-stream Pyramid crOss-fuSion TransformER network (POSTER), that aims to holistically solve all three issues. Specifically, we design a transformer-based cross-fusion method that enables effective collaboration of facial landmark features and image features to maximize proper attention to salient facial regions. Furthermore, POSTER employs a pyramid structure to promote scale invariance. Extensive experimental results demonstrate that our POSTER achieves new state-of-the-art results on RAF-DB (92.05%), FERPlus (91.62%), as well as AffectNet 7 class (67.31%) and 8 class (63.34%). The code is available at https://github.com/zczcwh/POSTER.

Kaikai An, Ce Zheng, Bofei Gao et al. · pku

Frame identification aims to find semantic frames associated with target words in a sentence. Recent researches measure the similarity or matching score between targets and candidate frames by modeling frame definitions. However, they either lack sufficient representation learning of the definitions or face challenges in efficiently selecting the most suitable frame from over 1000 candidate frames. Moreover, commonly used lexicon filtering ($lf$) to obtain candidate frames for the target may ignore out-of-vocabulary targets and cause inadequate frame modeling. In this paper, we propose CoFFTEA, a $\underline{Co}$arse-to-$\underline{F}$ine $\underline{F}$rame and $\underline{T}$arget $\underline{E}$ncoders $\underline{A}$rchitecture. With contrastive learning and dual encoders, CoFFTEA efficiently and effectively models the alignment between frames and targets. By employing a coarse-to-fine curriculum learning procedure, CoFFTEA gradually learns to differentiate frames with varying degrees of similarity. Experimental results demonstrate that CoFFTEA outperforms previous models by 0.93 overall scores and 1.53 R@1 without $lf$. Further analysis suggests that CoFFTEA can better model the relationships between frame and frame, as well as target and target. The code for our approach is available at https://github.com/pkunlp-icler/COFFTEA.

Ce Zheng, Xudong Chen, Runxin Xu et al.

Frame semantic parsing is a fundamental NLP task, which consists of three subtasks: frame identification, argument identification and role classification. Most previous studies tend to neglect relations between different subtasks and arguments and pay little attention to ontological frame knowledge defined in FrameNet. In this paper, we propose a Knowledge-guided Incremental semantic parser with Double-graph (KID). We first introduce Frame Knowledge Graph (FKG), a heterogeneous graph containing both frames and FEs (Frame Elements) built on the frame knowledge so that we can derive knowledge-enhanced representations for frames and FEs. Besides, we propose Frame Semantic Graph (FSG) to represent frame semantic structures extracted from the text with graph structures. In this way, we can transform frame semantic parsing into an incremental graph construction problem to strengthen interactions between subtasks and relations between arguments. Our experiments show that KID outperforms the previous state-of-the-art method by up to 1.7 F1-score on two FrameNet datasets. Our code is availavle at https://github.com/PKUnlp-icler/KID.

Shengnan Hu, Ce Zheng, Zixiang Zhou et al.

Human-centric visual understanding is an important desideratum for effective human-robot interaction. In order to navigate crowded public places, social robots must be able to interpret the activity of the surrounding humans. This paper addresses one key aspect of human-centric visual understanding, multi-person pose estimation. Achieving good performance on multi-person pose estimation in crowded scenes is difficult due to the challenges of occluded joints and instance separation. In order to tackle these challenges and overcome the limitations of image features in representing invisible body parts, we propose a novel prompt-based pose inference strategy called LAMP (Language Assisted Multi-person Pose estimation). By utilizing the text representations generated by a well-trained language model (CLIP), LAMP can facilitate the understanding of poses on the instance and joint levels, and learn more robust visual representations that are less susceptible to occlusion. This paper demonstrates that language-supervised training boosts the performance of single-stage multi-person pose estimation, and both instance-level and joint-level prompts are valuable for training. The code is available at https://github.com/shengnanh20/LAMP.

Qucheng Peng, Ce Zheng, Chen Chen

Human Pose Estimation (HPE) is widely used in various fields, including motion analysis, healthcare, and virtual reality. However, the great expenses of labeled real-world datasets present a significant challenge for HPE. To overcome this, one approach is to train HPE models on synthetic datasets and then perform domain adaptation (DA) on real-world data. Unfortunately, existing DA methods for HPE neglect data privacy and security by using both source and target data in the adaptation process. To this end, we propose a new task, named source-free domain adaptive HPE, which aims to address the challenges of cross-domain learning of HPE without access to source data during the adaptation process. We further propose a novel framework that consists of three models: source model, intermediate model, and target model, which explores the task from both source-protect and target-relevant perspectives. The source-protect module preserves source information more effectively while resisting noise, and the target-relevant module reduces the sparsity of spatial representations by building a novel spatial probability space, and pose-specific contrastive learning and information maximization are proposed on the basis of this space. Comprehensive experiments on several domain adaptive HPE benchmarks show that the proposed method outperforms existing approaches by a considerable margin. The codes are available at https://github.com/davidpengucf/SFDAHPE.

Ce Zheng, Yiming Wang, Baobao Chang

Frame Semantic Role Labeling (FSRL) identifies arguments and labels them with frame semantic roles defined in FrameNet. Previous researches tend to divide FSRL into argument identification and role classification. Such methods usually model role classification as naive multi-class classification and treat arguments individually, which neglects label semantics and interactions between arguments and thus hindering performance and generalization of models. In this paper, we propose a query-based framework named ArGument Extractor with Definitions in FrameNet (AGED) to mitigate these problems. Definitions of frames and frame elements (FEs) in FrameNet can be used to query arguments in text. Encoding text-definition pairs can guide models in learning label semantics and strengthening argument interactions. Experiments show that AGED outperforms previous state-of-the-art by up to 1.3 F1-score in two FrameNet datasets and the generalization power of AGED in zero-shot and fewshot scenarios. Our code and technical appendix is available at https://github.com/PKUnlp-icler/AGED.

Bofei Gao, Feifan Song, Yibo Miao et al. · pku

Large Language Models (LLMs) exhibit remarkably powerful capabilities. One of the crucial factors to achieve success is aligning the LLM's output with human preferences. This alignment process often requires only a small amount of data to efficiently enhance the LLM's performance. While effective, research in this area spans multiple domains, and the methods involved are relatively complex to understand. The relationships between different methods have been under-explored, limiting the development of the preference alignment. In light of this, we break down the existing popular alignment strategies into different components and provide a unified framework to study the current alignment strategies, thereby establishing connections among them. In this survey, we decompose all the strategies in preference learning into four components: model, data, feedback, and algorithm. This unified view offers an in-depth understanding of existing alignment algorithms and also opens up possibilities to synergize the strengths of different strategies. Furthermore, we present detailed working examples of prevalent existing algorithms to facilitate a comprehensive understanding for the readers. Finally, based on our unified perspective, we explore the challenges and future research directions for aligning large language models with human preferences.

Chen Sun, Shiyao Ma, Ce Zheng et al.

User selection has become crucial for decreasing the communication costs of federated learning (FL) over wireless networks. However, centralized user selection causes additional system complexity. This study proposes a network intrinsic approach of distributed user selection that leverages the radio resource competition mechanism in random access. Taking the carrier sensing multiple access (CSMA) mechanism as an example of random access, we manipulate the contention window (CW) size to prioritize certain users for obtaining radio resources in each round of training. Training data bias is used as a target scenario for FL with user selection. Prioritization is based on the distance between the newly trained local model and the global model of the previous round. To avoid excessive contribution by certain users, a counting mechanism is used to ensure fairness. Simulations with various datasets demonstrate that this method can rapidly achieve convergence similar to that of the centralized user selection approach.

Xianpeng Liu, Ce Zheng, Kelvin Cheng et al.

The main challenge of monocular 3D object detection is the accurate localization of 3D center. Motivated by a new and strong observation that this challenge can be remedied by a 3D-space local-grid search scheme in an ideal case, we propose a stage-wise approach, which combines the information flow from 2D-to-3D (3D bounding box proposal generation with a single 2D image) and 3D-to-2D (proposal verification by denoising with 3D-to-2D contexts) in a top-down manner. Specifically, we first obtain initial proposals from off-the-shelf backbone monocular 3D detectors. Then, we generate a 3D anchor space by local-grid sampling from the initial proposals. Finally, we perform 3D bounding box denoising at the 3D-to-2D proposal verification stage. To effectively learn discriminative features for denoising highly overlapped proposals, this paper presents a method of using the Perceiver I/O model to fuse the 3D-to-2D geometric information and the 2D appearance information. With the encoded latent representation of a proposal, the verification head is implemented with a self-attention module. Our method, named as MonoXiver, is generic and can be easily adapted to any backbone monocular 3D detectors. Experimental results on the well-established KITTI dataset and the challenging large-scale Waymo dataset show that MonoXiver consistently achieves improvement with limited computation overhead.

Wenhan Wu, Yilei Hua, Ce Zheng et al.

Fully supervised skeleton-based action recognition has achieved great progress with the blooming of deep learning techniques. However, these methods require sufficient labeled data which is not easy to obtain. In contrast, self-supervised skeleton-based action recognition has attracted more attention. With utilizing the unlabeled data, more generalizable features can be learned to alleviate the overfitting problem and reduce the demand of massive labeled training data. Inspired by the MAE, we propose a spatial-temporal masked autoencoder framework for self-supervised 3D skeleton-based action recognition (SkeletonMAE). Following MAE's masking and reconstruction pipeline, we utilize a skeleton-based encoder-decoder transformer architecture to reconstruct the masked skeleton sequences. A novel masking strategy, named Spatial-Temporal Masking, is introduced in terms of both joint-level and frame-level for the skeleton sequence. This pre-training strategy makes the encoder output generalizable skeleton features with spatial and temporal dependencies. Given the unmasked skeleton sequence, the encoder is fine-tuned for the action recognition task. Extensive experiments show that our SkeletonMAE achieves remarkable performance and outperforms the state-of-the-art methods on both NTU RGB+D and NTU RGB+D 120 datasets.

Ce Zheng, Xianpeng Liu, Guo-Jun Qi et al.

Transformer architectures have achieved SOTA performance on the human mesh recovery (HMR) from monocular images. However, the performance gain has come at the cost of substantial memory and computational overhead. A lightweight and efficient model to reconstruct accurate human mesh is needed for real-world applications. In this paper, we propose a pure transformer architecture named POoling aTtention TransformER (POTTER) for the HMR task from single images. Observing that the conventional attention module is memory and computationally expensive, we propose an efficient pooling attention module, which significantly reduces the memory and computational cost without sacrificing performance. Furthermore, we design a new transformer architecture by integrating a High-Resolution (HR) stream for the HMR task. The high-resolution local and global features from the HR stream can be utilized for recovering more accurate human mesh. Our POTTER outperforms the SOTA method METRO by only requiring 7% of total parameters and 14% of the Multiply-Accumulate Operations on the Human3.6M (PA-MPJPE metric) and 3DPW (all three metrics) datasets. The project webpage is https://zczcwh.github.io/potter_page.

Ce Zheng, Matias Mendieta, Taojiannan Yang et al.

Recently, vision transformers have shown great success in a set of human reconstruction tasks such as 2D human pose estimation (2D HPE), 3D human pose estimation (3D HPE), and human mesh reconstruction (HMR) tasks. In these tasks, feature map representations of the human structural information are often extracted first from the image by a CNN (such as HRNet), and then further processed by transformer to predict the heatmaps (encodes each joint's location into a feature map with a Gaussian distribution) for HPE or HMR. However, existing transformer architectures are not able to process these feature map inputs directly, forcing an unnatural flattening of the location-sensitive human structural information. Furthermore, much of the performance benefit in recent HPE and HMR methods has come at the cost of ever-increasing computation and memory needs. Therefore, to simultaneously address these problems, we propose FeatER, a novel transformer design that preserves the inherent structure of feature map representations when modeling attention while reducing memory and computational costs. Taking advantage of FeatER, we build an efficient network for a set of human reconstruction tasks including 2D HPE, 3D HPE, and HMR. A feature map reconstruction module is applied to improve the performance of the estimated human pose and mesh. Extensive experiments demonstrate the effectiveness of FeatER on various human pose and mesh datasets. For instance, FeatER outperforms the SOTA method MeshGraphormer by requiring 5% of Params and 16% of MACs on Human3.6M and 3DPW datasets. The project webpage is https://zczcwh.github.io/feater_page/.

Qitao Zhao, Ce Zheng, Mengyuan Liu et al.

The dominant paradigm in 3D human pose estimation that lifts a 2D pose sequence to 3D heavily relies on long-term temporal clues (i.e., using a daunting number of video frames) for improved accuracy, which incurs performance saturation, intractable computation and the non-causal problem. This can be attributed to their inherent inability to perceive spatial context as plain 2D joint coordinates carry no visual cues. To address this issue, we propose a straightforward yet powerful solution: leveraging the readily available intermediate visual representations produced by off-the-shelf (pre-trained) 2D pose detectors -- no finetuning on the 3D task is even needed. The key observation is that, while the pose detector learns to localize 2D joints, such representations (e.g., feature maps) implicitly encode the joint-centric spatial context thanks to the regional operations in backbone networks. We design a simple baseline named Context-Aware PoseFormer to showcase its effectiveness. Without access to any temporal information, the proposed method significantly outperforms its context-agnostic counterpart, PoseFormer, and other state-of-the-art methods using up to hundreds of video frames regarding both speed and precision. Project page: https://qitaozhao.github.io/ContextAware-PoseFormer

Puxun Tu, Hongfei Ye, Haochen Shi et al.

Phacoemulsification cataract surgery (PCS) is a routine procedure conducted using a surgical microscope, heavily reliant on the skill of the ophthalmologist. While existing PCS guidance systems extract valuable information from surgical microscopic videos to enhance intraoperative proficiency, they suffer from non-phasespecific guidance, leading to redundant visual information. In this study, our major contribution is the development of a novel phase-specific augmented reality (AR) guidance system, which offers tailored AR information corresponding to the recognized surgical phase. Leveraging the inherent quasi-standardized nature of PCS procedures, we propose a two-stage surgical microscopic video recognition network. In the first stage, we implement a multi-task learning structure to segment the surgical limbus region and extract limbus region-focused spatial feature for each frame. In the second stage, we propose the long-short spatiotemporal aggregation transformer (LS-SAT) network to model local fine-grained and global temporal relationships, and combine the extracted spatial features to recognize the current surgical phase. Additionally, we collaborate closely with ophthalmologists to design AR visual cues by utilizing techniques such as limbus ellipse fitting and regional restricted normal cross-correlation rotation computation. We evaluated the network on publicly available and in-house datasets, with comparison results demonstrating its superior performance compared to related works. Ablation results further validated the effectiveness of the limbus region-focused spatial feature extractor and the combination of temporal features. Furthermore, the developed system was evaluated in a clinical setup, with results indicating remarkable accuracy and real-time performance. underscoring its potential for clinical applications.

Wenhan Wu, Ce Zheng, Zihao Yang et al.

Recently, transformers have demonstrated great potential for modeling long-term dependencies from skeleton sequences and thereby gained ever-increasing attention in skeleton action recognition. However, the existing transformer-based approaches heavily rely on the naive attention mechanism for capturing the spatiotemporal features, which falls short in learning discriminative representations that exhibit similar motion patterns. To address this challenge, we introduce the Frequency-aware Mixed Transformer (FreqMixFormer), specifically designed for recognizing similar skeletal actions with subtle discriminative motions. First, we introduce a frequency-aware attention module to unweave skeleton frequency representations by embedding joint features into frequency attention maps, aiming to distinguish the discriminative movements based on their frequency coefficients. Subsequently, we develop a mixed transformer architecture to incorporate spatial features with frequency features to model the comprehensive frequency-spatial patterns. Additionally, a temporal transformer is proposed to extract the global correlations across frames. Extensive experiments show that FreqMiXFormer outperforms SOTA on 3 popular skeleton action recognition datasets, including NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets.

Ce Zheng, Xianpeng Liu, Qucheng Peng et al.

Human mesh recovery (HMR) provides rich human body information for various real-world applications. While image-based HMR methods have achieved impressive results, they often struggle to recover humans in dynamic scenarios, leading to temporal inconsistencies and non-smooth 3D motion predictions due to the absence of human motion. In contrast, video-based approaches leverage temporal information to mitigate this issue. In this paper, we present DiffMesh, an innovative motion-aware Diffusion-like framework for video-based HMR. DiffMesh establishes a bridge between diffusion models and human motion, efficiently generating accurate and smooth output mesh sequences by incorporating human motion within the forward process and reverse process in the diffusion model. Extensive experiments are conducted on the widely used datasets (Human3.6M \cite{h36m_pami} and 3DPW \cite{pw3d2018}), which demonstrate the effectiveness and efficiency of our DiffMesh. Visual comparisons in real-world scenarios further highlight DiffMesh's suitability for practical applications.

Qucheng Peng, Ce Zheng, Chen Chen

3D human pose data collected in controlled laboratory settings present challenges for pose estimators that generalize across diverse scenarios. To address this, domain generalization is employed. Current methodologies in domain generalization for 3D human pose estimation typically utilize adversarial training to generate synthetic poses for training. Nonetheless, these approaches exhibit several limitations. First, the lack of prior information about the target domain complicates the application of suitable augmentation through a single pose augmentor, affecting generalization on target domains. Moreover, adversarial training's discriminator tends to enforce similarity between source and synthesized poses, impeding the exploration of out-of-source distributions. Furthermore, the pose estimator's optimization is not exposed to domain shifts, limiting its overall generalization ability. To address these limitations, we propose a novel framework featuring two pose augmentors: the weak and the strong augmentors. Our framework employs differential strategies for generation and discrimination processes, facilitating the preservation of knowledge related to source poses and the exploration of out-of-source distributions without prior information about target poses. Besides, we leverage meta-optimization to simulate domain shifts in the optimization process of the pose estimator, thereby improving its generalization ability. Our proposed approach significantly outperforms existing methods, as demonstrated through comprehensive experiments on various benchmark datasets.Our code will be released at \url{https://github.com/davidpengucf/DAF-DG}.

Shimao Chen, Zirui Liu, Zhiying Wu et al.

As the foundation of large language models (LLMs), self-attention module faces the challenge of quadratic time and memory complexity with respect to sequence length. FlashAttention accelerates attention computation and reduces its memory usage by leveraging the GPU memory hierarchy. A promising research direction is to integrate FlashAttention with quantization methods. This paper introduces INT-FlashAttention, the first INT8 quantization architecture compatible with the forward workflow of FlashAttention, which significantly improves the inference speed of FlashAttention on Ampere GPUs. We implement our INT-FlashAttention prototype with fully INT8 activations and general matrix-multiplication (GEMM) kernels, making it the first attention operator with fully INT8 input. As a general token-level post-training quantization framework, INT-FlashAttention is also compatible with other data formats like INT4, etc. Experimental results show INT-FlashAttention achieves 72% faster inference speed and 82% smaller quantization error compared to standard FlashAttention with FP16 and FP8 data format.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Qucheng Peng, Ce Zheng, Zhengming Ding et al.

Human pose estimation (HPE) has received increasing attention recently due to its wide application in motion analysis, virtual reality, healthcare, etc. However, it suffers from the lack of labeled diverse real-world datasets due to the time- and labor-intensive annotation. To cope with the label deficiency issue, one common solution is to train the HPE models with easily available synthetic datasets (source) and apply them to real-world data (target) through domain adaptation (DA). Unfortunately, prevailing domain adaptation techniques within the HPE domain remain predominantly fixated on effecting alignment and aggregation between source and target features, often sidestepping the crucial task of excluding domain-specific representations. To rectify this, we introduce a novel framework that capitalizes on both representation aggregation and segregation for domain adaptive human pose estimation. Within this framework, we address the network architecture aspect by disentangling representations into distinct domain-invariant and domain-specific components, facilitating aggregation of domain-invariant features while simultaneously segregating domain-specific ones. Moreover, we tackle the discrepancy measurement facet by delving into various keypoint relationships and applying separate aggregation or segregation mechanisms to enhance alignment. Extensive experiments on various benchmarks, e.g., Human3.6M, LSP, H3D, and FreiHand, show that our method consistently achieves state-of-the-art performance. The project is available at \url{https://github.com/davidpengucf/EPIC}.

Ke Zhang, Xiaoning Zhao, Ce Zheng et al.

This study proposes Tool-RoCo, a novel benchmark for evaluating large language models (LLMs) in long-term multi-agent cooperation based on RoCo, a multi-robot cooperative benchmark. Recent research on LLM-based multi-agent systems has relied on predefined orchestration, while ignoring agent autonomy. Tool-RoCo treats other agents as tools and introduces cooperative tools, leveraging tool usage to evaluate multi-agent cooperation and self-organization. Tool usage means that each agent (LLM) selects a tool from a candidate set based on the current state, receives feedback, and adjusts its selection in subsequent rounds. To evaluate different autonomy levels, we propose four LLM paradigms: (1) centralized cooperation, where a single LLM allocates tools to all agents; (2) centralized self-organization, where a central LLM autonomously activates agents while keeping others inactive; (3) decentralized cooperation, where each agent has its own LLM and calls tools based on local information; and (4) self-organization, where a randomly chosen initial agent can request collaboration, activating additional agents via tool calls. Tool-RoCo includes three multi-robot tasks, SORT, PACK, and CABINET, to measure format and parameter accuracy and agent coordination through tool usage. The results using several LLMs showed that cooperative tools accounted for only 7.09% of all tools, indicating that LLM-based agents rarely invoked others as assistants. Moreover, activation tools accounted for 96.42%, suggesting that current LLMs tend to maintain active agents while seldom deactivating them for adaptive coordination. Tool-RoCo provides a systematic benchmark to evaluate LLM autonomy and cooperation in multi-agent tasks. Code and Demo: https://github.com/ColaZhang22/Tool-Roco

Ke Zhang, DanDan Zhu, Qiuhan Xu et al.

Training for multi-agent reinforcement learning(MARL) is a time-consuming process caused by distribution shift of each agent. One drawback is that strategy of each agent in MARL is independent but actually in cooperation. Thus, a vertical issue in multi-agent reinforcement learning is how to efficiently accelerate training process. To address this problem, current research has leveraged a centralized function(CF) across multiple agents to learn contribution of the team reward for each agent. However, CF based methods introduce joint error from other agents in estimation of value network. In so doing, inspired by federated learning, we propose three simple novel approaches called Average Periodically Parameter Sharing(A-PPS), Reward-Scalability Periodically Parameter Sharing(RS-PPS) and Partial Personalized Periodically Parameter Sharing(PP-PPS) mechanism to accelerate training of MARL. Agents share Q-value network periodically during the training process. Agents which has same identity adapt collected reward as scalability and update partial neural network during period to share different parameters. We apply our approaches in classical MARL method QMIX and evaluate our approaches on various tasks in StarCraft Multi-Agent Challenge(SMAC) environment. Performance of numerical experiments yield enormous enhancement, with an average improvement of 10\%-30\%, and enable to win tasks that QMIX cannot. Our code can be downloaded from https://github.com/ColaZhang22/PPS-QMIX

Lei Li, Jingjing Xu, Qingxiu Dong et al.

Language models~(LMs) gradually become general-purpose interfaces in the interactive and embodied world, where the understanding of physical concepts is an essential prerequisite. However, it is not yet clear whether LMs can understand physical concepts in the human world. To investigate this, we design a benchmark VEC that covers the tasks of (i) Visual concepts, such as the shape and material of objects, and (ii) Embodied Concepts, learned from the interaction with the world such as the temperature of objects. Our zero (few)-shot prompting results show that the understanding of certain visual concepts emerges as scaling up LMs, but there are still basic concepts to which the scaling law does not apply. For example, OPT-175B performs close to humans with a zero-shot accuracy of 85\% on the material concept, yet behaves like random guessing on the mass concept. Instead, vision-augmented LMs such as CLIP and BLIP achieve a human-level understanding of embodied concepts. Analysis indicates that the rich semantics in visual representation can serve as a valuable source of embodied knowledge. Inspired by this, we propose a distillation method to transfer embodied knowledge from VLMs to LMs, achieving performance gain comparable with that by scaling up the parameters of LMs 134x. Our dataset is available at \url{https://github.com/TobiasLee/VEC}

Ce Zheng, Lei Li, Qingxiu Dong et al.

Previous studies have shown that large language models (LLMs) like GPTs store massive factual knowledge in their parameters. However, the stored knowledge could be false or out-dated. Traditional knowledge editing methods refine LLMs via fine-tuning on texts containing specific knowledge. However, with the increasing scales of LLMs, these gradient-based approaches bring large computation costs. The trend of model-as-a-service also makes it impossible to modify knowledge in black-box LMs. Inspired by in-context learning (ICL), a new paradigm based on demonstration contexts without parameter updating, we explore whether ICL can edit factual knowledge. To answer this question, we give a comprehensive empirical study of ICL strategies. Experiments show that in-context knowledge editing (IKE), without any gradient and parameter updating, achieves a competitive success rate compared to gradient-based methods on GPT-J (6B) but with much fewer side effects, including less over-editing on similar but unrelated facts and less knowledge forgetting on previously stored knowledge. We also apply the method to larger LMs with tens or hundreds of parameters like OPT-175B, which shows the scalability of our method. The code is available at https://github.com/Zce1112zslx/IKE.

Haoran Zhang, Jianlong Yang, Ce Zheng et al.

Deep learning has been successfully applied to OCT segmentation. However, for data from different manufacturers and imaging protocols, and for different regions of interest (ROIs), it requires laborious and time-consuming data annotation and training, which is undesirable in many scenarios, such as surgical navigation and multi-center clinical trials. Here we propose an annotation-efficient learning method for OCT segmentation that could significantly reduce annotation costs. Leveraging self-supervised generative learning, we train a Transformer-based model to learn the OCT imagery. Then we connect the trained Transformer-based encoder to a CNN-based decoder, to learn the dense pixel-wise prediction in OCT segmentation. These training phases use open-access data and thus incur no annotation costs, and the pre-trained model can be adapted to different data and ROIs without re-training. Based on the greedy approximation for the k-center problem, we also introduce an algorithm for the selective annotation of the target data. We verified our method on publicly-available and private OCT datasets. Compared to the widely-used U-Net model with 100% training data, our method only requires ~10% of the data for achieving the same segmentation accuracy, and it speeds the training up to ~3.5 times. Furthermore, our proposed method outperforms other potential strategies that could improve annotation efficiency. We think this emphasis on learning efficiency may help improve the intelligence and application penetration of OCT-based technologies. Our code and pre-trained model are publicly available at https://github.com/SJTU-Intelligent-Optics-Lab/Annotation-efficient-learning-for-OCT-segmentation.

Ce Zheng, Sijie Zhu, Matias Mendieta et al.

Transformer architectures have become the model of choice in natural language processing and are now being introduced into computer vision tasks such as image classification, object detection, and semantic segmentation. However, in the field of human pose estimation, convolutional architectures still remain dominant. In this work, we present PoseFormer, a purely transformer-based approach for 3D human pose estimation in videos without convolutional architectures involved. Inspired by recent developments in vision transformers, we design a spatial-temporal transformer structure to comprehensively model the human joint relations within each frame as well as the temporal correlations across frames, then output an accurate 3D human pose of the center frame. We quantitatively and qualitatively evaluate our method on two popular and standard benchmark datasets: Human3.6M and MPI-INF-3DHP. Extensive experiments show that PoseFormer achieves state-of-the-art performance on both datasets. Code is available at \url{https://github.com/zczcwh/PoseFormer}

Ce Zheng, Wenhan Wu, Chen Chen et al.

Human pose estimation aims to locate the human body parts and build human body representation (e.g., body skeleton) from input data such as images and videos. It has drawn increasing attention during the past decade and has been utilized in a wide range of applications including human-computer interaction, motion analysis, augmented reality, and virtual reality. Although the recently developed deep learning-based solutions have achieved high performance in human pose estimation, there still remain challenges due to insufficient training data, depth ambiguities, and occlusion. The goal of this survey paper is to provide a comprehensive review of recent deep learning-based solutions for both 2D and 3D pose estimation via a systematic analysis and comparison of these solutions based on their input data and inference procedures. More than 250 research papers since 2014 are covered in this survey. Furthermore, 2D and 3D human pose estimation datasets and evaluation metrics are included. Quantitative performance comparisons of the reviewed methods on popular datasets are summarized and discussed. Finally, the challenges involved, applications, and future research directions are concluded. A regularly updated project page is provided: \url{https://github.com/zczcwh/DL-HPE}

Jiewei Lu, Zhun Fan, Ce Zheng et al.

Strabismus is one of the most influential ophthalmologic diseases in human's life. Timely detection of strabismus contributes to its prognosis and treatment. Telemedicine, which has great potential to alleviate the growing demand of the diagnosis of ophthalmologic diseases, is an effective method to achieve timely strabismus detection. In this paper, a tele strabismus dataset is established by the ophthalmologists. Then an end-to-end framework named as RF-CNN is proposed to achieve automated strabismus detection on the established tele strabismus dataset. RF-CNN first performs eye region segmentation on each individual image, and further classifies the segmented eye regions with deep neural networks. The experimental results on the established tele strabismus dataset demonstrates that the proposed RF-CNN can have a good performance on automated strabismus detection for telemedicine application. Code is made publicly available at: https://github.com/jieWeiLu/Strabismus-Detection-for-Telemedicine-Application.

Yiwen Zhao, Ce Zheng, Yufu Wang et al.

Human mesh recovery (HMR) models 3D human body from monocular videos, with recent works extending it to world-coordinate human trajectory and motion reconstruction. However, most existing methods remain offline, relying on future frames or global optimization, which limits their applicability in interactive feedback and perception-action loop scenarios such as AR/VR and telepresence. To address this, we propose OnlineHMR, a fully online framework that jointly satisfies four essential criteria of online processing, including system-level causality, faithfulness, temporal consistency, and efficiency. Built upon a two-branch architecture, OnlineHMR enables streaming inference via a causal key-value cache design and a curated sliding-window learning strategy. Meanwhile, a human-centric incremental SLAM provides online world-grounded alignment under physically plausible trajectory correction. Experimental results show that our method achieves performance comparable to existing chunk-based approaches on the standard EMDB benchmark and highly dynamic custom videos, while uniquely supporting online processing. Page and code are available at https://tsukasane.github.io/Video-OnlineHMR/.

Ce Zheng, Xinghan Wang, Jiahong Ning et al.

Federated inference enhances LLM performance in edge computing through weighted averaging of distributed model predictions. However, autoregressive LLM inference requires frequent full-model forward passes across workers, severely limiting decoding throughput. Distributed deployment further aggravates this due to a communication bottleneck: each worker must transmit full token probability distributions per draft token, dominating end-to-end latency. To address these challenges, we introduce speculative decoding to enable parallel LLM processing and propose a top-K compressed transmission scheme with two server-side reconstruction strategies. We theoretically analyze the robustness of our method in terms of local reconstruction error, aggregation bias, and acceptance-rate bias, and derive corresponding bounds. Experiments demonstrate that our scheme achieves high generation fidelity while significantly reducing communication overhead.

Xianpeng Liu, Ce Zheng, Ming Qian et al.

We present Multi-View Attentive Contextualization (MvACon), a simple yet effective method for improving 2D-to-3D feature lifting in query-based multi-view 3D (MV3D) object detection. Despite remarkable progress witnessed in the field of query-based MV3D object detection, prior art often suffers from either the lack of exploiting high-resolution 2D features in dense attention-based lifting, due to high computational costs, or from insufficiently dense grounding of 3D queries to multi-scale 2D features in sparse attention-based lifting. Our proposed MvACon hits the two birds with one stone using a representationally dense yet computationally sparse attentive feature contextualization scheme that is agnostic to specific 2D-to-3D feature lifting approaches. In experiments, the proposed MvACon is thoroughly tested on the nuScenes benchmark, using both the BEVFormer and its recent 3D deformable attention (DFA3D) variant, as well as the PETR, showing consistent detection performance improvement, especially in enhancing performance in location, orientation, and velocity prediction. It is also tested on the Waymo-mini benchmark using BEVFormer with similar improvement. We qualitatively and quantitatively show that global cluster-based contexts effectively encode dense scene-level contexts for MV3D object detection. The promising results of our proposed MvACon reinforces the adage in computer vision -- ``(contextualized) feature matters".

Sen Fang, Chen Chen, Lei Wang et al.

In this paper, we propose SignLLM, a multilingual Sign Language Production (SLP) large language model, which includes two novel multilingual SLP modes MLSF and Prompt2LangGloss that allow sign language gestures generation from query texts input and question-style prompts input respectively. Both modes can use a new RL loss based on reinforcement learning and a new RL module named Priority Learning Channel. These RL components can accelerate the training by enhancing the model's capability to sample high-quality data. To train SignLLM, we introduce Prompt2Sign, a comprehensive multilingual sign language dataset, which builds from public data, including American Sign Language (ASL) and seven others. This dataset standardizes information by extracting pose information from sign language videos into a unified compressed format. We extensively evaluate SignLLM, demonstrating that our model achieves state-of-the-art performance on SLP tasks across eight sign languages.

Liuyue Xie, George Z. Wei, Avik Kuthiala et al. · cmu

Frontier models have either been language-only or have primarily focused on vision and language modalities. Although recent advancements in models with vision and audio understanding capabilities have shown substantial progress, the field lacks a standardized evaluation framework for thoroughly assessing their cross-modality perception performance. We introduce MAVERIX~(Multimodal Audio-Visual Evaluation Reasoning IndeX), a novel benchmark with 700 videos and 2,556 questions explicitly designed to evaluate multimodal models through tasks that necessitate close integration of video and audio information. MAVERIX uniquely provides models with audiovisual tasks, closely mimicking the multimodal perceptual experiences available to humans during inference and decision-making processes. To our knowledge, MAVERIX is the first benchmark aimed explicitly at assessing comprehensive audiovisual integration. Experiments with state-of-the-art models, including Gemini 1.5 Pro and o1, show performance approaching human levels (around 70% accuracy), while human experts reach near-ceiling performance (95.1%). With standardized evaluation protocols, a rigorously annotated pipeline, and a public toolkit, MAVERIX establishes a challenging testbed for advancing audiovisual multimodal intelligence.

Aiyinsi Zuo, Haixi Zhang, Zirui Li et al.

Within the field of complicated multivariate time series forecasting (TSF), popular techniques frequently rely on intricate deep learning architectures, ranging from transformer-based designs to recurrent neural networks. However, recent findings suggest that simple Linear models can surpass sophisticated constructs on diverse datasets. These models directly map observation to multiple future time steps, thereby minimizing error accumulation in iterative multi-step prediction. Yet, these models fail to incorporate spatial and temporal information within the data, which is critical for capturing patterns and dependencies that drive insightful predictions. This oversight often leads to performance bottlenecks, especially under specific sequence lengths and dataset conditions, preventing their universal application. In response, we introduce the SpatioTemporal-Linear (STL) framework. STL seamlessly integrates time-embedded and spatially-informed bypasses to augment the Linear-based architecture. These extra routes offer a more robust and refined regression to the data, particularly when the amount of observation is limited and the capacity of simple linear layers to capture dependencies declines. Empirical evidence highlights STL's prowess, outpacing both Linear and Transformer benchmarks across varied observation and prediction durations and datasets. Such robustness accentuates its suitability across a spectrum of applications, including but not limited to, traffic trajectory and rare disease progression forecasting. Through this discourse, we not only validate the STL's distinctive capacities to become a more general paradigm in multivariate time-series prediction using deep-learning techniques but also stress the need to tackle data-scarce prediction scenarios for universal application. Code will be made available.

Jungbin Cho, Minsu Kim, Jisoo Kim et al.

Human motion is inherently diverse and semantically rich, while also shaped by the surrounding scene. However, existing motion generation approaches address either motion semantics or scene-awareness in isolation, since constructing large-scale datasets with both rich text--motion coverage and precise scene interactions is extremely challenging. In this work, we introduce SceneAdapt, a framework that injects scene awareness into text-conditioned motion models by leveraging disjoint scene--motion and text--motion datasets through two adaptation stages: inbetweening and scene-aware inbetweening. The key idea is to use motion inbetweening, learnable without text, as a proxy task to bridge two distinct datasets and thereby inject scene-awareness to text-to-motion models. In the first stage, we introduce keyframing layers that modulate motion latents for inbetweening while preserving the latent manifold. In the second stage, we add a scene-conditioning layer that injects scene geometry by adaptively querying local context through cross-attention. Experimental results show that SceneAdapt effectively injects scene awareness into text-to-motion models, and we further analyze the mechanisms through which this awareness emerges. Code and models will be released.

Jing Gao, Ce Zheng, Laszlo A. Jeni et al.

In-bed human mesh recovery can be crucial and enabling for several healthcare applications, including sleep pattern monitoring, rehabilitation support, and pressure ulcer prevention. However, it is difficult to collect large real-world visual datasets in this domain, in part due to privacy and expense constraints, which in turn presents significant challenges for training and deploying deep learning models. Existing in-bed human mesh estimation methods often rely heavily on real-world data, limiting their ability to generalize across different in-bed scenarios, such as varying coverings and environmental settings. To address this, we propose a Sim-to-Real Transfer Framework for in-bed human mesh recovery from overhead depth images, which leverages large-scale synthetic data alongside limited or no real-world samples. We introduce a diffusion model that bridges the gap between synthetic data and real data to support generalization in real-world in-bed pose and body inference scenarios. Extensive experiments and ablation studies validate the effectiveness of our framework, demonstrating significant improvements in robustness and adaptability across diverse healthcare scenarios.

Bofei Gao, Zefan Cai, Runxin Xu et al.

In recent progress, mathematical verifiers have achieved success in mathematical reasoning tasks by validating the correctness of solutions generated by policy models. However, existing verifiers are trained with binary classification labels, which are not informative enough for the model to accurately assess the solutions. To mitigate the aforementioned insufficiency of binary labels, we introduce step-wise natural language feedback as rationale labels, that is, the correctness of each step and the detailed explanations. In this paper, we propose Math-Minos, a natural language feedback-enhanced verifier by constructing automatically generated training data and a two-stage training paradigm for effective training and efficient inference. Our experiments reveal that a small set of natural language feedback can significantly boost the performance of the verifier in both verification and reinforcement learning. We have released the code and data for further exploration.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Rohan Anil, Andrew M. Dai, Orhan Firat et al.

We introduce PaLM 2, a new state-of-the-art language model that has better multilingual and reasoning capabilities and is more compute-efficient than its predecessor PaLM. PaLM 2 is a Transformer-based model trained using a mixture of objectives. Through extensive evaluations on English and multilingual language, and reasoning tasks, we demonstrate that PaLM 2 has significantly improved quality on downstream tasks across different model sizes, while simultaneously exhibiting faster and more efficient inference compared to PaLM. This improved efficiency enables broader deployment while also allowing the model to respond faster, for a more natural pace of interaction. PaLM 2 demonstrates robust reasoning capabilities exemplified by large improvements over PaLM on BIG-Bench and other reasoning tasks. PaLM 2 exhibits stable performance on a suite of responsible AI evaluations, and enables inference-time control over toxicity without additional overhead or impact on other capabilities. Overall, PaLM 2 achieves state-of-the-art performance across a diverse set of tasks and capabilities. When discussing the PaLM 2 family, it is important to distinguish between pre-trained models (of various sizes), fine-tuned variants of these models, and the user-facing products that use these models. In particular, user-facing products typically include additional pre- and post-processing steps. Additionally, the underlying models may evolve over time. Therefore, one should not expect the performance of user-facing products to exactly match the results reported in this report.

Yilei Hua, Wenhan Wu, Ce Zheng et al.

In recent years, remarkable results have been achieved in self-supervised action recognition using skeleton sequences with contrastive learning. It has been observed that the semantic distinction of human action features is often represented by local body parts, such as legs or hands, which are advantageous for skeleton-based action recognition. This paper proposes an attention-based contrastive learning framework for skeleton representation learning, called SkeAttnCLR, which integrates local similarity and global features for skeleton-based action representations. To achieve this, a multi-head attention mask module is employed to learn the soft attention mask features from the skeletons, suppressing non-salient local features while accentuating local salient features, thereby bringing similar local features closer in the feature space. Additionally, ample contrastive pairs are generated by expanding contrastive pairs based on salient and non-salient features with global features, which guide the network to learn the semantic representations of the entire skeleton. Therefore, with the attention mask mechanism, SkeAttnCLR learns local features under different data augmentation views. The experiment results demonstrate that the inclusion of local feature similarity significantly enhances skeleton-based action representation. Our proposed SkeAttnCLR outperforms state-of-the-art methods on NTURGB+D, NTU120-RGB+D, and PKU-MMD datasets.

Ce Zheng, Matias Mendieta, Pu Wang et al.

Existing deep learning-based human mesh reconstruction approaches have a tendency to build larger networks in order to achieve higher accuracy. Computational complexity and model size are often neglected, despite being key characteristics for practical use of human mesh reconstruction models (e.g. virtual try-on systems). In this paper, we present GTRS, a lightweight pose-based method that can reconstruct human mesh from 2D human pose. We propose a pose analysis module that uses graph transformers to exploit structured and implicit joint correlations, and a mesh regression module that combines the extracted pose feature with the mesh template to reconstruct the final human mesh. We demonstrate the efficiency and generalization of GTRS by extensive evaluations on the Human3.6M and 3DPW datasets. In particular, GTRS achieves better accuracy than the SOTA pose-based method Pose2Mesh while only using 10.2% of the parameters (Params) and 2.5% of the FLOPs on the challenging in-the-wild 3DPW dataset. Code will be publicly available.

Ce Zheng, Yecheng Lyu, Ming Li et al.

Deep learning based LiDAR odometry (LO) estimation attracts increasing research interests in the field of autonomous driving and robotics. Existing works feed consecutive LiDAR frames into neural networks as point clouds and match pairs in the learned feature space. In contrast, motivated by the success of image based feature extractors, we propose to transfer the LiDAR frames to image space and reformulate the problem as image feature extraction. With the help of scale-invariant feature transform (SIFT) for feature extraction, we are able to generate matched keypoint pairs (MKPs) that can be precisely returned to the 3D space. A convolutional neural network pipeline is designed for LiDAR odometry estimation by extracted MKPs. The proposed scheme, namely LodoNet, is then evaluated in the KITTI odometry estimation benchmark, achieving on par with or even better results than the state-of-the-art.