Long Zhang

Kesha Ou, Zhen Tian, Wayne Xin Zhao et al.

Click-through rate (CTR) prediction holds a pivotal role in online advertising and recommendation systems, where even small improvements can significantly boost revenue. Existing research primarily focuses on designing dual-stream architectures to capture effective complex feature interactions from both explicit and implicit perspectives. However, these approaches are faced with two major challenges: 1) the high complexity of feature interaction learning, which increases computational demands and the overfitting risk, and 2) the imbalance between explicit and implicit modules, where one module's output may dominate the final prediction. To address these issues, in this paper, we propose Dual-Stream MLP (DS-MLP), a novel feature interaction framework for the CTR prediction task. Specially, it leverages knowledge distillation to consolidate the capacity of learning explicit feature interaction into a main MLP network, while a parallel MLP simultaneously captures implicit feature interactions as a complement. To effectively optimize the dual-stream MLP architecture, we further design a specific learning approach with two alignment strategies for enhancing the compatibility of the two MLP components. Experiments demonstrate that DS-MLP, though merely a vanilla MLP structure (the final model), can achieve state-of-the-art performance across three widely used benchmarks, offering a scalable and efficient solution for large-scale recommendation systems. Our code is available at https://github.com/RUCAIBox/DS-MLP.

Long Zhang, Yizhe Zhao, Halvin Yang et al.

Integrated data and energy transfer (IDET) is a promising technique for supporting sustainable low-power wireless networks. To improve both communication reliability and energy transfer efficiency, this paper investigates a fluid index modulation (FIM) assisted IDET system, where the base station employs a two-dimensional fluid antenna system (FAS) and the receiver adopts a power-splitting architecture. In FIM, the information bits are delivered not only from the modulation symbols, but also the index of antenna position. Under finite-alphabet signaling, the average harvested power, bit error rate (BER), and achievable data rate are derived in closed form. A joint optimization problem is formulated to maximize the average harvested power subject to BER and achievable rate constraints by jointly optimizing the port selection, precoding vector, and power splitting ratio. An alternating optimization framework is developed, where the precoding vector and port selection are obtained via a Riemannian augmented Lagrangian method (RALM) and block coordinate descent (BCD) algorithm, respectively. Simulation results demonstrate that the proposed scheme achieves a superior rate-energy trade-off over benchmark schemes, while the proposed algorithm attains near-optimal performance with significantly lower complexity than exhaustive search.

Bingqian Li, Bowen Zheng, Xiaolei Wang et al.

Large language models (LLMs) have shown great promise in recommender systems, where supervised fine-tuning (SFT) is commonly used for adaptation. Subsequent studies further introduce preference learning to incorporate negative samples into the training process. However, existing methods rely on sequence-level, offline-generated negatives, making them less discriminative and informative when adapting LLMs to recommendation tasks with large negative item spaces. To address these challenges, we propose ILRec, a novel preference fine-tuning framework for LLM-based recommendation, leveraging self-hard negative signals extracted from intermediate layers to improve preference learning. Specifically, we identify self-hard negative tokens from intermediate layers as fine-grained negative supervision that dynamically reflects the model's preference learning process. To effectively integrate these signals into training, we design a two-stage framework comprising cross-layer preference optimization and cross-layer preference distillation, enabling the model to jointly discriminate informative negatives and enhance the quality of negative signals from intermediate layers. In addition, we introduce a lightweight collaborative filtering model to assign token-level rewards for negative signals, mitigating the risk of over-penalizing false negatives. Extensive experiments on three datasets demonstrate ILRec's effectiveness in enhancing the performance of LLM-based recommender systems.

Bingqian Li, Xiaolei Wang, Junyi Li et al.

Agentic recommender systems leverage Large Language Models (LLMs) to model complex user behaviors and support personalized decision-making. However, existing methods primarily model preference changes based on explicit user-item interactions, which are sparse, noisy, and unable to reflect the real-time, mutual influences among users and items. To address these limitations, we propose RecNet, a self-evolving preference propagation framework that proactively propagates real-time preference updates across related users and items. RecNet consists of two complementary phases. In the forward phase, the centralized preference routing mechanism leverages router agents to integrate preference updates and dynamically propagate them to the most relevant agents. To ensure accurate and personalized integration of propagated preferences, we further introduce a personalized preference reception mechanism, which combines a message buffer for temporary caching and an optimizable, rule-based filter memory to guide selective preference assimilation based on past experience and interests. In the backward phase, the feedback-driven propagation optimization mechanism simulates a multi-agent reinforcement learning framework, using LLMs for credit assignment, gradient analysis, and module-level optimization, enabling continuous self-evolution of propagation strategies. Extensive experiments on various scenarios demonstrate the effectiveness of RecNet in modeling preference propagation for recommender systems.

Jiayu Lin, Guanrong Chen, Bojun Jin et al.

In this paper we present the results of the AI-Debater 2023 Challenge held by the Chinese Conference on Affect Computing (CCAC 2023), and introduce the related datasets. We organize two tracks to handle the argumentative generation tasks in different scenarios, namely, Counter-Argument Generation (Track 1) and Claim-based Argument Generation (Track 2). Each track is equipped with its distinct dataset and baseline model respectively. In total, 32 competing teams register for the challenge, from which we received 11 successful submissions. In this paper, we will present the results of the challenge and a summary of the systems, highlighting commonalities and innovations among participating systems. Datasets and baseline models of the AI-Debater 2023 Challenge have been already released and can be accessed through the official website of the challenge.

Hao Jiang, Long Zhang, Guoquan Wang et al.

Local-life recommendation have witnessed rapid growth, providing users with convenient access to daily essentials. However, this domain faces two key challenges: (1) spatial constraints, driven by the requirements of the local-life scenario, where items are usually shown only to users within a limited geographic area, indirectly reducing their exposure probability; and (2) long-tail sparsity, where few popular items dominate user interactions, while many high-quality long-tail items are largely overlooked due to imbalanced interaction opportunities. Existing methods typically adopt a user-centric perspective, such as modeling spatial user preferences or enhancing long-tail representations with collaborative filtering signals. However, we argue that an item-centric perspective is more suitable for this domain, focusing on enhancing long-tail items representation that align with the spatially-constrained characteristics of local lifestyle services. To tackle this issue, we propose ReST, a Plug-And-Play Spatially-Constrained Representation Enhancement Framework for Long-Tail Local-Life Recommendation. Specifically, we first introduce a Meta ID Warm-up Network, which initializes fundamental ID representations by injecting their basic attribute-level semantic information. Subsequently, we propose a novel Spatially-Constrained ID Representation Enhancement Network (SIDENet) based on contrastive learning, which incorporates two efficient strategies: a spatially-constrained hard sampling strategy and a dynamic representation alignment strategy. This design adaptively identifies weak ID representations based on their attribute-level information during training. It additionally enhances them by capturing latent item relationships within the spatially-constrained characteristics of local lifestyle services, while preserving compatibility with popular items.

Yuxuan Chen, Long Zhang, Xu Zhu et al.

Merging natural language interfaces with computational fluid dynamics (CFD) workflows presents transformative opportunities for both industry and research. In this study, we introduce OptMetaOpenFOAM - a novel framework that bridges MetaOpenFOAM with external analysis and optimization tool libraries through a large language model (LLM)-driven chain-of-thought (COT) methodology. By automating complex CFD tasks via natural language inputs, the framework empowers non-expert users to perform sensitivity analyses and parameter optimizations with markedly improved efficiency. The test dataset comprises 11 distinct CFD analysis or optimization tasks, including a baseline simulation task derived from an OpenFOAM tutorial covering fluid dynamics, combustion, and heat transfer. Results confirm that OptMetaOpenFOAM can accurately interpret user requirements expressed in natural language and effectively invoke external tool libraries alongside MetaOpenFOAM to complete the tasks. Furthermore, validation on a non-OpenFOAM tutorial case - namely, a hydrogen combustion chamber - demonstrates that a mere 200-character natural language input can trigger a sequence of simulation, postprocessing, analysis, and optimization tasks spanning over 2,000 lines of code. These findings underscore the transformative potential of LLM-driven COT methodologies in linking external tool for advanced analysis and optimization, positioning OptMetaOpenFOAM as an effective tool that streamlines CFD simulations and enhances their convenience and efficiency for both industrial and research applications. Code is available at https://github.com/Terry-cyx/MetaOpenFOAM.

Zijun Min, Bingshuai Liu, Ante Wang et al.

Reinforcement Learning with Verifiable Rewards (RLVR) offers a promising framework for optimizing large language models in reasoning tasks. However, existing RLVR algorithms focus on different granularities, and each has complementary strengths and limitations. Group Relative Policy Optimization (GRPO) updates the policy with token-level importance ratios, which preserves fine-grained credit assignment but often suffers from high variance and instability. In contrast, Group Sequence Policy Optimization (GSPO) applies single sequence-level importance ratios across all tokens in a response that better matches sequence-level rewards, but sacrifices token-wise credit assignment. In this paper, we propose Dynamic Hybrid Policy Optimization (DHPO) to bridge GRPO and GSPO within a single clipped surrogate objective. DHPO combines token-level and sequence-level importance ratios using weighting mechanisms. We explore two variants of the mixing mechanism, including an averaged mixing and an entropy-guided mixing. To further stabilize training, we employ a branch-specific clipping strategy that constrains token-level and sequence-level ratios within separate trust regions before mixing, preventing outliers in either branch from dominating the update. Across seven challenging mathematical reasoning benchmarks, experiments on both dense and MoE models from the Qwen3 series show that DHPO consistently outperforms GRPO and GSPO. We will release our code upon acceptance of this paper.

Yu Chen, Yifei Han, Long Zhang et al.

Fine-tuning large pre-trained models for downstream tasks has become a fundamental approach in natural language processing. Fully fine-tuning all model parameters is computationally expensive and memory-intensive, especially in resource-constrained environments. Existing parameter-efficient fine-tuning methods reduce the number of trainable parameters but typically overlook the varying sensitivity of different model layers and the importance of training data. In this work, we propose TsqLoRA, a novel method that integrates data-quality-driven selection with sensitivity-aware low-rank adaptation, consisted of two main components: a quality-aware sampling mechanism for selecting the most informative training data, and a dynamic rank allocation module that adjusts the rank of each layer based on its sensitivity to parameter updates. The experimental results demonstrate that TsqLoRA improves fine-tuning efficiency while maintaining or even improving performance on a variety of NLP tasks. Our code will be available at https://github.com/Benjamin-Ricky/TsqLoRA.

Jiaxun Yang, Yifei Han, Long Zhang et al.

Chinese Patronizing and Condescending Language (CPCL) is an implicitly discriminatory toxic speech targeting vulnerable groups on Chinese video platforms. The existing dataset lacks user comments, which are a direct reflection of video content. This undermines the model's understanding of video content and results in the failure to detect some CPLC videos. To make up for this loss, this research reconstructs a new dataset PCLMMPLUS that includes 103k comment entries and expands the dataset size. We also propose the CPCLDetector model with alignment selection and knowledge-enhanced comment content modules. Extensive experiments show the proposed CPCLDetector outperforms the SOTA on PCLMM and achieves higher performance on PCLMMPLUS . CPLC videos are detected more accurately, supporting content governance and protecting vulnerable groups. Code and dataset are available at https://github.com/jiaxunyang256/PCLD.

Haofeng Huang, Yifei Han, Long Zhang et al.

Multimodal intent recognition (MMIR) suffers from weak semantic grounding and poor robustness under noisy or rare-class conditions. We propose MVCL-DAF++, which extends MVCL-DAF with two key modules: (1) Prototype-aware contrastive alignment, aligning instances to class-level prototypes to enhance semantic consistency; and (2) Coarse-to-fine attention fusion, integrating global modality summaries with token-level features for hierarchical cross-modal interaction. On MIntRec and MIntRec2.0, MVCL-DAF++ achieves new state-of-the-art results, improving rare-class recognition by +1.05\% and +4.18\% WF1, respectively. These results demonstrate the effectiveness of prototype-guided learning and coarse-to-fine fusion for robust multimodal understanding. The source code is available at https://github.com/chr1s623/MVCL-DAF-PlusPlus.

Hongyao Tu, Liang Zhang, Yujie Lin et al.

The goal of open relation extraction (OpenRE) is to develop an RE model that can generalize to new relations not encountered during training. Existing studies primarily formulate OpenRE as a clustering task. They first cluster all test instances based on the similarity between the instances, and then manually assign a new relation to each cluster. However, their reliance on human annotation limits their practicality. In this paper, we propose an OpenRE framework based on large language models (LLMs), which directly predicts new relations for test instances by leveraging their strong language understanding and generation abilities, without human intervention. Specifically, our framework consists of two core components: (1) a relation discoverer (RD), designed to predict new relations for test instances based on \textit{demonstrations} formed by training instances with known relations; and (2) a relation predictor (RP), used to select the most likely relation for a test instance from $n$ candidate relations, guided by \textit{demonstrations} composed of their instances. To enhance the ability of our framework to predict new relations, we design a self-correcting inference strategy composed of three stages: relation discovery, relation denoising, and relation prediction. In the first stage, we use RD to preliminarily predict new relations for all test instances. Next, we apply RP to select some high-reliability test instances for each new relation from the prediction results of RD through a cross-validation method. During the third stage, we employ RP to re-predict the relations of all test instances based on the demonstrations constructed from these reliable test instances. Extensive experiments on three OpenRE datasets demonstrate the effectiveness of our framework. We release our code at https://github.com/XMUDeepLIT/LLM-OREF.git.

Fei Huang, Fan Wu, Zeqing Zhang et al.

GeoGPT is an open large language model system built to advance research in the geosciences. To enhance its domain-specific capabilities, we integrated Retrieval Augmented Generation(RAG), which augments model outputs with relevant information retrieved from an external knowledge source. GeoGPT uses RAG to draw from the GeoGPT Library, a specialized corpus curated for geoscientific content, enabling it to generate accurate, context-specific answers. Users can also create personalized knowledge bases by uploading their own publication lists, allowing GeoGPT to retrieve and respond using user-provided materials. To further improve retrieval quality and domain alignment, we fine-tuned both the embedding model and a ranking model that scores retrieved passages by relevance to the query. These enhancements optimize RAG for geoscience applications and significantly improve the system's ability to deliver precise and trustworthy outputs. GeoGPT reflects a strong commitment to open science through its emphasis on collaboration, transparency, and community driven development. As part of this commitment, we have open-sourced two core RAG components-GeoEmbedding and GeoReranker-to support geoscientists, researchers, and professionals worldwide with powerful, accessible AI tools.

Anxiang Zeng, Haibo Zhang, Kaixiang Mo et al.

Recent R1-Zero-like research further demonstrates that reasoning extension has given large language models (LLMs) unprecedented reasoning capabilities, and Reinforcement Learning is the core technology to elicit its complex reasoning. However, conducting RL experiments directly on hyperscale models involves high computational costs and resource demands, posing significant risks. We propose the Compass-Thinker-7B model, which aims to explore the potential of Reinforcement Learning with less computational resources and costs, and provides insights for further research into RL recipes for larger models. Compass-Thinker-7B is trained from an open source model through a specially designed Reinforcement Learning Pipeline. We curate a dataset of 30k verifiable mathematics problems for the Reinforcement Learning Pipeline. By configuring data and training settings with different difficulty distributions for different stages, the potential of the model is gradually released and the training efficiency is improved. Extensive evaluations show that Compass-Thinker-7B possesses exceptional reasoning potential, and achieves superior performance on mathematics compared to the same-sized RL model. Especially in the challenging AIME2024 evaluation, Compass-Thinker-7B achieves 40% accuracy.

Deepika Tiwari, Long Zhang, Martin Monperrus et al.

In this paper, we propose to use production executions to improve the quality of testing for certain methods of interest for developers. These methods can be methods that are not covered by the existing test suite, or methods that are poorly tested. We devise an approach called PANKTI which monitors applications as they execute in production, and then automatically generates differential unit tests, as well as derived oracles, from the collected data. PANKTI's monitoring and generation focuses on one single programming language, Java. We evaluate it on three real-world, open-source projects: a videoconferencing system, a PDF manipulation library, and an e-commerce application. We show that PANKTI is able to generate differential unit tests by monitoring target methods in production, and that the generated tests improve the quality of the test suite of the application under consideration.

Long Zhang, Deepika Tiwari, Brice Morin et al.

Docker is a virtualization technique heavily used in the industry to build cloud-based systems. In the context of Docker, a system is said to be observable if engineers can get accurate information about its running state in production. In this paper, we present a novel approach, called POBS, to automatically improve the observability of Dockerized Java applications. POBS is based on automated transformations of Docker configuration files. Our approach injects additional modules in the production application, in order to provide better observability. We evaluate POBS by applying it on open-source Java applications which are containerized with Docker. Our key result is that 148/170 (87%) of Docker Java containers can be automatically augmented with better observability.

Hanyi Hu, Long Zhang, Shuan Li et al.

Financial frauds cause billions of losses annually and yet it lacks efficient approaches in detecting frauds considering user profile and their behaviors simultaneously in social network . A social network forms a graph structure whilst Graph neural networks (GNN), a promising research domain in Deep Learning, can seamlessly process non-Euclidean graph data . In financial fraud detection, the modus operandi of criminals can be identified by analyzing user profile and their behaviors such as transaction, loaning etc. as well as their social connectivity. Currently, most GNNs are incapable of selecting important neighbors since the neighbors' edge attributes (i.e., behaviors) are ignored. In this paper, we propose a novel behavior information aggregation network (BIAN) to combine the user behaviors with other user features. Different from its close "relatives" such as Graph Attention Networks (GAT) and Graph Transformer Networks (GTN), it aggregates neighbors based on neighboring edge attribute distribution, namely, user behaviors in financial social network. The experimental results on a real-world large-scale financial social network dataset, DGraph, show that BIAN obtains the 10.2% gain in AUROC comparing with the State-Of-The-Art models.

Long Zhang, Wei-neng Chen, Feng-feng Wei et al.

Language models often generate reasoning before giving a final answer, but the visible answer does not reveal when the model's answer preference became stable. We study this question through a narrow computable object: \emph{finite-answer preference stabilization}. For a model state and specified answer verbalizers, we project the model's own continuation probabilities onto a finite answer set; in binary tasks this yields an exact log-odds code, $δ(ξ)=S_θ(\mathrm{yes}\midξ)-S_θ(\mathrm{no}\midξ)$. This target defines parser-based answer onset, retrospective stabilization time, and lead without relying on greedy rollouts or learned probes. In controlled delayed-verdict tasks with Qwen3-4B-Instruct, the contextual finite-answer projection stabilizes before the answer is parseable, with 17--31 token mean lead in the main templates and positive, shorter lead in a parser-clean replication. The signal tracks the model's eventual output rather than truth, is linearly recoverable from compact hidden summaries, is partly separable from cursor progress, and transfers as shared information without a single invariant coordinate. Diagnostics separate the measurement from online stopping, verbalizer-free belief, and causal answer control; exact steering shows local sensitivity of $δ$ but not reliable generation control.

Long Zhang, Dai-jun Lin, Wei-neng Chen

Large language models (LLMs) generalize smoothly across continuous semantic spaces, yet strict logical reasoning demands the formation of discrete decision boundaries. Prevailing theories relying on linear isometric projections fail to resolve this fundamental tension. In this work, we argue that task context operates as a non-isometric dynamical operator that enforces a necessary "topological distortion." By applying Gram-Schmidt decomposition to residual-stream activations , we reveal a dual-modulation mechanism driving this process: a class-agnostic topological preservation that anchors global structure to prevent semantic collapse, and a specific algebraic divergence that directionally tears apart cross-class concepts to forge logical boundaries. We validate this geometric evolution across a gradient of tasks, from simple mapping to complex primality testing. Crucially, targeted specific vector ablation establishes a strict causal binding between this topology and model function: algebraically erasing the divergence component collapses parity classification accuracy from 100% to chance levels (38.57%). Furthermore, we uncover a three-phase layer-wise geometric dynamic and demonstrate that under social pressure prompts, models fail to generate sufficient divergence. This results in a "manifold entanglement" that geometrically explains sycophancy and hallucination. Ultimately, our findings revise the linear-isometric presumption, demonstrating that the emergence of discrete logic in LLMs is purchased at an irreducible cost of topological deformation.

Anxiang Zeng, Haibo Zhang, Hailing Zhang et al.

We present CompassMax-V3-Thinking, a hundred-billion-scale MoE reasoning model trained with a new RL framework built on one principle: each prompt must matter. Scaling RL to this size exposes critical inefficiencies-zero-variance prompts that waste rollouts, unstable importance sampling over long horizons, advantage inversion from standard reward models, and systemic bottlenecks in rollout processing. To overcome these challenges, we introduce several unified innovations: (1) Multi-Stage Zero-Variance Elimination, which filters out non-informative prompts and stabilizes group-based policy optimization (e.g. GRPO) by removing wasted rollouts; (2) ESPO, an entropy-adaptive optimization method that balances token-level and sequence-level importance sampling to maintain stable learning dynamics; (3) a Router Replay strategy that aligns training-time MoE router decisions with inference-time behavior to mitigate train-infer discrepancies, coupled with a reward model adjustment to prevent advantage inversion; (4) a high-throughput RL system with FP8-precision rollouts, overlapped reward computation, and length-aware scheduling to eliminate performance bottlenecks. Together, these contributions form a cohesive pipeline that makes RL on hundred-billion-scale MoE models stable and efficient. The resulting model delivers strong performance across both internal and public evaluations.

Long Zhang, Zi-bo Qin, Wei-neng Chen

Large language models (LLMs) increasingly fuse heterogeneous inputs in ubiquitous systems. Yet, how LLMs implicitly allocate authority when sensor measurements and user claims conflict remains unexamined, raising critical reliability concerns for deployments where physical sensing must retain priority. Unlike explicit traditional fusion, LLMs bury authority allocation within learned representations. We discover this allocation is severely format-dependent: numerical sensor data fails to integrate into answer-relevant model directions, allowing natural-language claims to dominate the final decision, a phenomenon we term \textbf{Authority Inversion}.To diagnose and mitigate this, we develop a geometric framework of context integration, introduce two computable audit metrics, specifically the Context Integration Ratio (CIR) and Authority Alignment Index (AAI), and propose Geometric Authority Calibration (GAC), an inference-time layer-level intervention to suppress misplaced user authority. Evaluating four models (4B to 35B parameters, three architectures) across four datasets totaling 576 conflict instances reveals extreme inversion: on numerical tasks, models exhibit near-zero sensor trust (AAI = -0.805, Cohen's d = -2.14), unaffected by model capacity. Validating our geometric framework, theory-guided causal injection flips 80.2\% of incorrect decisions (vs. <0.4\% for random controls). Practically, GAC improves HAR accuracy from 0 -- 1.6\% to 21.9 -- 27.5\%, outperforming prompting baselines. Ultimately, authority allocation in LLM-mediated systems must be explicitly audited and application-specifically configured rather than left implicit.

Ryan Wong, Jiawei Wang, Junjie Zhao et al.

From professional research to everyday planning, many tasks are bottlenecked by wide-scale information seeking, which is more repetitive than cognitively complex. With the rapid development of Large Language Models (LLMs), automated search agents powered by LLMs offer a promising solution to liberate humans from this tedious work. However, the capability of these agents to perform such "wide-context" collection reliably and completely remains largely unevaluated due to a lack of suitable benchmarks. To bridge this gap, we introduce WideSearch, a new benchmark engineered to evaluate agent reliability on these large-scale collection tasks. The benchmark features 200 manually curated questions (100 in English, 100 in Chinese) from over 15 diverse domains, grounded in real user queries. Each task requires agents to collect large-scale atomic information, which could be verified one by one objectively, and arrange it into a well-organized output. A rigorous five-stage quality control pipeline ensures the difficulty, completeness, and verifiability of the dataset. We benchmark over 10 state-of-the-art agentic search systems, including single-agent, multi-agent frameworks, and end-to-end commercial systems. Most systems achieve overall success rates near 0\%, with the best performer reaching just 5\%. However, given sufficient time, cross-validation by multiple human testers can achieve a near 100\% success rate. These results demonstrate that present search agents have critical deficiencies in large-scale information seeking, underscoring urgent areas for future research and development in agentic search. Our dataset, evaluation pipeline, and benchmark results have been publicly released at https://widesearch-seed.github.io/

Peipei Song, Long Zhang, Long Lan et al.

Partially relevant video retrieval (PRVR) is a practical yet challenging task in text-to-video retrieval, where videos are untrimmed and contain much background content. The pursuit here is of both effective and efficient solutions to capture the partial correspondence between text queries and untrimmed videos. Existing PRVR methods, which typically focus on modeling multi-scale clip representations, however, suffer from content independence and information redundancy, impairing retrieval performance. To overcome these limitations, we propose a simple yet effective approach with active moment discovering (AMDNet). We are committed to discovering video moments that are semantically consistent with their queries. By using learnable span anchors to capture distinct moments and applying masked multi-moment attention to emphasize salient moments while suppressing redundant backgrounds, we achieve more compact and informative video representations. To further enhance moment modeling, we introduce a moment diversity loss to encourage different moments of distinct regions and a moment relevance loss to promote semantically query-relevant moments, which cooperate with a partially relevant retrieval loss for end-to-end optimization. Extensive experiments on two large-scale video datasets (\ie, TVR and ActivityNet Captions) demonstrate the superiority and efficiency of our AMDNet. In particular, AMDNet is about 15.5 times smaller (\#parameters) while 6.0 points higher (SumR) than the up-to-date method GMMFormer on TVR.

Long Zhang, Yuchen Xia, Bingqing Wei et al.

The advent of Large Multimodal Models (LMMs) offers a promising technology to tackle the limitations of modular design in autonomous driving, which often falters in open-world scenarios requiring sustained environmental understanding and logical reasoning. Besides, embodied artificial intelligence facilitates policy optimization through closed-loop interactions to achieve the continuous learning capability, thereby advancing autonomous driving toward embodied intelligent (El) driving. However, such capability will be constrained by relying solely on LMMs to enhance EI driving without joint decision-making. This article introduces a novel semantics and policy dual-driven hybrid decision framework to tackle this challenge, ensuring continuous learning and joint decision. The framework merges LMMs for semantic understanding and cognitive representation, and deep reinforcement learning (DRL) for real-time policy optimization. We starts by introducing the foundational principles of EI driving and LMMs. Moreover, we examine the emerging opportunities this framework enables, encompassing potential benefits and representative use cases. A case study is conducted experimentally to validate the performance superiority of our framework in completing lane-change planning task. Finally, several future research directions to empower EI driving are identified to guide subsequent work.

Yun Wang, Long Zhang, Jingren Liu et al.

Video Large Language Models (Video-LLMs) excel at general video understanding but struggle with long-form videos due to context window limits. Consequently, recent approaches focus on keyframe retrieval, condensing lengthy videos into a small set of informative frames. Despite their practicality, these methods simplify the problem to static text image matching, overlooking spatio temporal relationships crucial for capturing scene transitions and contextual continuity, and may yield redundant keyframes with limited information, diluting salient cues essential for accurate video question answering. To address these limitations, we introduce Video-EM, a training free framework inspired by the principles of human episodic memory, designed to facilitate robust and contextually grounded reasoning. Rather than treating keyframes as isolated visual entities, Video-EM explicitly models them as temporally ordered episodic events, capturing both spatial relationships and temporal dynamics necessary for accurately reconstructing the underlying narrative. Furthermore, the framework leverages chain of thought (CoT) thinking with LLMs to iteratively identify a minimal yet highly informative subset of episodic memories, enabling efficient and accurate question answering by Video-LLMs. Extensive evaluations on the Video-MME, EgoSchema, HourVideo, and LVBench benchmarks confirm the superiority of Video-EM, which achieves highly competitive results with performance gains of 4-9 percent over respective baselines while utilizing fewer frames.

Lei Huang, Hao Guo, Linzhi Peng et al.

We introduce SessionRec, a novel next-session prediction paradigm (NSPP) for generative sequential recommendation, addressing the fundamental misalignment between conventional next-item prediction paradigm (NIPP) and real-world recommendation scenarios. Unlike NIPP's item-level autoregressive generation that contradicts actual session-based user interactions, our framework introduces a session-aware representation learning through hierarchical sequence aggregation (intra/inter-session), reducing attention computation complexity while enabling implicit modeling of massive negative interactions, and a session-based prediction objective that better captures users' diverse interests through multi-item recommendation in next sessions. Moreover, we found that incorporating a rank loss for items within the session under the next session prediction paradigm can significantly improve the ranking effectiveness of generative sequence recommendation models. We also verified that SessionRec exhibits clear power-law scaling laws similar to those observed in LLMs. Extensive experiments conducted on public datasets and online A/B test in Meituan App demonstrate the effectiveness of SessionRec. The proposed paradigm establishes new foundations for developing industrial-scale generative recommendation systems through its model-agnostic architecture and computational efficiency.

Kesha Ou, Chenghao Wu, Xiaolei Wang et al.

The technical foundations of recommender systems have progressed from collaborative filtering to complex neural models and, more recently, large language models. Despite these technological advances, deployed systems often underserve their users by simply presenting a list of items, leaving the burden of exploration, comparison, and synthesis entirely on the user. This paper argues that this traditional "tool-based" paradigm fundamentally limits user experience, as the system acts as a passive filter rather than an active assistant. To address this limitation, we propose a novel deep research paradigm for recommendation, which replaces conventional item lists with comprehensive, user-centric reports. We instantiate this paradigm through RecPilot, a multi-agent framework comprising two core components: a user trajectory simulation agent that autonomously explores the item space, and a self-evolving report generation agent that synthesizes the findings into a coherent, interpretable report tailored to support user decisions. This approach reframes recommendation as a proactive, agent-driven service. Extensive experiments on public datasets demonstrate that RecPilot not only achieves strong performance in modeling user behaviors but also generates highly persuasive reports that substantially reduce user effort in item evaluation, validating the potential of this new interaction paradigm.

Long Zhang, Peipei Song, Jianfeng Dong et al.

Partially Relevant Video Retrieval (PRVR) aims to retrieve untrimmed videos partially relevant to a given query. The core challenge lies in learning robust query-video alignment against spurious semantic correlations arising from inherent data uncertainty: 1) query ambiguity, where the query incompletely characterizes the target video and often contains uninformative tokens, and 2) partial video relevance, where abundant query-irrelevant segments introduce contextual noise in cross-modal alignment. Existing methods often focus on enhancing multi-scale clip representations and retrieving the most relevant clip. However, the inherent data uncertainty in PRVR renders them vulnerable to distractor videos with spurious similarities, leading to suboptimal performance. To fill this research gap, we propose Robust Alignment Learning (RAL) framework, which explicitly models the uncertainty in data. Key innovations include: 1) we pioneer probabilistic modeling for PRVR by encoding videos and queries as multivariate Gaussian distributions. This not only quantifies data uncertainty but also enables proxy-level matching to capture the variability in cross-modal correspondences; 2) we consider the heterogeneous informativeness of query words and introduce learnable confidence gates to dynamically weight similarity. As a plug-and-play solution, RAL can be seamlessly integrated into the existing architectures. Extensive experiments across diverse retrieval backbones demonstrate its effectiveness.

Jianyu Zhang, Yongwang Zhao, Long Zhang et al.

Large language models (LLMs) for formal theorem proving have become a prominent research focus. At present, the proving ability of these LLMs is mainly evaluated through proof pass rates on datasets such as miniF2F. However, this evaluation method overlooks the varying importance of theorems. As a result, it fails to highlight the real performance disparities between LLMs and leads to high evaluation costs. This study proposes a psychometric-based evaluation method for theorem proving with LLMs, comprising two main components: Dataset Annotation and Adaptive Evaluation. First, we propose a metric calculation method to annotate the dataset with difficulty and discrimination metrics. Specifically, we annotate each theorem in the miniF2F dataset and grade them into varying difficulty levels according to the performance of LLMs, resulting in an enhanced dataset: miniF2F-Graded. Experimental results show that the difficulty grading in miniF2F-Graded better reflects the theorem difficulty perceived by LLMs. Secondly, we design an adaptive evaluation method to dynamically select the most suitable theorems for testing based on the annotated metrics and the real-time performance of LLMs. We apply this method to evaluate 10 LLMs. The results show that our method finely highlights the performance disparities between LLMs. It also reduces evaluation costs by using only 23% of the theorems in the dataset.

Long Zhang, Fangwei Lin

Large Language Models (LLMs) frequently prioritize conflicting in-context information over pre-existing parametric memory, a phenomenon often termed sycophancy or compliance. However, the mechanistic realization of this behavior remains obscure, specifically how the model resolves these knowledge conflicts through compliance, and whether this suppression arises from signal magnitude dilution or directional geometric alteration within the residual stream. To resolve this, we conducted a layer-wise geometric analysis across Qwen-4B, Llama-3.1-8B, and GLM-4-9B, decomposing the residual stream updates induced by counter-factual contexts into radial (norm-based) and angular (cosine-based) components. Our empirical results reject the universality of the "Manifold Dilution" hypothesis, as two of the three architectures maintained stable residual norms despite exhibiting significant performance degradation on factual queries. Instead, we observed that compliance is consistently characterized by "Orthogonal Interference," where the conflicting context injects a steering vector that is quasi-orthogonal to the ground-truth direction, effectively rotating the hidden state representation. This suggests that models do not "unlearn" or suppress the magnitude of internal truths but rather employ a mechanism of geometric displacement to bypass the correct unembedding vector, effectively simulating adoption while preserving the original structural magnitude. These findings challenge scalar confidence metrics for detecting hallucinations and underscore the necessity of vectorial monitoring to distinguish between genuine knowledge integration and superficial in-context mimicry.

Yuhan Chen, Yuxuan Liu, Long Zhang et al.

Multi-turn interaction remains challenging for online reinforcement learning. A common solution is trajectory-level optimization, which treats each trajectory as a single training sample. However, this approach can be inefficient and yield misleading learning signals: it applies uniform sampling across tasks regardless of difficulty, penalizes correct intermediate actions in failed trajectories, and incurs high sample-collection costs. To address these issues, we propose STEP (Success-rate-aware Trajectory-Efficient Policy optimization), a framework that dynamically allocates sampling based on per-task success rates and performs step-level optimization. STEP maintains a smoothed success-rate record to guide adaptive trajectory resampling, allocating more effort to harder tasks. It then computes success-rate-weighted advantages and decomposes trajectories into step-level samples. Finally, it applies a step-level GRPO augmentation to refine updates for low-success tasks. Experiments on OSWorld and AndroidWorld show that STEP substantially improves sample efficiency and training stability over trajectory-level GRPO, converging faster and generalizing better under the same sampling budget.

Jiaqi Yan, Ruilong Ren, Jingren Liu et al.

Egocentric AI assistants in real-world settings must process multi-modal inputs (video, audio, text), respond in real time, and retain evolving long-term memory. However, existing benchmarks typically evaluate these abilities in isolation, lack realistic streaming scenarios, or support only short-term tasks. We introduce \textbf{TeleEgo}, a long-duration, streaming, omni-modal benchmark for evaluating egocentric AI assistants in realistic daily contexts. The dataset features over 14 hours per participant of synchronized egocentric video, audio, and text across four domains: work \& study, lifestyle \& routines, social activities, and outings \& culture. All data is aligned on a unified global timeline and includes high-quality visual narrations and speech transcripts, curated through human refinement.TeleEgo defines 12 diagnostic subtasks across three core capabilities: Memory (recalling past events), Understanding (interpreting the current moment), and Cross-Memory Reasoning (linking distant events). It contains 3,291 human-verified QA items spanning multiple question formats (single-choice, binary, multi-choice, and open-ended), evaluated strictly in a streaming setting. We propose two key metrics -- Real-Time Accuracy and Memory Persistence Time -- to jointly assess correctness, temporal responsiveness, and long-term retention. TeleEgo provides a realistic and comprehensive evaluation to advance the development of practical AI assistants.

Yuanli Wu, Long Zhang, Yue Du et al.

We propose a rubric-guided, pseudo-labeled, and prompt-driven zero-shot video summarization framework that bridges large language models with structured semantic reasoning. A small subset of human annotations is converted into high-confidence pseudo labels and organized into dataset-adaptive rubrics defining clear evaluation dimensions such as thematic relevance, action detail, and narrative progression. During inference, boundary scenes, including the opening and closing segments, are scored independently based on their own descriptions, while intermediate scenes incorporate concise summaries of adjacent segments to assess narrative continuity and redundancy. This design enables the language model to balance local salience with global coherence without any parameter tuning. Across three benchmarks, the proposed method achieves stable and competitive results, with F1 scores of 57.58 on SumMe, 63.05 on TVSum, and 53.79 on QFVS, surpassing zero-shot baselines by +0.85, +0.84, and +0.37, respectively. These outcomes demonstrate that rubric-guided pseudo labeling combined with contextual prompting effectively stabilizes LLM-based scoring and establishes a general, interpretable, and training-free paradigm for both generic and query-focused video summarization.

Lingwen Deng, Yifei Han, Long Zhang et al.

Parameter-Preserving Knowledge Editing (PPKE) enables updating models with new or corrected information without retraining or parameter adjustment. Recent PPKE approaches based on knowledge graphs (KG) to extend knowledge editing (KE) capabilities to multi-hop question answering (MHQA). However, these methods often lack consistency, leading to knowledge contamination, unstable updates, and retrieval behaviors that fail to reflect the intended edits. Such inconsistencies undermine the reliability of PPKE in multi-hop reasoning. We present CAPE-KG, Consistency-Aware Parameter-Preserving Editing with Knowledge Graphs, a novel consistency-aware framework for PPKE on MHQA. CAPE-KG ensures KG construction, update, and retrieval are always aligned with the requirements of the MHQA task, maintaining coherent reasoning over both unedited and edited knowledge. Extensive experiments on the MQuAKE benchmark show accuracy improvements in PPKE performance for MHQA, demonstrating the effectiveness of addressing consistency in PPKE.

Yang Sun, Zhiyong Xie, Dan Luo et al.

Retrieval-augmented generation (RAG) incorporates external knowledge into large language models (LLMs), improving their adaptability to downstream tasks and enabling information updates. Surprisingly, recent empirical evidence demonstrates that injecting noise into retrieved relevant documents paradoxically facilitates exploitation of external knowledge and improves generation quality. Although counterintuitive and challenging to apply in practice, this phenomenon enables granular control and rigorous analysis of how LLMs integrate external knowledge. Therefore, in this paper, we intervene on noise injection and establish a layer-specific functional demarcation within the LLM: shallow layers specialize in local context modeling, intermediate layers focus on integrating long-range external factual knowledge, and deeper layers primarily rely on parametric internal knowledge. Building on this insight, we propose Layer Fused Decoding (LFD), a simple decoding strategy that directly combines representations from an intermediate layer with final-layer decoding outputs to fully exploit the external factual knowledge. To identify the optimal intermediate layer, we introduce an internal knowledge score (IKS) criterion that selects the layer with the lowest IKS value in the latter half of layers. Experimental results across multiple benchmarks demonstrate that LFD helps RAG systems more effectively surface retrieved context knowledge with minimal cost.

Tong Zhang, Wei Ye, Baosong Yang et al.

Low-frequency word prediction remains a challenge in modern neural machine translation (NMT) systems. Recent adaptive training methods promote the output of infrequent words by emphasizing their weights in the overall training objectives. Despite the improved recall of low-frequency words, their prediction precision is unexpectedly hindered by the adaptive objectives. Inspired by the observation that low-frequency words form a more compact embedding space, we tackle this challenge from a representation learning perspective. Specifically, we propose a frequency-aware token-level contrastive learning method, in which the hidden state of each decoding step is pushed away from the counterparts of other target words, in a soft contrastive way based on the corresponding word frequencies. We conduct experiments on widely used NIST Chinese-English and WMT14 English-German translation tasks. Empirical results show that our proposed methods can not only significantly improve the translation quality but also enhance lexical diversity and optimize word representation space. Further investigation reveals that, comparing with related adaptive training strategies, the superiority of our method on low-frequency word prediction lies in the robustness of token-level recall across different frequencies without sacrificing precision.

Long Zhang, Javier Ron, Benoit Baudry et al.

In this paper, we present ChaosETH, a chaos engineering approach for resilience assessment of Ethereum blockchain clients. ChaosETH operates in the following manner: First, it monitors Ethereum clients to determine their normal behavior. Then, it injects system call invocation errors into one single Ethereum client at a time, and observes the behavior resulting from perturbation. Finally, ChaosETH compares the behavior recorded before, during, and after perturbation to assess the impact of the injected system call invocation errors. The experiments are performed on the two most popular Ethereum client implementations: GoEthereum and Nethermind. We assess the impact of 22 different system call errors on those Ethereum clients with respect to 15 application-level metrics. Our results reveal a broad spectrum of resilience characteristics of Ethereum clients w.r.t. system call invocation errors, ranging from direct crashes to full resilience. The experiments clearly demonstrate the feasibility of applying chaos engineering principles to blockchain systems.

Bin Cao, Zixin Wang, Long Zhang et al.

In the past decade, blockchain has shown a promising vision greatly to build the trust without any powerful third party in a secure, decentralized and salable manner. However, due to the wide application and future development from cryptocurrency to Internet of Things, blockchain is an extremely complex system enabling integration with mathematics, finance, computer science, communication and network engineering, etc. As a result, it is a challenge for engineer, expert and researcher to fully understand the blockchain process in a systematic view from top to down. First, this article introduces how blockchain works, the research activity and challenge, and illustrates the roadmap involving the classic methodology with typical blockchain use cases and topics. Second, in blockchain system, how to adopt stochastic process, game theory, optimization, machine learning and cryptography to study blockchain running process and design blockchain protocol/algorithm are discussed in details. Moreover, the advantage and limitation using these methods are also summarized as the guide of future work to further considered. Finally, some remaining problems from technical, commercial and political views are discussed as the open issues. The main findings of this article will provide an overview in a methodology perspective to study theoretical model for blockchain fundamentals understanding, design network service for blockchain-based mechanisms and algorithms, as well as apply blockchain for Internet of Things, etc.

Mingrui Cao, Long Zhang, Bin Cao

Due to the distributed characteristics of Federated Learning (FL), the vulnerability of global model and coordination of devices are the main obstacle. As a promising solution of decentralization, scalability and security, leveraging blockchain in FL has attracted much attention in recent years. However, the traditional consensus mechanisms designed for blockchain like Proof of Work (PoW) would cause extreme resource consumption, which reduces the efficiency of FL greatly, especially when the participating devices are wireless and resource-limited. In order to address device asynchrony and anomaly detection in FL while avoiding the extra resource consumption caused by blockchain, this paper introduces a framework for empowering FL using Direct Acyclic Graph (DAG)-based blockchain systematically (DAG-FL). Accordingly, DAG-FL is first introduced from a three-layer architecture in details, and then two algorithms DAG-FL Controlling and DAG-FL Updating are designed running on different nodes to elaborate the operation of DAG-FL consensus mechanism. After that, a Poisson process model is formulated to discuss that how to set deployment parameters to maintain DAG-FL stably in different federated learning tasks. The extensive simulations and experiments show that DAG-FL can achieve better performance in terms of training efficiency and model accuracy compared with the typical existing on-device federated learning systems as the benchmarks.

Long Zhang, Brice Morin, Benoit Baudry et al.

In this paper, we present a novel fault injection framework for system call invocation errors, called Phoebe. Phoebe is unique as follows. First, Phoebe enables developers to have full observability of system call invocations. Second, Phoebe generates error models that are realistic in the sense that they mimic errors that naturally happen in production. Third, Phoebe is able to automatically conduct experiments to systematically assess the reliability of applications with respect to system call invocation errors in production. We evaluate the effectiveness and runtime overhead of Phoebe on two real-world applications in a production environment. The results show that Phoebe successfully generates realistic error models and is able to detect important reliability weaknesses with respect to system call invocation errors. To our knowledge, this novel concept of "realistic error injection", which consists of grounding fault injection on production errors, has never been studied before.

Jesper Simonsson, Long Zhang, Brice Morin et al.

In this paper, we present a novel fault injection system called ChaosOrca for system calls in containerized applications. ChaosOrca aims at evaluating a given application's self-protection capability with respect to system call errors. The unique feature of ChaosOrca is that it conducts experiments under production-like workload without instrumenting the application. We exhaustively analyze all kinds of system calls and utilize different levels of monitoring techniques to reason about the behaviour under perturbation. We evaluate ChaosOrca on three real-world applications: a file transfer client, a reverse proxy server and a micro-service oriented web application. Our results show that it is promising to detect weaknesses of resilience mechanisms related to system calls issues.

Long Zhang, Xuechao Sun, Yong Li et al.

Deep neural networks (DNNs) have been widely used in the fields such as natural language processing, computer vision and image recognition. But several studies have been shown that deep neural networks can be easily fooled by artificial examples with some perturbations, which are widely known as adversarial examples. Adversarial examples can be used to attack deep neural networks or to improve the robustness of deep neural networks. A common way of generating adversarial examples is to first generate some noises and then add them into original examples. In practice, different examples have different noise-sensitive. To generate an effective adversarial example, it may be necessary to add a lot of noise to low noise-sensitive example, which may make the adversarial example meaningless. In this paper, we propose a noise-sensitivity-analysis-based test prioritization technique to pick out examples by their noise sensitivity. We construct an experiment to validate our approach on four image sets and two DNN models, which shows that examples are sensitive to noise and our method can effectively pick out examples by their noise sensitivity.

Long Zhang, Martin Monperrus

In this paper, we present a novel resilience improvement system for Java applications. The unique feature of this system is to combine automated monitoring, automated perturbation injection, and automated resilience improvement. The latter is achieved thanks to the failure-oblivious computing, a concept introduced in 2004 by Rinard and colleagues. We design and implement the system as agents for the Java virtual machine. We evaluate the system on two real-world applications: a file transfer client and an email server. Our results show that it is possible to automatically improve the resilience of Java applications with respect to uncaught or mishandled exceptions.

Long Zhang, Brice Morin, Philipp Haller et al.

Software systems contain resilience code to handle those failures and unexpected events happening in production. It is essential for developers to understand and assess the resilience of their systems. Chaos engineering is a technology that aims at assessing resilience and uncovering weaknesses by actively injecting perturbations in production. In this paper, we propose a novel design and implementation of a chaos engineering system in Java called ChaosMachine. It provides a unique and actionable analysis on exception-handling capabilities in production, at the level of try-catch blocks. To evaluate our approach, we have deployed ChaosMachine on top of 3 large-scale and well-known Java applications totaling 630k lines of code. Our results show that ChaosMachine reveals both strengths and weaknesses of the resilience code of a software system at the level of exception handling.