Jiarui Zhang

Tong Wu, Jiarui Zhang, Xiao Fu et al.

Recent advances in modeling 3D objects mostly rely on synthetic datasets due to the lack of large-scale realscanned 3D databases. To facilitate the development of 3D perception, reconstruction, and generation in the real world, we propose OmniObject3D, a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects. OmniObject3D has several appealing properties: 1) Large Vocabulary: It comprises 6,000 scanned objects in 190 daily categories, sharing common classes with popular 2D datasets (e.g., ImageNet and LVIS), benefiting the pursuit of generalizable 3D representations. 2) Rich Annotations: Each 3D object is captured with both 2D and 3D sensors, providing textured meshes, point clouds, multiview rendered images, and multiple real-captured videos. 3) Realistic Scans: The professional scanners support highquality object scans with precise shapes and realistic appearances. With the vast exploration space offered by OmniObject3D, we carefully set up four evaluation tracks: a) robust 3D perception, b) novel-view synthesis, c) neural surface reconstruction, and d) 3D object generation. Extensive studies are performed on these four benchmarks, revealing new observations, challenges, and opportunities for future research in realistic 3D vision.

Liyuan Zhang, Jiarui Zhang, Jinwei Yao et al.

Speculative decoding accelerates autoregressive large language model inference by drafting multiple tokens and verifying them in a single target-model forward pass. Recent diffusion-based drafters generate an entire block of tokens in parallel but usually commit to a single draft sequence per verification: once the first mismatch occurs, all subsequent draft tokens are discarded, resulting in a limited acceptance rate. Naively batching more draft candidate sequences only introduces a marginal improvement, as redundant or poorly placed branches increase the cost of drafting and verification without proportionally increasing the number of accepted tokens. We propose D^2SD, a dual diffusion draft speculative decoding framework that organizes candidates into a confidence-guided prefix tree, where the first diffusion drafter generates a block along with per-position confidence scores that are used to identify the most likely rejection boundary and select the top-K prefix ranges for recovery; the second variable-prefix diffusion drafter re-anchors at each selected prefix and proposes alternative continuations in one batched pass; the resulting shared-prefix candidates are jointly verified via cascade attention. Empirically, D^2SD shows clear improvements over both the underlying diffusion approach and strong autoregressive speculative decoding baselines.

Jiarui Zhang, Filip Ilievski, Kaixin Ma et al. · cmu

Intelligent Traffic Monitoring (ITMo) technologies hold the potential for improving road safety/security and for enabling smart city infrastructure. Understanding traffic situations requires a complex fusion of perceptual information with domain-specific and causal commonsense knowledge. Whereas prior work has provided benchmarks and methods for traffic monitoring, it remains unclear whether models can effectively align these information sources and reason in novel scenarios. To address this assessment gap, we devise three novel text-based tasks for situational reasoning in the traffic domain: i) BDD-QA, which evaluates the ability of Language Models (LMs) to perform situational decision-making, ii) TV-QA, which assesses LMs' abilities to reason about complex event causality, and iii) HDT-QA, which evaluates the ability of models to solve human driving exams. We adopt four knowledge-enhanced methods that have shown generalization capability across language reasoning tasks in prior work, based on natural language inference, commonsense knowledge-graph self-supervision, multi-QA joint training, and dense retrieval of domain information. We associate each method with a relevant knowledge source, including knowledge graphs, relevant benchmarks, and driving manuals. In extensive experiments, we benchmark various knowledge-aware methods against the three datasets, under zero-shot evaluation; we provide in-depth analyses of model performance on data partitions and examine model predictions categorically, to yield useful insights on traffic understanding, given different background knowledge and reasoning strategies.

Jiarui Zhang, Filip Ilievski, Kaixin Ma et al. · cmu

Self-supervision based on the information extracted from large knowledge graphs has been shown to improve the generalization of language models, in zero-shot evaluation on various downstream language reasoning tasks. Since these improvements are reported in aggregate, however, little is known about (i) how to select the appropriate knowledge for solid performance across tasks, (ii) how to combine this knowledge with neural language models, and (iii) how these pairings affect granular task performance. In this paper, we study the effect of knowledge sampling strategies and sizes that can be used to generate synthetic data for adapting language models. We study the effect of different synthetic datasets on language models with various architectures and sizes. The resulting models are evaluated against four task properties: domain overlap, answer similarity, vocabulary overlap, and answer length. Our experiments show that encoder-decoder models benefit from more data to learn from, whereas sampling strategies that balance across different aspects yield best performance. Most of the improvement occurs on questions with short answers and dissimilar answer candidates, which corresponds to the characteristics of the data used for pre-training.

Jiarui Zhang, Filip Ilievski, Aravinda Kollaa et al. · cmu

Understanding novel situations in the traffic domain requires an intricate combination of domain-specific and causal commonsense knowledge. Prior work has provided sufficient perception-based modalities for traffic monitoring, in this paper, we focus on a complementary research aspect of Intelligent Transportation: traffic understanding. We scope our study to text-based methods and datasets given the abundant commonsense knowledge that can be extracted using language models from large corpus and knowledge graphs. We adopt three knowledge-driven approaches for zero-shot QA over traffic situations, based on prior natural language inference methods, commonsense models with knowledge graph self-supervision, and dense retriever-based models. We constructed two text-based multiple-choice question answering sets: BDD-QA for evaluating causal reasoning in the traffic domain and HDT-QA for measuring the possession of domain knowledge akin to human driving license tests. Among the methods, Unified-QA reaches the best performance on the BDD-QA dataset with the adaptation of multiple formats of question answers. Language models trained with inference information and commonsense knowledge are also good at predicting the cause and effect in the traffic domain but perform badly at answering human-driving QA sets. For such sets, DPR+Unified-QA performs the best due to its efficient knowledge extraction.

Yuhang Zhang, Jiarui Zhang, Bowen Jian et al.

The rapid iteration of autonomous driving algorithms has created a growing demand for high-fidelity, replayable, and diagnosable testing data. However, many public datasets lack real vehicle dynamics feedback and closed-loop interaction with surrounding traffic and road infrastructure, limiting their ability to reflect deployment readiness. To address this gap, we present OVPD (OnSite Virtual-Physical Dataset), a virtual-physical fusion testing dataset released from the 2025 OnSite Autonomous Driving Challenge. Centered on real-vehicle-in-the-loop testing, OVPD integrates virtual background traffic with vehicle-infrastructure perception to build controllable and interactive closed-loop test environments on a proving ground. The dataset contains 20 testing clips from 20 teams over a scenario chain of 15 atomic scenarios, totaling nearly 3 hours of multi-modal data, including vehicle trajectories and states, control commands, and digital-twin-rendered surround-view observations. OVPD supports long-tail planning and decision-making validation, open-loop or platform-enabled closed-loop evaluation, and comprehensive assessment across safety, efficiency, comfort, rule compliance, and traffic impact, providing actionable evidence for failure diagnosis and iterative improvement. The dataset is available via: https://huggingface.co/datasets/Yuhang253820/Onsite_OPVD

Jiarui Zhang, Xiangyu Liu, Yong Hu et al.

Multi-turn dialogue is the predominant form of interaction with large language models (LLMs). While LLM routing is effective in single-turn settings, existing methods fail to maximize cumulative performance in multi-turn dialogue due to interaction dynamics and delayed rewards. To address this challenge, we move from myopic, single-turn selection to long-horizon sequential routing for multi-turn dialogue. Accordingly, we propose DialRouter, which first performs MCTS to explore dialogue branches induced by different LLM selections and collect trajectories with high cumulative rewards. DialRouter then learns a lightweight routing policy from search-derived data, augmented with retrieval-based future state approximation, enabling multi-turn routing without online search. Experiments on both open-domain and domain-specific dialogue tasks across diverse candidate sets of both open-source and closed-source LLMs demonstrate that DialRouter significantly outperforms single LLMs and existing routing baselines in task success rate, while achieving a superior performance-cost trade-off when combined with a cost-aware reward.

Zi Ye, Yibin Wen, Xiaoya Fan et al.

Agricultural decision-making increasingly requires multimodal systems that can transform visual observations into reliable, executable actions. However, existing agricultural multimodal benchmarks mainly evaluate final-answer correctness and provide limited support for assessing whether models can use external tools to complete precision-sensitive workflows. In this paper, we introduce AgroTools, a benchmark for evaluating tool-augmented multimodal agents in agriculture. AgroTools contains 539 question-answer instances paired with 1,097 heterogeneous agricultural images, spanning five task families and an executable environment of 14 agricultural tools. Each query is annotated with structured tool-use traces, enabling a dual-view evaluation of both process-level execution quality and outcome-level task success. We benchmark 9 open-source and 4 closed-source multimodal large language models on AgroTools. Results show that current models remain far from reliable in agricultural tool-use settings, with clear bottlenecks in tool planning, argument generation, execution recovery, and final-answer synthesis. We hope AgroTools will support future research on multimodal agents for high-precision agricultural applications. The benchmark and evaluation are available at https://huggingface.co/datasets/AgroTools/AgroTools.

Jiarui Zhang, Yuliang Liu, Zijun Wu et al.

Document parsing is a core task in document intelligence, supporting applications such as information extraction, retrieval-augmented generation, and automated document analysis. However, real-world documents often feature complex layouts with multi-level tables, embedded images or formulas, and cross-page structures, which remain challenging for existing OCR systems. We introduce MonkeyOCR v1.5, a unified vision-language framework that enhances both layout understanding and content recognition through a two-stage pipeline. The first stage employs a large multimodal model to jointly predict layout and reading order, leveraging visual information to ensure sequential consistency. The second stage performs localized recognition of text, formulas, and tables within detected regions, maintaining high visual fidelity while reducing error propagation. To address complex table structures, we propose a visual consistency-based reinforcement learning scheme that evaluates recognition quality via render-and-compare alignment, improving structural accuracy without manual annotations. Additionally, two specialized modules, Image-Decoupled Table Parsing and Type-Guided Table Merging, are introduced to enable reliable parsing of tables containing embedded images and reconstruction of tables crossing pages or columns. Comprehensive experiments on OmniDocBench v1.5 demonstrate that MonkeyOCR v1.5 achieves state-of-the-art performance, outperforming PPOCR-VL and MinerU 2.5 while showing exceptional robustness in visually complex document scenarios. A trial link can be found at https://github.com/Yuliang-Liu/MonkeyOCR .

Yingjie Chen, Jiarui Zhang, Tao Wang et al.

Facial action units (AUs) play an indispensable role in human emotion analysis. We observe that although AU-based high-level emotion analysis is urgently needed by real-world applications, frame-level AU results provided by previous works cannot be directly used for such analysis. Moreover, as AUs are dynamic processes, the utilization of global temporal information is important but has been gravely ignored in the literature. To this end, we propose EventFormer for AU event detection, which is the first work directly detecting AU events from a video sequence by viewing AU event detection as a multiple class-specific sets prediction problem. Extensive experiments conducted on a commonly used AU benchmark dataset, BP4D, show the superiority of EventFormer under suitable metrics.

Zhang Li, Zhibo Lin, Qiang Liu et al.

We introduce Multilingual Document Parsing Benchmark, the first benchmark for multilingual digital and photographed document parsing. Document parsing has made remarkable strides, yet almost exclusively on clean, digital, well-formatted pages in a handful of dominant languages. No systematic benchmark exists to evaluate how models perform on digital and photographed documents across diverse scripts and low-resource languages. MDPBench comprises 3,400 document images spanning 17 languages, diverse scripts, and varied photographic conditions, with high-quality annotations produced through a rigorous pipeline of expert model labeling, manual correction, and human verification. To ensure fair comparison and prevent data leakage, we maintain separate public and private evaluation splits. Our comprehensive evaluation of both open-source and closed-source models uncovers a striking finding: while closed-source models (notably Gemini3-Pro) prove relatively robust, open-source alternatives suffer dramatic performance collapse, particularly on non-Latin scripts and real-world photographed documents, with an average drop of 17.8% on photographed documents and 14.0% on non-Latin scripts. These results reveal significant performance imbalances across languages and conditions, and point to concrete directions for building more inclusive, deployment-ready parsing systems. Source available at https://github.com/Yuliang-Liu/MultimodalOCR.

Zongxia Li, Hongyang Du, Chengsong Huang et al.

Self-evolving has emerged as a key paradigm for improving foundational models such as Large Language Models (LLMs) and Vision Language Models (VLMs) with minimal human intervention. While recent approaches have demonstrated that LLM agents can self-evolve from scratch with little to no data, VLMs introduce an additional visual modality that typically requires at least some seed data, such as images, to bootstrap the self-evolution process. In this work, we present Multi-model Multimodal Zero (MM-Zero), the first RL-based framework to achieve zero-data self-evolution for VLM reasoning. Moving beyond prior dual-role (Proposer and Solver) setups, MM-Zero introduces a multi-role self-evolving training framework comprising three specialized roles: a Proposer that generates abstract visual concepts and formulates questions; a Coder that translates these concepts into executable code (e.g., Python, SVG) to render visual images; and a Solver that performs multimodal reasoning over the generated visual content. All three roles are initialized from the same base model and trained using Group Relative Policy Optimization (GRPO), with carefully designed reward mechanisms that integrate execution feedback, visual verification, and difficulty balancing. Our experiments show that MM-Zero improves VLM reasoning performance across a wide range of multimodal benchmarks. MM-Zero establishes a scalable path toward self-evolving multi-model systems for multimodal models, extending the frontier of self-improvement beyond the conventional two-model paradigm.

Baixin Xu, Jiarui Zhang, Kwan-Yee Lin et al.

Reconstructing 3D human heads in low-view settings presents technical challenges, mainly due to the pronounced risk of overfitting with limited views and high-frequency signals. To address this, we propose geometry decomposition and adopt a two-stage, coarse-to-fine training strategy, allowing for progressively capturing high-frequency geometric details. We represent 3D human heads using the zero level-set of a combined signed distance field, comprising a smooth template, a non-rigid deformation, and a high-frequency displacement field. The template captures features that are independent of both identity and expression and is co-trained with the deformation network across multiple individuals with sparse and randomly selected views. The displacement field, capturing individual-specific details, undergoes separate training for each person. Our network training does not require 3D supervision or object masks. Experimental results demonstrate the effectiveness and robustness of our geometry decomposition and two-stage training strategy. Our method outperforms existing neural rendering approaches in terms of reconstruction accuracy and novel view synthesis under low-view settings. Moreover, the pre-trained template serves a good initialization for our model when encountering unseen individuals.

Hang Zeng, Xiangyu Liu, Yong Hu et al.

Large language models (LLMs) offer strong capabilities but raise cost and privacy concerns, whereas small language models (SLMs) facilitate efficient and private local inference yet suffer from limited capacity. To synergize the complementary strengths, we introduce a dynamic collaboration framework, where an SLM learns to proactively decide how to request an LLM during multi-step reasoning, while the LLM provides adaptive feedback instead of acting as a passive tool. We further systematically investigate how collaboration strategies are shaped by SLM and LLM capabilities as well as efficiency and privacy constraints. Evaluation results reveal a distinct scaling effect: stronger SLMs become more self-reliant, while stronger LLMs enable fewer and more informative interactions. In addition, the learned dynamic collaboration strategies significantly outperform static pipelines and standalone inference, and transfer robustly to unseen LLMs.

Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara et al.

Multimodal Large Language Models (MLLMs) have recently achieved promising zero-shot accuracy on visual question answering (VQA) -- a fundamental task affecting various downstream applications and domains. Given the great potential for the broad use of these models, it is important to investigate their limitations in dealing with different image and question properties. In this work, we investigate whether MLLMs can perceive small details as well as large details in images. In particular, we show that their zero-shot accuracy in answering visual questions is very sensitive to the size of the visual subject of the question, declining up to 46% with size. Furthermore, we show that this effect is causal by observing that human visual cropping can significantly mitigate their sensitivity to size. Inspired by the usefulness of human cropping, we then propose five automatic visual cropping methods -- leveraging either external localization models or the decision process of the given MLLM itself -- as inference time mechanisms to improve the zero-shot performance of MLLMs. We study their effectiveness on four popular VQA datasets, and a subset of the VQAv2 dataset tailored towards fine visual details. Our findings suggest that MLLMs should be used with caution in detail-sensitive VQA applications, and that visual cropping is a promising direction to improve their zero-shot performance. To facilitate further investigation of MLLMs' behaviors, our code and data are publicly released.

Jiarui Zhang

In modern commercial systems, including Recommendation, Ranking, and E-Commerce platforms, there is a trend towards improving customer experiences by incorporating Personalization context as input into Large Language Models (LLMs). However, LLMs often struggle to effectively parse and utilize sparse and complex personal context without additional processing or contextual enrichment, underscoring the need for more sophisticated context understanding mechanisms. In this work, we propose Guided Profile Generation (GPG), a general method designed to generate personal profiles in natural language. As is observed, intermediate guided profile generation enables LLMs to summarize, and extract the important, distinctive features from the personal context into concise, descriptive sentences, precisely tailoring their generation more closely to an individual's unique habits and preferences. Our experimental results show that GPG improves LLM's personalization ability across different tasks, for example, it increases 37% accuracy in predicting personal preference compared to directly feeding the LLMs with raw personal context.

Mingzhe Lu, Yiwen Wang, Yanbing Liu et al.

Computational narrative analysis aims to capture rhythm, tension, and emotional dynamics in literary texts. Existing large language models can generate long stories but overly focus on causal coherence, neglecting the complex story arcs and orchestration inherent in human narratives. This suggests a structural misalignment between model- and human-generated narratives. We therefore position narrative analysis as a diagnostic proxy for generation and propose VISTA Space, a high-dimensional framework for narrative orchestration that unifies human and model perspectives while jointly characterizing narrative function and structure in a common space. We further introduce LitVISTA, a structurally annotated benchmark grounded in literary texts, which operationalizes VISTA Space for systematic evaluation of models' narrative orchestration capabilities. Under an oracle setting with gold event anchors, we evaluate frontier LLMs including GPT, Claude, Grok, and Gemini. Results reveal systematic deficiencies, as current models struggle to jointly capture narrative function and structure and fail to form an integrated global view of literary narrative orchestration. End-to-end analysis further shows that failures are dominated by anchor identification and localization errors. Even advanced thinking modes yield mixed and often limited gains for literary narrative understanding.

Zhuoling Li, Hossein Rahmani, Jiarui Zhang et al.

The rapid growth of the text-to-image (T2I) community has fostered a thriving online ecosystem of expert models, which are variants of pretrained diffusion models specialized for diverse generative abilities. Yet, existing model merging methods remain limited in fully leveraging abundant online expert resources and still struggle to meet diverse in-the-wild user needs. We present DiffGraph, a novel agent-driven graph-based model merging framework, which automatically harnesses online experts and flexibly merges them for diverse user needs. Our DiffGraph constructs a scalable graph and organizes ever-expanding online experts within it through node registration and calibration. Then, DiffGraph dynamically activates specific subgraphs based on user needs, enabling flexible combinations of different experts to achieve user-desired generation. Extensive experiments show the efficacy of our method.

Xin Shen, Zhishu Jiang, Jiaye Yang et al. · baidu, tsinghua

Immersive conversational systems in production face a persistent trade-off between responsiveness and long-horizon task capability. Real-time interaction is achievable for lightweight turns, but requests involving planning and tool invocation (e.g., search and media generation) produce heavy-tail execution latency that degrades turn-taking, persona consistency, and user trust. To address this challenge, we propose DuCCAE (Conversation while Collaboration with Augmentation and Evolution), a hybrid engine for immersive conversation deployed within Baidu Search, serving millions of users. DuCCAE decouples real-time response generation from asynchronous agentic execution and synchronizes them via a shared state that maintains session context and execution traces, enabling asynchronous results to be integrated back into the ongoing dialogue. The system orchestrates five subsystems-Info, Conversation, Collaboration, Augmentation, and Evolution-to support multi-agent collaboration and continuous improvement. We evaluate DuCCAE through a comprehensive framework that combines offline benchmarking on the Du-Interact dataset and large-scale production evaluation within Baidu Search. Experimental results demonstrate that DuCCAE outperforms strong baselines in agentic execution reliability and dialogue quality while reducing latency to fit strict real-time budgets. Crucially, deployment metrics since June 2025 confirm substantial real-world effectiveness, evidenced by a tripling of Day-7 user retention to 34.2% and a surge in the complex task completion rate to 65.2%. Our hybrid architecture successfully preserves conversational continuity while enabling reliable agentic execution, offering practical guidelines for deploying scalable agentic systems in industrial settings.

Zhuoling Li, Jiarui Zhang, Ping Hu et al.

Method illustrations (MIs) play a crucial role in conveying the core ideas of scientific papers, yet their generation remains a labor-intensive process. Here, we take inspiration from human authors' drawing practices and correspondingly propose \textbf{FigAgent}, a novel multi-agent framework for high-quality automatic MI generation. Our FigAgent distills drawing experiences from similar components across MIs and encapsulates them into reusable drawing middlewares that can be orchestrated for MI generation, while evolving these middlewares to adapt to dynamically evolving drawing requirements. Besides, a novel Explore-and-Select drawing strategy is introduced to mimic the human-like trial-and-error manner for gradually constructing MIs with complex structures. Extensive experiments show the efficacy of our method.

Jiarui Zhang, Junqi Hu, Zurong Mai et al.

Agricultural multimodal reasoning requires robust spatial understanding across varying scales, from ground-level close-ups to top-down UAV and satellite imagery. Existing Multi-modal Large Language Models (MLLMs) suffer from a significant "terrestrial-centric" bias, causing scale confusion and logic drift during complex agricultural planning. To address this, we introduce the first large-scale AgroOmni (288K), a multi-view training corpus designed to capture diverse spatial topologies and scales in modern precision agriculture. Built on this dataset, we propose AgroNVILA, an MLLM that utilizes a novel Perception-Reasoning Decoupling (PRD) architecture. On the perception side, we incorporate a View-Conditioned Meta-Net (VCMN), which injects macroscopic spatial context into visual tokens, resolving scale ambiguities with minimal computational overhead. On the reasoning side, Agriculture-aware Relative Policy Optimization (ARPO) leverages reinforcement learning to align the model's decision-making with expert agricultural logic, preventing statistical shortcuts. Extensive experiments demonstrate that AgroNVILA outperforms state-of-the-art MLLMs, achieving significant improvements (+15.18%) in multi-altitude agricultural reasoning, reflecting its robust capability for holistic agricultural spatial planning.

Jiarui Zhang, Jinyi Hu, Mahyar Khayatkhoei et al. · tsinghua

Multimodal Large Language Models (MLLMs) have recently shown remarkable perceptual capability in answering visual questions, however, little is known about the limits of their perception. In particular, while prior works have provided anecdotal evidence of MLLMs' sensitivity to object size, this phenomenon and its underlying causes have not been explored comprehensively. In this work, we quantitatively study the perception of small visual objects in several state-of-the-art MLLMs and reveal a pervasive limitation in answering questions about small objects in images. Next, we identify four independent factors that can contribute to this limitation -- object quality, size, distractors, and location -- and conduct controlled intervention studies to measure the effect of each factor on MLLMs' perception. In particular, we find that lower object quality and smaller object size can both independently reduce MLLMs' ability to answer visual questions. More surprisingly, we find that the location of the object in the image and the presence of visual distractors can also significantly reduce MLLMs' question answering accuracy. Our study provides a better understanding of the perceptual limitation of MLLMs and contributes new evaluation protocols for analyzing the perception of future MLLMs. To facilitate further investigations, we release our code and data.

Kian Ahrabian, Zhivar Sourati, Kexuan Sun et al.

While large language models (LLMs) are still being adopted to new domains and utilized in novel applications, we are experiencing an influx of the new generation of foundation models, namely multi-modal large language models (MLLMs). These models integrate verbal and visual information, opening new possibilities to demonstrate more complex reasoning abilities at the intersection of the two modalities. However, despite the revolutionizing prospect of MLLMs, our understanding of their reasoning abilities is limited. In this study, we assess the nonverbal abstract reasoning abilities of open-source and closed-source MLLMs using variations of Raven's Progressive Matrices. Our experiments reveal the challenging nature of such problems for MLLMs while showcasing the immense gap between open-source and closed-source models. We also uncover critical shortcomings of visual and textual perceptions, subjecting the models to low-performance ceilings. Finally, to improve MLLMs' performance, we experiment with different methods, such as Chain-of-Thought prompting, leading to a significant (up to 100%) boost in performance. Our code and datasets are available at https://github.com/usc-isi-i2/isi-mmlm-rpm.

Zhen Fan, Peng Dai, Zhuo Su et al.

Egocentric human pose estimation (HPE) using wearable sensors is essential for VR/AR applications. Most methods rely solely on either egocentric-view images or sparse Inertial Measurement Unit (IMU) signals, leading to inaccuracies due to self-occlusion in images or the sparseness and drift of inertial sensors. Most importantly, the lack of real-world datasets containing both modalities is a major obstacle to progress in this field. To overcome the barrier, we propose EMHI, a multimodal \textbf{E}gocentric human \textbf{M}otion dataset with \textbf{H}ead-Mounted Display (HMD) and body-worn \textbf{I}MUs, with all data collected under the real VR product suite. Specifically, EMHI provides synchronized stereo images from downward-sloping cameras on the headset and IMU data from body-worn sensors, along with pose annotations in SMPL format. This dataset consists of 885 sequences captured by 58 subjects performing 39 actions, totaling about 28.5 hours of recording. We evaluate the annotations by comparing them with optical marker-based SMPL fitting results. To substantiate the reliability of our dataset, we introduce MEPoser, a new baseline method for multimodal egocentric HPE, which employs a multimodal fusion encoder, temporal feature encoder, and MLP-based regression heads. The experiments on EMHI show that MEPoser outperforms existing single-modal methods and demonstrates the value of our dataset in solving the problem of egocentric HPE. We believe the release of EMHI and the method could advance the research of egocentric HPE and expedite the practical implementation of this technology in VR/AR products.

Zhang Li, Yuliang Liu, Qiang Liu et al.

We introduce MonkeyOCR, a vision-language model for document parsing that advances the state of the art by leveraging a Structure-Recognition-Relation (SRR) triplet paradigm. This design simplifies what would otherwise be a complex multi-tool pipeline (as in MinerU's modular approach) and avoids the inefficiencies of processing full pages with giant end-to-end models (e.g., large multimodal LLMs like Qwen-VL). In SRR, document parsing is abstracted into three fundamental questions - "Where is it?" (structure), "What is it?" (recognition), and "How is it organized?" (relation) - corresponding to layout analysis, content identification, and logical ordering. This focused decomposition balances accuracy and speed: it enables efficient, scalable processing without sacrificing precision. To train and evaluate this approach, we introduce the MonkeyDoc (the most comprehensive document parsing dataset to date), with 3.9 million instances spanning over ten document types in both Chinese and English. Experiments show that MonkeyOCR outperforms MinerU by an average of 5.1%, with particularly notable improvements on challenging content such as formulas (+15.0%) and tables (+8.6%). Remarkably, our 3B-parameter model surpasses much larger and top-performing models, including Qwen2.5-VL (72B) and Gemini 2.5 Pro, achieving state-of-the-art average performance on English document parsing tasks. In addition, MonkeyOCR processes multi-page documents significantly faster (0.84 pages per second compared to 0.65 for MinerU and 0.12 for Qwen2.5-VL-7B). The 3B model can be efficiently deployed for inference on a single NVIDIA 3090 GPU. Code and models will be released at https://github.com/Yuliang-Liu/MonkeyOCR.

Jiarui Zhang, Ruixu Geng, Xiaolong Du et al.

Passive non-line-of-sight (NLOS) imaging has witnessed rapid development in recent years, due to its ability to image objects that are out of sight. The light transport condition plays an important role in this task since changing the conditions will lead to different imaging models. Existing learning-based NLOS methods usually train independent models for different light transport conditions, which is computationally inefficient and impairs the practicality of the models. In this work, we propose NLOS-LTM, a novel passive NLOS imaging method that effectively handles multiple light transport conditions with a single network. We achieve this by inferring a latent light transport representation from the projection image and using this representation to modulate the network that reconstructs the hidden image from the projection image. We train a light transport encoder together with a vector quantizer to obtain the light transport representation. To further regulate this representation, we jointly learn both the reconstruction network and the reprojection network during training. A set of light transport modulation blocks is used to modulate the two jointly trained networks in a multi-scale way. Extensive experiments on a large-scale passive NLOS dataset demonstrate the superiority of the proposed method. The code is available at https://github.com/JerryOctopus/NLOS-LTM.

Jiarui Zhang, Xiangyu Liu, Yong Hu et al.

Retrieval-Augmented Generation (RAG) significantly improves the performance of Large Language Models (LLMs) on knowledge-intensive tasks. However, varying response quality across LLMs under RAG necessitates intelligent routing mechanisms, which select the most suitable model for each query from multiple retrieval-augmented LLMs via a dedicated router model. We observe that external documents dynamically affect LLMs' ability to answer queries, while existing routing methods, which rely on static parametric knowledge representations, exhibit suboptimal performance in RAG scenarios. To address this, we formally define the new retrieval-augmented LLM routing problem, incorporating the influence of retrieved documents into the routing framework. We propose RAGRouter, a RAG-aware routing design, which leverages document embeddings and RAG capability embeddings with contrastive learning to capture knowledge representation shifts and enable informed routing decisions. Extensive experiments on diverse knowledge-intensive tasks and retrieval settings, covering open and closed-source LLMs, show that RAGRouter outperforms the best individual LLM and existing routing methods. With an extended score-threshold-based mechanism, it also achieves strong performance-efficiency trade-offs under low-latency constraints. The code and data are available at https://github.com/OwwO99/RAGRouter.

Yibin Wen, Qingmei Li, Zi Ye et al.

Recent advancements in Vision-Language Models (VLMs) have significantly transformed various industries. In agriculture, these dual-modal capabilities offer promising applications such as precision farming, crop monitoring, pest detection, and environmental sustainability. While several Visual Question Answering (VQA) datasets and benchmarks have been developed to evaluate VLM performance, they often fail to adequately assess the critical reasoning and problem-solving skills required in complex agricultural contexts. To address this gap, we introduce AgriCoT, a VQA dataset that incorporates Chain-of-Thought (CoT) reasoning, specifically designed to evaluate the reasoning capabilities of VLMs. With 4,535 carefully curated samples, AgriCoT offers a comprehensive and robust evaluation of reasoning abilities for VLMs, particularly in zero-shot scenarios, by focusing on their capacity to engage in logical reasoning and effective problem-solving. Our evaluations, conducted with 26 representative VLMs, including both proprietary and open-source models, reveal that while some proprietary models excel at answering questions, there is a notable and significant gap in their reasoning capabilities. This underscores the importance of incorporating CoT for more precise and effective assessments. Our dataset are available at https://huggingface.co/datasets/wenyb/AgriCoT.

Qingmei Li, Yang Zhang, Zurong Mai et al.

Large Multimodal Models (LMMs) has demonstrated capabilities across various domains, but comprehensive benchmarks for agricultural remote sensing (RS) remain scarce. Existing benchmarks designed for agricultural RS scenarios exhibit notable limitations, primarily in terms of insufficient scene diversity in the dataset and oversimplified task design. To bridge this gap, we introduce AgroMind, a comprehensive agricultural remote sensing benchmark covering four task dimensions: spatial perception, object understanding, scene understanding, and scene reasoning, with a total of 13 task types, ranging from crop identification and health monitoring to environmental analysis. We curate a high-quality evaluation set by integrating eight public datasets and one private farmland plot dataset, containing 27,247 QA pairs and 19,615 images. The pipeline begins with multi-source data pre-processing, including collection, format standardization, and annotation refinement. We then generate a diverse set of agriculturally relevant questions through the systematic definition of tasks. Finally, we employ LMMs for inference, generating responses, and performing detailed examinations. We evaluated 20 open-source LMMs and 4 closed-source models on AgroMind. Experiments reveal significant performance gaps, particularly in spatial reasoning and fine-grained recognition, it is notable that human performance lags behind several leading LMMs. By establishing a standardized evaluation framework for agricultural RS, AgroMind reveals the limitations of LMMs in domain knowledge and highlights critical challenges for future work. Data and code can be accessed at https://rssysu.github.io/AgroMind/.

Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara et al.

Visual Question Answering is a challenging task, as it requires seamless interaction between perceptual, linguistic, and background knowledge systems. While the recent progress of visual and natural language models like BLIP has led to improved performance on this task, we lack understanding of the ability of such models to perform on different kinds of questions and reasoning types. As our initial analysis of BLIP-family models revealed difficulty with answering fine-detail questions, we investigate the following question: Can visual cropping be employed to improve the performance of state-of-the-art visual question answering models on fine-detail questions? Given the recent success of the BLIP-family models, we study a zero-shot and a fine-tuned BLIP model. We define three controlled subsets of the popular VQA-v2 benchmark to measure whether cropping can help model performance. Besides human cropping, we devise two automatic cropping strategies based on multi-modal embedding by CLIP and BLIP visual QA model gradients. Our experiments demonstrate that the performance of BLIP model variants can be significantly improved through human cropping, and automatic cropping methods can produce comparable benefits. A deeper dive into our findings indicates that the performance enhancement is more pronounced in zero-shot models than in fine-tuned models and more salient with smaller bounding boxes than larger ones. We perform case studies to connect quantitative differences with qualitative observations across question types and datasets. Finally, we see that the cropping enhancement is robust, as we gain an improvement of 4.59% (absolute) in the general VQA-random task by simply inputting a concatenation of the original and gradient-based cropped images. We make our code available to facilitate further innovation on visual cropping methods for question answering.

Wenhao Li, Fanchao Qi, Maosong Sun et al.

Poetry is one of the most important art forms of human languages. Recently many studies have focused on incorporating some linguistic features of poetry, such as style and sentiment, into its understanding or generation system. However, there is no focus on understanding or evaluating the semantics of poetry. Therefore, we propose a novel task to assess a model's semantic understanding of poetry by poem matching. Specifically, this task requires the model to select one line of Chinese classical poetry among four candidates according to the modern Chinese translation of a line of poetry. To construct this dataset, we first obtain a set of parallel data of Chinese classical poetry and modern Chinese translation. Then we retrieve similar lines of poetry with the lines in a poetry corpus as negative choices. We name the dataset Chinese Classical Poetry Matching Dataset (CCPM) and release it at https://github.com/THUNLP-AIPoet/CCPM. We hope this dataset can further enhance the study on incorporating deep semantics into the understanding and generation system of Chinese classical poetry. We also preliminarily run two variants of BERT on this dataset as the baselines for this dataset.

Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara et al.

Multimodal Large Language Models (MLLMs) have experienced rapid progress in visual recognition tasks in recent years. Given their potential integration into many critical applications, it is important to understand the limitations of their visual perception. In this work, we study whether MLLMs can perceive small visual details as effectively as large ones when answering questions about images. We observe that their performance is very sensitive to the size of the visual subject of the question, and further show that this effect is in fact causal by conducting an intervention study. Next, we study the attention patterns of MLLMs when answering visual questions, and intriguingly find that they consistently know where to look, even when they provide the wrong answer. Based on these findings, we then propose training-free visual intervention methods that leverage the internal knowledge of any MLLM itself, in the form of attention and gradient maps, to enhance its perception of small visual details. We evaluate our proposed methods on two widely-used MLLMs and seven visual question answering benchmarks and show that they can significantly improve MLLMs' accuracy without requiring any training. Our results elucidate the risk of applying MLLMs to visual recognition tasks concerning small details and indicate that visual intervention using the model's internal state is a promising direction to mitigate this risk.

Mingzhe Lu, Yanbing Liu, Qihao Wang et al.

Movie scripts pose a fundamental challenge for automatic summarization due to their non-linear, cross-cut narrative structure, which makes surface-level saliency methods ineffective at preserving core story progression. To address this, we introduce S^2tory (Story Spine Distillation), a narratology-grounded framework that leverages character development trajectories to identify plot nuclei, the essential events that drive the narrative forward, while filtering out peripheral satellite events that merely enrich atmosphere or emotion. Our Narrative Expert Agent (NEAgent) performs theory-constrained reasoning, whose distilled knowledge conditions a small model to identify plot nuclei. Another model then uses these plot nuclei to generate the summary. Experiments on the MovieSum dataset demonstrate state-of-the-art semantic fidelity at approximately 3.5x compression, and zero-shot evaluation on BookSum confirms strong out-of-domain generalization. Human evaluation further validates that narratological theory provides an indispensable foundation for modeling complex, non-linear narratives.

Yifan Jiang, Jiarui Zhang, Kexuan Sun et al.

While multi-modal large language models (MLLMs) have shown significant progress on many popular visual reasoning benchmarks, whether they possess abstract visual reasoning abilities remains an open question. Similar to the Sudoku puzzles, abstract visual reasoning (AVR) problems require finding high-level patterns (e.g., repetition constraints) that control the input shapes (e.g., digits) in a specific task configuration (e.g., matrix). However, existing AVR benchmarks only considered a limited set of patterns (addition, conjunction), input shapes (rectangle, square), and task configurations (3 by 3 matrices). To evaluate MLLMs' reasoning abilities comprehensively, we introduce MARVEL, a multidimensional AVR benchmark with 770 puzzles composed of six core knowledge patterns, geometric and abstract shapes, and five different task configurations. To inspect whether the model accuracy is grounded in perception and reasoning, MARVEL complements the general AVR question with perception questions in a hierarchical evaluation framework. We conduct comprehensive experiments on MARVEL with nine representative MLLMs in zero-shot and few-shot settings. Our experiments reveal that all models show near-random performance on the AVR question, with significant performance gaps (40%) compared to humans across all patterns and task configurations. Further analysis of perception questions reveals that MLLMs struggle to comprehend the visual features (near-random performance) and even count the panels in the puzzle ( <45%), hindering their ability for abstract reasoning. We release our entire code and dataset.

Jiarui Zhang, Ollie Liu, Tianyu Yu et al. · tsinghua

Multimodal large language models (MLLMs) have made rapid progress in recent years, yet continue to struggle with low-level visual perception (LLVP) -- particularly the ability to accurately describe the geometric details of an image. This capability is crucial for applications in areas such as robotics, medical image analysis, and manufacturing. In this paper, we first introduce Geoperception, a benchmark designed to evaluate an MLLM's ability to accurately transcribe 2D geometric information from an image. Using this benchmark, we demonstrate the limitations of leading MLLMs, and then conduct a comprehensive empirical study to explore strategies for improving their performance on geometric tasks. Our findings highlight the benefits of certain model architectures, training techniques, and data strategies, including the use of high-fidelity synthetic data and multi-stage training with a data curriculum. Notably, we find that a data curriculum enables models to learn challenging geometry understanding tasks which they fail to learn from scratch. Leveraging these insights, we develop Euclid, a family of models specifically optimized for strong low-level geometric perception. Although purely trained on synthetic multimodal data, Euclid shows strong generalization ability to novel geometry shapes. For instance, Euclid outperforms the best closed-source model, Gemini-1.5-Pro, by up to 58.56% on certain Geoperception benchmark tasks and 10.65% on average across all tasks.

Jinta Weng, Jiarui Zhang, Yue Hu et al.

Large language models (LLMs) can be used as accessible and intelligent chatbots by constructing natural language queries and directly inputting the prompt into the large language model. However, different prompt' constructions often lead to uncertainty in the answers and thus make it hard to utilize the specific knowledge of LLMs (like ChatGPT). To alleviate this, we use an interpretable structure to explain the prompt learning principle in LLMs, which certificates that the effectiveness of language models is determined by position changes of the task's related tokens. Therefore, we propose MTPrompt, a multi-dimensional task prompt learning method consisting based on task-related object, summary, and task description information. By automatically building and searching for appropriate prompts, our proposed MTPrompt achieves the best results on few-shot samples setting and five different datasets. In addition, we demonstrate the effectiveness and stability of our method in different experimental settings and ablation experiments. In interaction with large language models, embedding more task-related information into prompts will make it easier to stimulate knowledge embedded in large language models.

Junyu Lai, Jiakun Zhang, Shuo Xu et al.

Recent advancements in large language models (LLMs) have sparked considerable interest in automated theorem proving and a prominent line of research integrates stepwise LLM-based provers into tree search. In this paper, we introduce a novel proof-state exploration approach for training data synthesis, designed to produce diverse tactics across a wide range of intermediate proof states, thereby facilitating effective one-shot fine-tuning of LLM as the policy model. We also propose an adaptive beam size strategy, which effectively takes advantage of our data synthesis method and achieves a trade-off between exploration and exploitation during tree search. Evaluations on the MiniF2F and ProofNet benchmarks demonstrate that our method outperforms strong baselines under the stringent Pass@1 metric, attaining an average pass rate of $60.74\%$ on MiniF2F and $21.18\%$ on ProofNet. These results underscore the impact of large-scale synthetic data in advancing automated theorem proving.

Yingjie Chen, Jiarui Zhang, Tao Wang et al.

With the increasing need for facial behavior analysis, semi-supervised AU intensity estimation using only keyframe annotations has emerged as a practical and effective solution to relieve the burden of annotation. However, the lack of annotations makes the spurious correlation problem caused by AU co-occurrences and subject variation much more prominent, leading to non-robust intensity estimation that is entangled among AUs and biased among subjects. We observe that trend information inherent in keyframe annotations could act as extra supervision and raising the awareness of AU-specific facial appearance changing trends during training is the key to learning invariant AU-specific features. To this end, we propose \textbf{T}rend-\textbf{A}ware \textbf{S}upervision (TAS), which pursues three kinds of trend awareness, including intra-trend ranking awareness, intra-trend speed awareness, and inter-trend subject awareness. TAS alleviates the spurious correlation problem by raising trend awareness during training to learn AU-specific features that represent the corresponding facial appearance changes, to achieve intensity estimation invariance. Experiments conducted on two commonly used AU benchmark datasets, BP4D and DISFA, show the effectiveness of each kind of awareness. And under trend-aware supervision, the performance can be improved without extra computational or storage costs during inference.

Jiarui Zhang, Shaojuan Wu, Xiaowang Zhang et al.

Stance detection classifies stance relations (namely, Favor, Against, or Neither) between comments and targets. Pretrained language models (PLMs) are widely used to mine the stance relation to improve the performance of stance detection through pretrained knowledge. However, PLMs also embed ``bad'' pretrained knowledge concerning stance into the extracted stance relation semantics, resulting in pretrained stance bias. It is not trivial to measure pretrained stance bias due to its weak quantifiability. In this paper, we propose Relative Counterfactual Contrastive Learning (RCCL), in which pretrained stance bias is mitigated as relative stance bias instead of absolute stance bias to overtake the difficulty of measuring bias. Firstly, we present a new structural causal model for characterizing complicated relationships among context, PLMs and stance relations to locate pretrained stance bias. Then, based on masked language model prediction, we present a target-aware relative stance sample generation method for obtaining relative bias. Finally, we use contrastive learning based on counterfactual theory to mitigate pretrained stance bias and preserve context stance relation. Experiments show that the proposed method is superior to stance detection and debiasing baselines.

Yu Bai, Zhe Wang, Jiarui Zhang et al.

The trade-off between clean accuracy and adversarial robustness is a pervasive phenomenon in deep learning, yet its geometric origin remains elusive. In this work, we utilize Symmetry-Breaking Dimensional Expansion (SBDE) as a controlled probe to investigate the mechanism underlying this trade-off. SBDE expands input images by inserting constant-valued pixels, which breaks translational symmetry and consistently improves clean accuracy (e.g., from $90.47\%$ to $95.63\%$ on CIFAR-10 with ResNet-18) by reducing parameter degeneracy. However, this accuracy gain comes at the cost of reduced robustness against iterative white-box attacks. By employing a test-time \emph{mask projection} that resets the inserted auxiliary pixels to their training values, we demonstrate that the vulnerability stems almost entirely from the inserted dimensions. The projection effectively neutralizes the attacks and restores robustness, revealing that the model achieves high accuracy by creating \emph{sharp boundaries} (steep loss gradients) specifically along the auxiliary axes. Our findings provide a concrete geometric explanation for the accuracy-robustness paradox: the optimization landscape deepens the basin of attraction to improve accuracy but inevitably erects steep walls along the auxiliary degrees of freedom, creating a fragile sensitivity to off-manifold perturbations.

Junqiao Fan, Yunjiao Zhou, Yizhuo Yang et al.

Human mesh reconstruction (HMR) provides direct insights into body-environment interaction, which enables various immersive applications. While existing large-scale HMR datasets rely heavily on line-of-sight RGB input, vision-based sensing is limited by occlusion, lighting variation, and privacy concerns. To overcome these limitations, recent efforts have explored radio-frequency (RF) mmWave radar for privacy-preserving indoor human sensing. However, current radar datasets are constrained by sparse skeleton labels, limited scale, and simple in-place actions. To advance the HMR research community, we introduce M4Human, the current largest-scale (661K-frame) ($9\times$ prior largest) multimodal benchmark, featuring high-resolution mmWave radar, RGB, and depth data. M4Human provides both raw radar tensors (RT) and processed radar point clouds (RPC) to enable research across different levels of RF signal granularity. M4Human includes high-quality motion capture (MoCap) annotations with 3D meshes and global trajectories, and spans 20 subjects and 50 diverse actions, including in-place, sit-in-place, and free-space sports or rehabilitation movements. We establish benchmarks on both RT and RPC modalities, as well as multimodal fusion with RGB-D modalities. Extensive results highlight the significance of M4Human for radar-based human modeling while revealing persistent challenges under fast, unconstrained motion. The dataset and code will be released after the paper publication.

Jiarui Zhang, Zhihao Li, Chong Wang et al.

Neural fields (NFs) have demonstrated remarkable performance in scene reconstruction, powering various tasks such as novel view synthesis. However, existing NF methods relying on RGB or LiDAR inputs often exhibit severe fragility to adverse weather, particularly when applied in outdoor scenarios like autonomous driving. In contrast, millimeter-wave radar is inherently robust to environmental changes, while unfortunately, its integration with NFs remains largely underexplored. Besides, as outdoor driving scenarios frequently involve moving objects, making spatiotemporal modeling essential for temporally consistent novel view synthesis. To this end, we introduce RF4D, a radar-based neural field framework specifically designed for novel view synthesis in outdoor dynamic scenes. RF4D explicitly incorporates temporal information into its representation, significantly enhancing its capability to model moving objects. We further introduce a feature-level flow module that predicts latent temporal offsets between adjacent frames, enforcing temporal coherence in dynamic scene modeling. Moreover, we propose a radar-specific power rendering formulation closely aligned with radar sensing physics, improving synthesis accuracy and interoperability. Extensive experiments on public radar datasets demonstrate the superior performance of RF4D in terms of radar measurement synthesis quality and occupancy estimation accuracy, achieving especially pronounced improvements in dynamic outdoor scenarios.

Prateek Chhikara, Jiarui Zhang, Filip Ilievski et al.

Communication via natural language is a key aspect of machine intelligence, and it requires computational models to learn and reason about world concepts, with varying levels of supervision. Significant progress has been made on fully-supervised non-interactive tasks, such as question-answering and procedural text understanding. Yet, various sequential interactive tasks, as in text-based games, have revealed limitations of existing approaches in terms of coherence, contextual awareness, and their ability to learn effectively from the environment. In this paper, we propose a knowledge-injection framework for improved functional grounding of agents in text-based games. Specifically, we consider two forms of domain knowledge that we inject into learning-based agents: memory of previous correct actions and affordances of relevant objects in the environment. Our framework supports two representative model classes: reinforcement learning agents and language model agents. Furthermore, we devise multiple injection strategies for the above domain knowledge types and agent architectures, including injection via knowledge graphs and augmentation of the existing input encoding strategies. We experiment with four models on the 10 tasks in the ScienceWorld text-based game environment, to illustrate the impact of knowledge injection on various model configurations and challenging task settings. Our findings provide crucial insights into the interplay between task properties, model architectures, and domain knowledge for interactive contexts.

Junxuan Chen, Xiang Li, Jiarui Zhang et al.

Recently, document-level neural machine translation (NMT) has become a hot topic in the community of machine translation. Despite its success, most of existing studies ignored the discourse structure information of the input document to be translated, which has shown effective in other tasks. In this paper, we propose to improve document-level NMT with the aid of discourse structure information. Our encoder is based on a hierarchical attention network (HAN). Specifically, we first parse the input document to obtain its discourse structure. Then, we introduce a Transformer-based path encoder to embed the discourse structure information of each word. Finally, we combine the discourse structure information with the word embedding before it is fed into the encoder. Experimental results on the English-to-German dataset show that our model can significantly outperform both Transformer and Transformer+HAN.

Haofu Liao, Wei-An Lin, Jiarui Zhang et al.

We propose to tackle the problem of multiview 2D/3D rigid registration for intervention via a Point-Of-Interest Network for Tracking and Triangulation ($\text{POINT}^2$). $\text{POINT}^2$ learns to establish 2D point-to-point correspondences between the pre- and intra-intervention images by tracking a set of random POIs. The 3D pose of the pre-intervention volume is then estimated through a triangulation layer. In $\text{POINT}^2$, the unified framework of the POI tracker and the triangulation layer enables learning informative 2D features and estimating 3D pose jointly. In contrast to existing approaches, $\text{POINT}^2$ only requires a single forward-pass to achieve a reliable 2D/3D registration. As the POI tracker is shift-invariant, $\text{POINT}^2$ is more robust to the initial pose of the 3D pre-intervention image. Extensive experiments on a large-scale clinical cone-beam CT (CBCT) dataset show that the proposed $\text{POINT}^2$ method outperforms the existing learning-based method in terms of accuracy, robustness and running time. Furthermore, when used as an initial pose estimator, our method also improves the robustness and speed of the state-of-the-art optimization-based approaches by ten folds.