Tianhao Li

Xiangpeng Wei, Haoran Wei, Huan Lin et al.

Large language models (LLMs) demonstrate remarkable ability to comprehend, reason, and generate following nature language instructions. However, the development of LLMs has been primarily focused on high-resource languages, such as English, thereby limiting their applicability and research in other languages. Consequently, we present PolyLM, a multilingual LLM trained on 640 billion (B) tokens, avaliable in two model sizes: 1.7B and 13B. To enhance its multilingual capabilities, we 1) integrate bilingual data into training data; and 2) adopt a curriculum learning strategy that increases the proportion of non-English data from 30% in the first stage to 60% in the final stage during pre-training. Further, we propose a multilingual self-instruct method which automatically generates 132.7K diverse multilingual instructions for model fine-tuning. To assess the model's performance, we collect several existing multilingual tasks, including multilingual understanding, question answering, generation, and translation. Extensive experiments show that PolyLM surpasses other open-source models such as LLaMA and BLOOM on multilingual tasks while maintaining comparable performance in English. Our models, alone with the instruction data and multilingual benchmark, are available at: \url{https://modelscope.cn/models/damo/nlp_polylm_13b_text_generation}.

An Yang, Baosong Yang, Binyuan Hui et al.

This report introduces the Qwen2 series, the latest addition to our large language models and large multimodal models. We release a comprehensive suite of foundational and instruction-tuned language models, encompassing a parameter range from 0.5 to 72 billion, featuring dense models and a Mixture-of-Experts model. Qwen2 surpasses most prior open-weight models, including its predecessor Qwen1.5, and exhibits competitive performance relative to proprietary models across diverse benchmarks on language understanding, generation, multilingual proficiency, coding, mathematics, and reasoning. The flagship model, Qwen2-72B, showcases remarkable performance: 84.2 on MMLU, 37.9 on GPQA, 64.6 on HumanEval, 89.5 on GSM8K, and 82.4 on BBH as a base language model. The instruction-tuned variant, Qwen2-72B-Instruct, attains 9.1 on MT-Bench, 48.1 on Arena-Hard, and 35.7 on LiveCodeBench. Moreover, Qwen2 demonstrates robust multilingual capabilities, proficient in approximately 30 languages, spanning English, Chinese, Spanish, French, German, Arabic, Russian, Korean, Japanese, Thai, Vietnamese, and more, underscoring its versatility and global reach. To foster community innovation and accessibility, we have made the Qwen2 model weights openly available on Hugging Face and ModelScope, and the supplementary materials including example code on GitHub. These platforms also include resources for quantization, fine-tuning, and deployment, facilitating a wide range of applications and research endeavors.

Tianhao Li, Sandesh Shetty, Advaith Kamath et al.

Large pre-trained language models (LLMs) have been shown to have significant potential in few-shot learning across various fields, even with minimal training data. However, their ability to generalize to unseen tasks in more complex fields, such as biology, has yet to be fully evaluated. LLMs can offer a promising alternative approach for biological inference, particularly in cases where structured data and sample size are limited, by extracting prior knowledge from text corpora. Our proposed few-shot learning approach uses LLMs to predict the synergy of drug pairs in rare tissues that lack structured data and features. Our experiments, which involved seven rare tissues from different cancer types, demonstrated that the LLM-based prediction model achieved significant accuracy with very few or zero samples. Our proposed model, the CancerGPT (with $\sim$ 124M parameters), was even comparable to the larger fine-tuned GPT-3 model (with $\sim$ 175B parameters). Our research is the first to tackle drug pair synergy prediction in rare tissues with limited data. We are also the first to utilize an LLM-based prediction model for biological reaction prediction tasks.

Gregor von Laszewski, Wesley Brewer, Jeyan Thiyagalingam et al.

Benchmarks are a cornerstone of modern machine learning, enabling reproducibility, comparison, and scientific progress. However, AI benchmarks are increasingly complex, requiring dynamic, AI-focused workflows. Rapid evolution in model architectures, scale, datasets, and deployment contexts makes evaluation a moving target. Large language models often memorize static benchmarks, causing a gap between benchmark results and real-world performance. Beyond traditional static benchmarks, continuous adaptive benchmarking frameworks are needed to align scientific assessment with deployment risks. This calls for skills and education in AI Benchmark Carpentry. From our experience with MLCommons, educational initiatives, and programs like the DOE's Trillion Parameter Consortium, key barriers include high resource demands, limited access to specialized hardware, lack of benchmark design expertise, and uncertainty in relating results to application domains. Current benchmarks often emphasize peak performance on top-tier hardware, offering limited guidance for diverse, real-world scenarios. Benchmarking must become dynamic, incorporating evolving models, updated data, and heterogeneous platforms while maintaining transparency, reproducibility, and interpretability. Democratization requires both technical innovation and systematic education across levels, building sustained expertise in benchmark design and use. Benchmarks should support application-relevant comparisons, enabling informed, context-sensitive decisions. Dynamic, inclusive benchmarking will ensure evaluation keeps pace with AI evolution and supports responsible, reproducible, and accessible AI deployment. Community efforts can provide a foundation for AI Benchmark Carpentry.

Boyi Deng, Xu Wang, Yaoning Wang et al.

Large language models have achieved remarkable capabilities across diverse tasks, yet their internal decision-making processes remain largely opaque, limiting our ability to inspect, control, and systematically improve them. This opacity motivates a growing body of research in mechanistic interpretability, with sparse autoencoders (SAEs) emerging as one of the most promising tools for decomposing model activations into sparse, interpretable feature representations. We introduce Qwen-Scope, an open-source suite of SAEs built on the Qwen model family, comprising 14 groups of SAEs across 7 model variants from the Qwen3 and Qwen3.5 series, covering both dense and mixture-of-expert architectures. Built on top of these SAEs, we show that SAEs can go beyond post-hoc analysis to serve as practical interfaces for model development along four directions: (i) inference-time steering, where SAE feature directions control language, concepts, and preferences without modifying model weights; (ii) evaluation analysis, where activated SAE features provide a representation-level proxy for benchmark redundancy and capability coverage; (iii) data-centric workflows, where SAE features support multilingual toxicity classification and safety-oriented data synthesis; and (iv) post-training optimization, where SAE-derived signals are incorporated into supervised fine-tuning and reinforcement learning objectives to mitigate undesirable behaviors such as code-switching and repetition. Together, these results demonstrate that SAEs can serve not only as post-hoc analysis tools, but also as reusable representation-level interfaces for diagnosing, controlling, evaluating, and improving large language models. By open-sourcing Qwen-Scope, we aim to support mechanistic research and accelerate practical workflows that connect model internals to downstream behavior.

Yiruo Cheng, Kelong Mao, Tianhao Li et al.

Conversational shopping agents represent a critical consumer-facing application of Large Language Model (LLM)-powered agents, yet how to effectively apply post-training Reinforcement Learning (RL) to optimize such agents remains underexplored. This work investigates RL-based optimization for shopping agents in real-world scenarios, where agents must simultaneously satisfy multiple interdependent objectives spanning objective metrics (product correctness), subjective qualities (persuasiveness), outcome rewards (final response quality), and process rewards (tool efficiency). We present a complete methodology to address this challenge. Specifically, we first construct SmartShopBench, a benchmark that captures diverse shopping intents with a hierarchical evaluation that decomposes complex quality requirements into measurable levels. Building on this evaluation framework, we design Hierarchical Reward Modeling (HRM) to structure mixed reward types through conditional gating that reflects their logical dependencies. To enable efficient training, we further propose Dynamic Contrastive Policy Optimization (DCPO), which balances response quality with operational efficiency through dynamic trajectory selection based on reward and reasoning length. Extensive experiments demonstrate that our RL-trained agent, namely ChatShopBuddy, consistently outperforms larger models relying on generic reasoning, achieving superior stability rather than merely higher peaks. Our work provides valuable guidance for applying RL to real-world conversational agents.

Yaqi Wang, Yifan Zhang, Xiaodiao Chen et al.

Computer-aided design (CAD) tools are increasingly popular in modern dental practice, particularly for treatment planning or comprehensive prognosis evaluation. In particular, the 2D panoramic X-ray image efficiently detects invisible caries, impacted teeth and supernumerary teeth in children, while the 3D dental cone beam computed tomography (CBCT) is widely used in orthodontics and endodontics due to its low radiation dose. However, there is no open-access 2D public dataset for children's teeth and no open 3D dental CBCT dataset, which limits the development of automatic algorithms for segmenting teeth and analyzing diseases. The Semi-supervised Teeth Segmentation (STS) Challenge, a pioneering event in tooth segmentation, was held as a part of the MICCAI 2023 ToothFairy Workshop on the Alibaba Tianchi platform. This challenge aims to investigate effective semi-supervised tooth segmentation algorithms to advance the field of dentistry. In this challenge, we provide two modalities including the 2D panoramic X-ray images and the 3D CBCT tooth volumes. In Task 1, the goal was to segment tooth regions in panoramic X-ray images of both adult and pediatric teeth. Task 2 involved segmenting tooth sections using CBCT volumes. Limited labelled images with mostly unlabelled ones were provided in this challenge prompt using semi-supervised algorithms for training. In the preliminary round, the challenge received registration and result submission by 434 teams, with 64 advancing to the final round. This paper summarizes the diverse methods employed by the top-ranking teams in the STS MICCAI 2023 Challenge.

Haobo Zhang, Yutao Zhu, Kelong Mao et al.

Large Language Models (LLMs) have revolutionized recommendation agents by providing superior reasoning and flexible decision-making capabilities. However, existing methods mainly follow a passive information acquisition paradigm, where agents either rely on static pre-defined workflows or perform reasoning with constrained information. It limits the agent's ability to identify information sufficiency, often leading to suboptimal recommendations when faced with fragmented user profiles or sparse item metadata. To address these limitations, we propose RecThinker, an agentic framework for tool-augmented reasoning in recommendation, which shifts recommendation from passive processing to autonomous investigation by dynamically planning reasoning paths and proactively acquiring essential information via autonomous tool-use. Specifically, RecThinker adopts an Analyze-Plan-Act paradigm, which first analyzes the sufficiency of user-item information and autonomously invokes tool-calling sequences to bridge information gaps between available knowledge and reasoning requirements. We develop a suite of specialized tools for RecThinker, enabling the model to acquire user-side, item-side, and collaborative information for better reasoning and user-item matching. Furthermore, we introduce a self-augmented training pipeline, comprising a Supervised Fine-Tuning (SFT) stage to internalize high-quality reasoning trajectories and a Reinforcement Learning (RL) stage to optimize for decision accuracy and tool-use efficiency. Extensive experiments on multiple benchmark datasets demonstrate that RecThinker consistently outperforms strong baselines in the recommendation scenario.

An Yang, Anfeng Li, Baosong Yang et al. · tsinghua

In this work, we present Qwen3, the latest version of the Qwen model family. Qwen3 comprises a series of large language models (LLMs) designed to advance performance, efficiency, and multilingual capabilities. The Qwen3 series includes models of both dense and Mixture-of-Expert (MoE) architectures, with parameter scales ranging from 0.6 to 235 billion. A key innovation in Qwen3 is the integration of thinking mode (for complex, multi-step reasoning) and non-thinking mode (for rapid, context-driven responses) into a unified framework. This eliminates the need to switch between different models--such as chat-optimized models (e.g., GPT-4o) and dedicated reasoning models (e.g., QwQ-32B)--and enables dynamic mode switching based on user queries or chat templates. Meanwhile, Qwen3 introduces a thinking budget mechanism, allowing users to allocate computational resources adaptively during inference, thereby balancing latency and performance based on task complexity. Moreover, by leveraging the knowledge from the flagship models, we significantly reduce the computational resources required to build smaller-scale models, while ensuring their highly competitive performance. Empirical evaluations demonstrate that Qwen3 achieves state-of-the-art results across diverse benchmarks, including tasks in code generation, mathematical reasoning, agent tasks, etc., competitive against larger MoE models and proprietary models. Compared to its predecessor Qwen2.5, Qwen3 expands multilingual support from 29 to 119 languages and dialects, enhancing global accessibility through improved cross-lingual understanding and generation capabilities. To facilitate reproducibility and community-driven research and development, all Qwen3 models are publicly accessible under Apache 2.0.

Qwen, An Yang, Baosong Yang et al.

In this report, we introduce Qwen2.5, a comprehensive series of large language models (LLMs) designed to meet diverse needs. Compared to previous iterations, Qwen 2.5 has been significantly improved during both the pre-training and post-training stages. In terms of pre-training, we have scaled the high-quality pre-training datasets from the previous 7 trillion tokens to 18 trillion tokens. This provides a strong foundation for common sense, expert knowledge, and reasoning capabilities. In terms of post-training, we implement intricate supervised finetuning with over 1 million samples, as well as multistage reinforcement learning. Post-training techniques enhance human preference, and notably improve long text generation, structural data analysis, and instruction following. To handle diverse and varied use cases effectively, we present Qwen2.5 LLM series in rich sizes. Open-weight offerings include base and instruction-tuned models, with quantized versions available. In addition, for hosted solutions, the proprietary models currently include two mixture-of-experts (MoE) variants: Qwen2.5-Turbo and Qwen2.5-Plus, both available from Alibaba Cloud Model Studio. Qwen2.5 has demonstrated top-tier performance on a wide range of benchmarks evaluating language understanding, reasoning, mathematics, coding, human preference alignment, etc. Specifically, the open-weight flagship Qwen2.5-72B-Instruct outperforms a number of open and proprietary models and demonstrates competitive performance to the state-of-the-art open-weight model, Llama-3-405B-Instruct, which is around 5 times larger. Qwen2.5-Turbo and Qwen2.5-Plus offer superior cost-effectiveness while performing competitively against GPT-4o-mini and GPT-4o respectively. Additionally, as the foundation, Qwen2.5 models have been instrumental in training specialized models such as Qwen2.5-Math, Qwen2.5-Coder, QwQ, and multimodal models.

Hao Zhou, Tianhao Li, Zhijun Wang et al.

Expanding Large Language Models~(LLMs) to new languages is a costly endeavor, demanding extensive Continued Pre-Training~(CPT) and data-intensive alignment. While recent data-free merging techniques attempt to bypass alignment by fusing a multilingual CPT-enhanced model with its instruct counterpart, they are plagued by a critical trade-off: mitigating parameter conflicts to preserve original abilities inevitably dilutes new language acquisition, and vice-versa. To resolve this conflict, we introduce \method, which upcycles a dense model into a Mixture-of-Experts~(MoE) architecture, allocating different experts to different languages. Alignment ability is then transferred by grafting a MoE-expanded parameter delta~($Δ_{\text{post}}$) to the CPT-enhanced base model, bypassing the complex alignment phase. Experiments demonstrate \method's superiority even against baselines with similar FLOPs or number of parameters; it improves performance on expanded languages while effectively preserving original capabilities. We further show our approach is highly applicable across different models and Post-training deltas.

Haolin Qin, Tingfa Xu, Tianhao Li et al.

UAV tracking faces significant challenges in real-world scenarios, such as small-size targets and occlusions, which limit the performance of RGB-based trackers. Multispectral images (MSI), which capture additional spectral information, offer a promising solution to these challenges. However, progress in this field has been hindered by the lack of relevant datasets. To address this gap, we introduce the first large-scale Multispectral UAV Single Object Tracking dataset (MUST), which includes 250 video sequences spanning diverse environments and challenges, providing a comprehensive data foundation for multispectral UAV tracking. We also propose a novel tracking framework, UNTrack, which encodes unified spectral, spatial, and temporal features from spectrum prompts, initial templates, and sequential searches. UNTrack employs an asymmetric transformer with a spectral background eliminate mechanism for optimal relationship modeling and an encoder that continuously updates the spectrum prompt to refine tracking, improving both accuracy and efficiency. Extensive experiments show that our proposed UNTrack outperforms state-of-the-art UAV trackers. We believe our dataset and framework will drive future research in this area. The dataset is available on https://github.com/q2479036243/MUST-Multispectral-UAV-Single-Object-Tracking.

Qian Xiong, Yuekai Huang, Bo Yang et al.

LLMs have advanced tool-using agents for real-world applications, yet they often lead to unexpected behaviors or results. Beyond obvious failures, the subtle issue of "intent deviation" severely hinders reliable evaluation and performance improvement. Existing post-training methods generally leverage either real system samples or virtual data simulated by LLMs. However, the former is costly due to reliance on hand-crafted user requests, while the latter suffers from distribution shift from the real tools in the wild. Additionally, both methods lack negative samples tailored to intent deviation scenarios, hindering effective guidance on preference learning. We introduce RISE, a "Real-to-Virtual" method designed to mitigate intent deviation. Anchoring on verified tool primitives, RISE synthesizes virtual trajectories and generates diverse negative samples through mutation on critical parameters. With synthetic data, RISE fine-tunes backbone LLMs via the two-stage training for intent alignment. Evaluation results demonstrate that data synthesized by RISE achieve promising results in eight metrics covering user requires, execution trajectories and agent responses. Integrating with training, RISE achieves an average 35.28% improvement in Acctask (task completion) and 23.27% in Accintent (intent alignment), outperforming SOTA baselines by 1.20--42.09% and 1.17--54.93% respectively.

Yujia Zheng, Tianhao Li, Haotian Huang et al.

Prompt-based adversarial attacks have become an effective means to assess the robustness of large language models (LLMs). However, existing approaches often treat prompts as monolithic text, overlooking their structural heterogeneity-different prompt components contribute unequally to adversarial robustness. Prior works like PromptRobust assume prompts are value-neutral, but our analysis reveals that complex, domain-specific prompts with rich structures have components with differing vulnerabilities. To address this gap, we introduce PromptAnatomy, an automated framework that dissects prompts into functional components and generates diverse, interpretable adversarial examples by selectively perturbing each component using our proposed method, ComPerturb. To ensure linguistic plausibility and mitigate distribution shifts, we further incorporate a perplexity (PPL)-based filtering mechanism. As a complementary resource, we annotate four public instruction-tuning datasets using the PromptAnatomy framework, verified through human review. Extensive experiments across these datasets and five advanced LLMs demonstrate that ComPerturb achieves state-of-the-art attack success rates. Ablation studies validate the complementary benefits of prompt dissection and PPL filtering. Our results underscore the importance of prompt structure awareness and controlled perturbation for reliable adversarial robustness evaluation in LLMs. Code and data are available at https://github.com/Yujiaaaaa/PACP.

Awais Naeem, Tianhao Li, Huang-Ru Liao et al.

Accurate diagnosis and prognosis assisted by pathology images are essential for cancer treatment selection and planning. Despite the recent trend of adopting deep-learning approaches for analyzing complex pathology images, they fall short as they often overlook the domain-expert understanding of tissue structure and cell composition. In this work, we focus on a challenging Open-ended Pathology VQA (PathVQA-Open) task and propose a novel framework named Path-RAG, which leverages HistoCartography to retrieve relevant domain knowledge from pathology images and significantly improves performance on PathVQA-Open. Admitting the complexity of pathology image analysis, Path-RAG adopts a human-centered AI approach by retrieving domain knowledge using HistoCartography to select the relevant patches from pathology images. Our experiments suggest that domain guidance can significantly boost the accuracy of LLaVA-Med from 38% to 47%, with a notable gain of 28% for H&E-stained pathology images in the PathVQA-Open dataset. For longer-form question and answer pairs, our model consistently achieves significant improvements of 32.5% in ARCH-Open PubMed and 30.6% in ARCH-Open Books on H\&E images. Our code and dataset is available here (https://github.com/embedded-robotics/path-rag).

Tianhao Li, Tingfa Xu, Ying Wang et al.

Drone-based multi-object tracking is essential yet highly challenging due to small targets, severe occlusions, and cluttered backgrounds. Existing RGB-based tracking algorithms heavily depend on spatial appearance cues such as color and texture, which often degrade in aerial views, compromising reliability. Multispectral imagery, capturing pixel-level spectral reflectance, provides crucial cues that enhance object discriminability under degraded spatial conditions. However, the lack of dedicated multispectral UAV datasets has hindered progress in this domain. To bridge this gap, we introduce MMOT, the first challenging benchmark for drone-based multispectral multi-object tracking. It features three key characteristics: (i) Large Scale - 125 video sequences with over 488.8K annotations across eight categories; (ii) Comprehensive Challenges - covering diverse conditions such as extreme small targets, high-density scenarios, severe occlusions, and complex motion; and (iii) Precise Oriented Annotations - enabling accurate localization and reduced ambiguity under aerial perspectives. To better extract spectral features and leverage oriented annotations, we further present a multispectral and orientation-aware MOT scheme adapting existing methods, featuring: (i) a lightweight Spectral 3D-Stem integrating spectral features while preserving compatibility with RGB pretraining; (ii) an orientation-aware Kalman filter for precise state estimation; and (iii) an end-to-end orientation-adaptive transformer. Extensive experiments across representative trackers consistently show that multispectral input markedly improves tracking performance over RGB baselines, particularly for small and densely packed objects. We believe our work will advance drone-based multispectral multi-object tracking research. Our MMOT, code, and benchmarks are publicly available at https://github.com/Annzstbl/MMOT.

Tao Feng, Tingfa Xu, Haolin Qin et al.

Visual object tracking in real-world scenarios presents numerous challenges including occlusion, interference from similar objects and complex backgrounds-all of which limit the effectiveness of RGB-based trackers. Multispectral imagery, which captures pixel-level spectral reflectance, enhances target discriminability. However, the availability of multispectral tracking datasets remains limited. To bridge this gap, we introduce MSITrack, the largest and most diverse multispectral single object tracking dataset to date. MSITrack offers the following key features: (i) More Challenging Attributes-including interference from similar objects and similarity in color and texture between targets and backgrounds in natural scenarios, along with a wide range of real-world tracking challenges; (ii) Richer and More Natural Scenes-spanning 55 object categories and 300 distinct natural scenes, MSITrack far exceeds the scope of existing benchmarks. Many of these scenes and categories are introduced to the multispectral tracking domain for the first time; (iii) Larger Scale-300 videos comprising over 129k frames of multispectral imagery. To ensure annotation precision, each frame has undergone meticulous processing, manual labeling and multi-stage verification. Extensive evaluations using representative trackers demonstrate that the multispectral data in MSITrack significantly improves performance over RGB-only baselines, highlighting its potential to drive future advancements in the field. The MSITrack dataset is publicly available at: https://github.com/Fengtao191/MSITrack.

Zhenzhi Wang, Limin Wang, Tao Wu et al.

Temporal grounding aims to localize a video moment which is semantically aligned with a given natural language query. Existing methods typically apply a detection or regression pipeline on the fused representation with the research focus on designing complicated prediction heads or fusion strategies. Instead, from a perspective on temporal grounding as a metric-learning problem, we present a Mutual Matching Network (MMN), to directly model the similarity between language queries and video moments in a joint embedding space. This new metric-learning framework enables fully exploiting negative samples from two new aspects: constructing negative cross-modal pairs in a mutual matching scheme and mining negative pairs across different videos. These new negative samples could enhance the joint representation learning of two modalities via cross-modal mutual matching to maximize their mutual information. Experiments show that our MMN achieves highly competitive performance compared with the state-of-the-art methods on four video grounding benchmarks. Based on MMN, we present a winner solution for the HC-STVG challenge of the 3rd PIC workshop. This suggests that metric learning is still a promising method for temporal grounding via capturing the essential cross-modal correlation in a joint embedding space. Code is available at https://github.com/MCG-NJU/MMN.

Tianhao Li, Limin Wang, Gangshan Wu

Deep learning has achieved remarkable progress for visual recognition on large-scale balanced datasets but still performs poorly on real-world long-tailed data. Previous methods often adopt class re-balanced training strategies to effectively alleviate the imbalance issue, but might be a risk of over-fitting tail classes. The recent decoupling method overcomes over-fitting issues by using a multi-stage training scheme, yet, it is still incapable of capturing tail class information in the feature learning stage. In this paper, we show that soft label can serve as a powerful solution to incorporate label correlation into a multi-stage training scheme for long-tailed recognition. The intrinsic relation between classes embodied by soft labels turns out to be helpful for long-tailed recognition by transferring knowledge from head to tail classes. Specifically, we propose a conceptually simple yet particularly effective multi-stage training scheme, termed as Self Supervised to Distillation (SSD). This scheme is composed of two parts. First, we introduce a self-distillation framework for long-tailed recognition, which can mine the label relation automatically. Second, we present a new distillation label generation module guided by self-supervision. The distilled labels integrate information from both label and data domains that can model long-tailed distribution effectively. We conduct extensive experiments and our method achieves the state-of-the-art results on three long-tailed recognition benchmarks: ImageNet-LT, CIFAR100-LT and iNaturalist 2018. Our SSD outperforms the strong LWS baseline by from $2.7\%$ to $4.5\%$ on various datasets. The code is available at https://github.com/MCG-NJU/SSD-LT.

Tianhao Li, Limin Wang

Current video representations heavily rely on learning from manually annotated video datasets which are time-consuming and expensive to acquire. We observe videos are naturally accompanied by abundant text information such as YouTube titles and Instagram captions. In this paper, we leverage this visual-textual connection to learn spatiotemporal features in an efficient weakly-supervised manner. We present a general cross-modal pair discrimination (CPD) framework to capture this correlation between a video and its associated text. Specifically, we adopt noise-contrastive estimation to tackle the computational issue imposed by the huge amount of pair instance classes and design a practical curriculum learning strategy. We train our CPD models on both standard video dataset (Kinetics-210k) and uncurated web video dataset (Instagram-300k) to demonstrate its effectiveness. Without further fine-tuning, the learnt models obtain competitive results for action classification on Kinetics under the linear classification protocol. Moreover, our visual model provides an effective initialization to fine-tune on downstream tasks, which yields a remarkable performance gain for action recognition on UCF101 and HMDB51, compared with the existing state-of-the-art self-supervised training methods. In addition, our CPD model yields a new state of the art for zero-shot action recognition on UCF101 by directly utilizing the learnt visual-textual embeddings. The code will be made available at https://github.com/MCG-NJU/CPD-Video.

Tienyu Chang, Tianhao Li, Ruogu Fang et al.

Optical coherence tomography (OCT), a commonly used retinal imaging modality, plays a central role in retinal disease diagnosis by providing high-resolution visualization of retinal layers. While deep learning (DL) has achieved expert-level accuracy in OCT-based retinal disease detection, its "black box" nature poses challenges for clinical adoption, where explainability is essential for clinical trust and regulatory approval. Existing post-hoc explainable AI (XAI) methods often struggle to delineate fine-grained lesion structures, respect anatomical boundaries, or suppress noise, limiting the trustworthiness of their explanations. To bridge these gaps, we propose a Structure-Aware Interpretable Learning (SAIL) framework that integrates retinal anatomical priors at the representation level and couples them with semantic features via a fusion design. Without modifying standard post-hoc explainability methods, this representation yields sharper and more anatomically aligned attribution maps. Comprehensive experiments on diverse OCT datasets demonstrate that our structure-aware method consistently enhances interpretability, producing clinically meaningful and anatomy-aware explanations. Ablation studies further show that strong interpretability requires both structural priors and semantic features, and that properly fusing the two is critical to achieve the best explanation quality. Together, these results highlight structure-aware representations as a key step toward reliable explainability in OCT.

Zaid Nasser, Mikhail Iumanov, Tianhao Li et al.

We present KM-ViPE (Knowledge Mapping Video Pose Engine), a real-time open-vocabulary SLAM framework for uncalibrated monocular cameras in dynamic environments. Unlike systems requiring depth sensors and offline calibration, KM-ViPE operates directly on raw RGB streams, making it ideal for ego-centric applications and harvesting internet-scale video data for training. KM-ViPE tightly couples DINO visual features with geometric constraints through a high-level features based adaptive robust kernel that handles both moving objects and movable static objects (e.g., moving furniture in ego-centric views). The system performs simultaneous online localization and open-vocabulary semantic mapping by fusing geometric and deep visual features aligned with language embeddings. Our results are competitive with state-of-the-art approaches, while existing solutions either operate offline, need depth data and/or odometry estimation, or lack dynamic scene robustness. KM-ViPE benefits from internet-scale training and uniquely combines online operation, uncalibrated monocular input, and robust handling of dynamic scenes, which makes it a good fit for autonomous robotics and AR/VR applications and advances practical spatial intelligence capabilities for embodied AI.

Shaona Ghosh, Heather Frase, Adina Williams et al. · deepmind, stanford

The rapid advancement and deployment of AI systems have created an urgent need for standard safety-evaluation frameworks. This paper introduces AILuminate v1.0, the first comprehensive industry-standard benchmark for assessing AI-product risk and reliability. Its development employed an open process that included participants from multiple fields. The benchmark evaluates an AI system's resistance to prompts designed to elicit dangerous, illegal, or undesirable behavior in 12 hazard categories, including violent crimes, nonviolent crimes, sex-related crimes, child sexual exploitation, indiscriminate weapons, suicide and self-harm, intellectual property, privacy, defamation, hate, sexual content, and specialized advice (election, financial, health, legal). Our method incorporates a complete assessment standard, extensive prompt datasets, a novel evaluation framework, a grading and reporting system, and the technical as well as organizational infrastructure for long-term support and evolution. In particular, the benchmark employs an understandable five-tier grading scale (Poor to Excellent) and incorporates an innovative entropy-based system-response evaluation. In addition to unveiling the benchmark, this report also identifies limitations of our method and of building safety benchmarks generally, including evaluator uncertainty and the constraints of single-turn interactions. This work represents a crucial step toward establishing global standards for AI risk and reliability evaluation while acknowledging the need for continued development in areas such as multiturn interactions, multimodal understanding, coverage of additional languages, and emerging hazard categories. Our findings provide valuable insights for model developers, system integrators, and policymakers working to promote safer AI deployment.

Zaid Nasser, Mikhail Iumanov, Tianhao Li et al.

We present RADIO-ViPE (Reduce All Domains Into One -- Video Pose Engine), an online semantic SLAM system that enables geometry-aware open-vocabulary grounding, associating arbitrary natural language queries with localized 3D regions and objects in dynamic environments. Unlike existing approaches that require calibrated, posed RGB-D input, RADIO-ViPE operates directly on raw monocular RGB video streams, requiring no prior camera intrinsics, depth sensors, or pose initialization. The system tightly couples multi-modal embeddings -- spanning vision and language -- derived from agglomerative foundation models (e.g., RADIO) with geometric scene information. This coupling takes place in initialization, optimization and factor graph connections to improve the consistency of the map from multiple modalities. The optimization is wrapped within adaptive robust kernels, designed to handle both actively moving objects and agent-displaced scene elements (e.g., furniture rearranged during ego-centric session). Experiments demonstrate that RADIO-ViPE achieves state-of-the-art results on the dynamic TUM-RGBD benchmark while maintaining competitive performance against offline open-vocabulary methods that rely on calibrated data and static scene assumptions. RADIO-ViPE bridges a critical gap in real-world deployment, enabling robust open-vocabulary semantic grounding for autonomous robotics and unconstrained in-the-wild video streams. Project page: https://be2rlab.github.io/radio_vipe

Tianhao Li, Tianyu Zeng, Yujia Zheng et al.

Deep learning-based medical image segmentation models, such as U-Net, rely on high-quality annotated datasets to achieve accurate predictions. However, the increasing use of generative models for synthetic data augmentation introduces potential risks, particularly in the absence of rigorous quality control. In this paper, we investigate the impact of synthetic MRI data on the robustness and segmentation accuracy of U-Net models for brain tumor segmentation. Specifically, we generate synthetic T1-contrast-enhanced (T1-Ce) MRI scans using a GAN-based model with a shared encoding-decoding framework and shortest-path regularization. To quantify the effect of synthetic data contamination, we train U-Net models on progressively "poisoned" datasets, where synthetic data proportions range from 16.67% to 83.33%. Experimental results on a real MRI validation set reveal a significant performance degradation as synthetic data increases, with Dice coefficients dropping from 0.8937 (33.33% synthetic) to 0.7474 (83.33% synthetic). Accuracy and sensitivity exhibit similar downward trends, demonstrating the detrimental effect of synthetic data on segmentation robustness. These findings underscore the importance of quality control in synthetic data integration and highlight the risks of unregulated synthetic augmentation in medical image analysis. Our study provides critical insights for the development of more reliable and trustworthy AI-driven medical imaging systems.

Qian Xiong, Yuekai Huang, Ziyou Jiang et al.

The emergence of the tool agent paradigm has broadened the capability boundaries of the Large Language Model (LLM), enabling it to complete more complex tasks. However, the effectiveness of this paradigm is limited due to the issue of parameter failure during its execution. To explore this phenomenon and propose corresponding suggestions, we first construct a parameter failure taxonomy in this paper. We derive five failure categories from the invocation chain of a mainstream tool agent. Then, we explore the correlation between three different input sources and failure categories by applying 15 input perturbation methods to the input. Experimental results show that parameter name hallucination failure primarily stems from inherent LLM limitations, while issues with input sources mainly cause other failure patterns. To improve the reliability and effectiveness of tool-agent interactions, we propose corresponding improvement suggestions, including standardizing tool return formats, improving error feedback mechanisms, and ensuring parameter consistency.

Kyle Cox, Jiawei Xu, Yikun Han et al.

An interesting behavior in large language models (LLMs) is prompt sensitivity. When provided with different but semantically equivalent versions of the same prompt, models may produce very different distributions of answers. This suggests that the uncertainty reflected in a model's output distribution for one prompt may not reflect the model's uncertainty about the meaning of the prompt. We model prompt sensitivity as a type of generalization error, and show that sampling across the semantic ``concept space'' with paraphrasing perturbations improves uncertainty calibration without compromising accuracy. Additionally, we introduce a new metric for uncertainty decomposition in black-box LLMs that improves upon entropy-based decomposition by modeling semantic continuities in natural language generation. We show that this decomposition metric can be used to quantify how much LLM uncertainty is attributed to prompt sensitivity. Our work introduces a new way to improve uncertainty calibration in prompt-sensitive language models, and provides evidence that some LLMs fail to exhibit consistent general reasoning about the meanings of their inputs.

Linjuan Wu, Haoran Wei, Huan Lin et al.

Large language models (LLMs) exhibit remarkable multilingual capabilities despite English-dominated pre-training, attributed to cross-lingual mechanisms during pre-training. Existing methods for enhancing cross-lingual transfer remain constrained by parallel resources, suffering from limited linguistic and domain coverage. We propose Cross-lingual In-context Pre-training (CrossIC-PT), a simple and scalable approach that enhances cross-lingual transfer by leveraging semantically related bilingual texts via simple next-word prediction. We construct CrossIC-PT samples by interleaving semantic-related bilingual Wikipedia documents into a single context window. To access window size constraints, we implement a systematic segmentation policy to split long bilingual document pairs into chunks while adjusting the sliding window mechanism to preserve contextual coherence. We further extend data availability through a semantic retrieval framework to construct CrossIC-PT samples from web-crawled corpus. Experimental results demonstrate that CrossIC-PT improves multilingual performance on three models (Llama-3.1-8B, Qwen2.5-7B, and Qwen2.5-1.5B) across six target languages, yielding performance gains of 3.79%, 3.99%, and 1.95%, respectively, with additional improvements after data augmentation.

Yuyao Ge, Lingrui Mei, Zenghao Duan et al.

The advancement of large language models (LLMs) has catalyzed a paradigm shift from code generation assistance to autonomous coding agents, enabling a novel development methodology termed "Vibe Coding" where developers validate AI-generated implementations through outcome observation rather than line-by-line code comprehension. Despite its transformative potential, the effectiveness of this emergent paradigm remains under-explored, with empirical evidence revealing unexpected productivity losses and fundamental challenges in human-AI collaboration. To address this gap, this survey provides the first comprehensive and systematic review of Vibe Coding with large language models, establishing both theoretical foundations and practical frameworks for this transformative development approach. Drawing from systematic analysis of over 1000 research papers, we survey the entire vibe coding ecosystem, examining critical infrastructure components including LLMs for coding, LLM-based coding agent, development environment of coding agent, and feedback mechanisms. We first introduce Vibe Coding as a formal discipline by formalizing it through a Constrained Markov Decision Process that captures the dynamic triadic relationship among human developers, software projects, and coding agents. Building upon this theoretical foundation, we then synthesize existing practices into five distinct development models: Unconstrained Automation, Iterative Conversational Collaboration, Planning-Driven, Test-Driven, and Context-Enhanced Models, thus providing the first comprehensive taxonomy in this domain. Critically, our analysis reveals that successful Vibe Coding depends not merely on agent capabilities but on systematic context engineering, well-established development environments, and human-agent collaborative development models.

Tianhao Li, Yang Li, Mengtian Li et al.

Accurately perceiving dynamic environments is a fundamental task for autonomous driving and robotic systems. Existing methods inadequately utilize temporal information, relying mainly on local temporal interactions between adjacent frames and failing to leverage global sequence information effectively. To address this limitation, we investigate how to effectively aggregate global temporal features from temporal sequences, aiming to achieve occupancy representations that efficiently utilize global temporal information from historical observations. For this purpose, we propose a global temporal aggregation denoising network named GTAD, introducing a global temporal information aggregation framework as a new paradigm for holistic 3D scene understanding. Our method employs an in-model latent denoising network to aggregate local temporal features from the current moment and global temporal features from historical sequences. This approach enables the effective perception of both fine-grained temporal information from adjacent frames and global temporal patterns from historical observations. As a result, it provides a more coherent and comprehensive understanding of the environment. Extensive experiments on the nuScenes and Occ3D-nuScenes benchmark and ablation studies demonstrate the superiority of our method.

Jinhao Duan, Xinyu Zhao, Zhuoxuan Zhang et al.

Although Large Language Models (LLMs) succeed in human-guided conversations such as instruction following and question answering, the potential of LLM-guided conversations-where LLMs direct the discourse and steer the conversation's objectives-remains under-explored. In this study, we first characterize LLM-guided conversation into three fundamental components: (i) Goal Navigation; (ii) Context Management; (iii) Empathetic Engagement, and propose GuideLLM as an installation. We then implement an interviewing environment for the evaluation of LLM-guided conversation. Specifically, various topics are involved in this environment for comprehensive interviewing evaluation, resulting in around 1.4k turns of utterances, 184k tokens, and over 200 events mentioned during the interviewing for each chatbot evaluation. We compare GuideLLM with 6 state-of-the-art LLMs such as GPT-4o and Llama-3-70b-Instruct, from the perspective of interviewing quality, and autobiography generation quality. For automatic evaluation, we derive user proxies from multiple autobiographies and employ LLM-as-a-judge to score LLM behaviors. We further conduct a human-involved experiment by employing 45 human participants to chat with GuideLLM and baselines. We then collect human feedback, preferences, and ratings regarding the qualities of conversation and autobiography. Experimental results indicate that GuideLLM significantly outperforms baseline LLMs in automatic evaluation and achieves consistent leading performances in human ratings.

Tianhao Li, Shangjie Li, Binbin Xie et al.

The advent of large language models (LLMs) has predominantly catered to high-resource languages, leaving a disparity in performance for low-resource languages. Conventional Continual Training (CT) approaches to bridge this gap often undermine a model's original linguistic proficiency when expanding to multilingual contexts. Addressing this issue, we introduce a novel MoE-CT architecture, a paradigm that innovatively separates the base model's learning from the multilingual expansion process. Our design freezes the original LLM parameters, thus safeguarding its performance in high-resource languages, while an appended MoE module, trained on diverse language datasets, augments low-resource language proficiency. Our approach significantly outperforms conventional CT methods, as evidenced by our experiments, which show marked improvements in multilingual benchmarks without sacrificing the model's original language performance. Moreover, our MoE-CT framework demonstrates enhanced resistance to forgetting and superior transfer learning capabilities. By preserving the base model's integrity and focusing on strategic parameter expansion, our methodology advances multilingual language modeling and represents a significant step forward for low-resource language inclusion in LLMs, indicating a fruitful direction for future research in language technologies.

Mingquan Lin, Tianhao Li, Zhaoyi Sun et al.

Purpose: Limited studies exploring concrete methods or approaches to tackle and enhance model fairness in the radiology domain. Our proposed AI model utilizes supervised contrastive learning to minimize bias in CXR diagnosis. Materials and Methods: In this retrospective study, we evaluated our proposed method on two datasets: the Medical Imaging and Data Resource Center (MIDRC) dataset with 77,887 CXR images from 27,796 patients collected as of April 20, 2023 for COVID-19 diagnosis, and the NIH Chest X-ray (NIH-CXR) dataset with 112,120 CXR images from 30,805 patients collected between 1992 and 2015. In the NIH-CXR dataset, thoracic abnormalities include atelectasis, cardiomegaly, effusion, infiltration, mass, nodule, pneumonia, pneumothorax, consolidation, edema, emphysema, fibrosis, pleural thickening, or hernia. Our proposed method utilizes supervised contrastive learning with carefully selected positive and negative samples to generate fair image embeddings, which are fine-tuned for subsequent tasks to reduce bias in chest X-ray (CXR) diagnosis. We evaluated the methods using the marginal AUC difference ($δ$ mAUC). Results: The proposed model showed a significant decrease in bias across all subgroups when compared to the baseline models, as evidenced by a paired T-test (p<0.0001). The $δ$ mAUC obtained by our method were 0.0116 (95\% CI, 0.0110-0.0123), 0.2102 (95% CI, 0.2087-0.2118), and 0.1000 (95\% CI, 0.0988-0.1011) for sex, race, and age on MIDRC, and 0.0090 (95\% CI, 0.0082-0.0097) for sex and 0.0512 (95% CI, 0.0512-0.0532) for age on NIH-CXR, respectively. Conclusion: Employing supervised contrastive learning can mitigate bias in CXR diagnosis, addressing concerns of fairness and reliability in deep learning-based diagnostic methods.

Ajay Jaiswal, Tianhao Li, Cyprian Zander et al.

Computer-aided diagnosis plays a salient role in more accessible and accurate cardiopulmonary diseases classification and localization on chest radiography. Millions of people get affected and die due to these diseases without an accurate and timely diagnosis. Recently proposed contrastive learning heavily relies on data augmentation, especially positive data augmentation. However, generating clinically-accurate data augmentations for medical images is extremely difficult because the common data augmentation methods in computer vision, such as sharp, blur, and crop operations, can severely alter the clinical settings of medical images. In this paper, we proposed a novel and simple data augmentation method based on patient metadata and supervised knowledge to create clinically accurate positive and negative augmentations for chest X-rays. We introduce an end-to-end framework, SCALP, which extends the self-supervised contrastive approach to a supervised setting. Specifically, SCALP pulls together chest X-rays from the same patient (positive keys) and pushes apart chest X-rays from different patients (negative keys). In addition, it uses ResNet-50 along with the triplet-attention mechanism to identify cardiopulmonary diseases, and Grad-CAM++ to highlight the abnormal regions. Our extensive experiments demonstrate that SCALP outperforms existing baselines with significant margins in both classification and localization tasks. Specifically, the average classification AUCs improve from 82.8% (SOTA using DenseNet-121) to 83.9% (SCALP using ResNet-50), while the localization results improve on average by 3.7% over different IoU thresholds.

Wei Wang, Weiqing Min, Tianhao Li et al.

Background: Maintaining a healthy diet is vital to avoid health-related issues, e.g., undernutrition, obesity and many non-communicable diseases. An indispensable part of the health diet is dietary assessment. Traditional manual recording methods are not only burdensome but time-consuming, and contain substantial biases and errors. Recent advances in Artificial Intelligence (AI), especially computer vision technologies, have made it possible to develop automatic dietary assessment solutions, which are more convenient, less time-consuming and even more accurate to monitor daily food intake. Scope and approach: This review presents Vision-Based Dietary Assessment (VBDA) architectures, including multi-stage architecture and end-to-end one. The multi-stage dietary assessment generally consists of three stages: food image analysis, volume estimation and nutrient derivation. The prosperity of deep learning makes VBDA gradually move to an end-to-end implementation, which applies food images to a single network to directly estimate the nutrition. The recently proposed end-to-end methods are also discussed. We further analyze existing dietary assessment datasets, indicating that one large-scale benchmark is urgently needed, and finally highlight critical challenges and future trends for VBDA. Key findings and conclusions: After thorough exploration, we find that multi-task end-to-end deep learning approaches are one important trend of VBDA. Despite considerable research progress, many challenges remain for VBDA due to the meal complexity. We also provide the latest ideas for future development of VBDA, e.g., fine-grained food analysis and accurate volume estimation. This review aims to encourage researchers to propose more practical solutions for VBDA.

Zhishun Wang, Wei Lu, Kaixin Zhang et al.

It is a difficult task for both professional investors and individual traders continuously making profit in stock market. With the development of computer science and deep reinforcement learning, Buy\&Hold (B\&H) has been oversteped by many artificial intelligence trading algorithms. However, the information and process are not enough, which limit the performance of reinforcement learning algorithms. Thus, we propose a parallel-network continuous quantitative trading model with GARCH and PPO to enrich the basical deep reinforcement learning model, where the deep learning parallel network layers deal with 3 different frequencies data (including GARCH information) and proximal policy optimization (PPO) algorithm interacts actions and rewards with stock trading environment. Experiments in 5 stocks from Chinese stock market show our method achieves more extra profit comparing with basical reinforcement learning methods and bench models.

Wei Peng, Chongxuan Huang, Tianhao Li et al.

Existing data augmentation approaches for neural machine translation (NMT) have predominantly relied on back-translating in-domain (IND) monolingual corpora. These methods suffer from issues associated with a domain information gap, which leads to translation errors for low frequency and out-of-vocabulary terminology. This paper proposes a dictionary-based data augmentation (DDA) method for cross-domain NMT. DDA synthesizes a domain-specific dictionary with general domain corpora to automatically generate a large-scale pseudo-IND parallel corpus. The generated pseudo-IND data can be used to enhance a general domain trained baseline. The experiments show that the DDA-enhanced NMT models demonstrate consistent significant improvements, outperforming the baseline models by 3.75-11.53 BLEU. The proposed method is also able to further improve the performance of the back-translation based and IND-finetuned NMT models. The improvement is associated with the enhanced domain coverage produced by DDA.