Zixuan Yang

Zixuan Yang, Yibo Zhao, Weicong Liu et al.

Matching submissions with suitable reviewers at scale is a growing challenge for major venues, yet existing approaches either rely on coarse proxy signals that conflate general relatedness with true suitability, or require expensive human annotations that are difficult to scale for training. We propose MERIT, a two-stage framework that bridges this gap by converting criterion-level expertise matching into scalable suitability supervision. In the first stage, we train a reviewer assessor via reinforcement learning to identify the expertise dimensions a paper requires, match them against the reviewer's prior work, and produce a suitability decision, with rewards provided by an LLM judge guided by paper-specific expertise rubrics. In the second stage, we distill the assessor's predictions into an embedding-based retriever for efficient large-scale assignment. Experiments show that our 4B reviewer assessor outperforms larger general-purpose LLMs on suitability classification, and the resulting retriever achieves state-of-the-art performance across LR-Bench and the CMU Gold dataset. Our code is available at https://github.com/Luli3220/MERIT.

Yiqun Chen, Lingyong Yan, Zixuan Yang et al.

Agentic search has emerged as a promising paradigm for complex information seeking by enabling Large Language Models (LLMs) to interleave reasoning with tool use. However, prevailing systems rely on monolithic agents that suffer from structural bottlenecks, including unconstrained reasoning outputs that inflate trajectories, sparse outcome-level rewards that complicate credit assignment, and stochastic search noise that destabilizes learning. To address these challenges, we propose \textbf{M-ASK} (Multi-Agent Search and Knowledge), a framework that explicitly decouples agentic search into two complementary roles: Search Behavior Agents, which plan and execute search actions, and Knowledge Management Agents, which aggregate, filter, and maintain a compact internal context. This decomposition allows each agent to focus on a well-defined subtask and reduces interference between search and context construction. Furthermore, to enable stable coordination, M-ASK employs turn-level rewards to provide granular supervision for both search decisions and knowledge updates. Experiments on multi-hop QA benchmarks demonstrate that M-ASK outperforms strong baselines, achieving not only superior answer accuracy but also significantly more stable training dynamics.\footnote{The source code for M-ASK is available at https://github.com/chenyiqun/M-ASK.}

Tao Wang, Yushu Zhang, Zixuan Yang et al.

Massive captured face images are stored in the database for the identification of individuals. However, these images can be observed unintentionally by data managers, which is not at the will of individuals and may cause privacy violations. Existing protection schemes can maintain identifiability but slightly change the facial appearance, rendering it still susceptible to the visual perception of the original identity by data managers. In this paper, we propose an effective identity hider for human vision protection, which can significantly change appearance to visually hide identity while allowing identification for face recognizers. Concretely, the identity hider benefits from two specially designed modules: 1) The virtual face generation module generates a virtual face with a new appearance by manipulating the latent space of StyleGAN2. In particular, the virtual face has a similar parsing map to the original face, supporting other vision tasks such as head pose detection. 2) The appearance transfer module transfers the appearance of the virtual face into the original face via attribute replacement. Meanwhile, identity information can be preserved well with the help of the disentanglement networks. In addition, diversity and background preservation are supported to meet the various requirements. Extensive experiments demonstrate that the proposed identity hider achieves excellent performance on privacy protection and identifiability preservation.

Zixuan Yang, Yiqun Chen, Wei Yang et al.

Reinforcement learning in open-ended long-form generation is challenging because reliable reference answers and automatic metrics are often unavailable. Existing rubric-based methods typically rely on pointwise LLM-as-a-judge scoring, but absolute scores are difficult to calibrate across complex responses, may provide weak discrimination among same-query rollouts, and can become saturated during optimization. We propose Tournament-GRPO, a group-wise reward framework that converts rubric-guided LLM judgments into relative rewards through repeated multi-round tournaments among same-query rollouts. Tournament-GRPO compares candidates within groups, accumulates tournament outcomes, and normalizes them into group-wise rewards for GRPO training. Experiments on Deep Research Bench show that Tournament-GRPO consistently outperforms existing reward-design baselines, achieving a 4.52-point overall-score improvement over the strongest baseline. Further analyses show that tournament rewards provide a favorable effectiveness--efficiency trade-off and that tournament design affects training dynamics. These results suggest that rubric-guided tournament comparison provides an effective reward signal for reinforcement learning in open-ended long-form generation.

Kangkang Lu, Yanhua Yu, Hao Fei et al.

In recent years, spectral graph neural networks, characterized by polynomial filters, have garnered increasing attention and have achieved remarkable performance in tasks such as node classification. These models typically assume that eigenvalues for the normalized Laplacian matrix are distinct from each other, thus expecting a polynomial filter to have a high fitting ability. However, this paper empirically observes that normalized Laplacian matrices frequently possess repeated eigenvalues. Moreover, we theoretically establish that the number of distinguishable eigenvalues plays a pivotal role in determining the expressive power of spectral graph neural networks. In light of this observation, we propose an eigenvalue correction strategy that can free polynomial filters from the constraints of repeated eigenvalue inputs. Concretely, the proposed eigenvalue correction strategy enhances the uniform distribution of eigenvalues, thus mitigating repeated eigenvalues, and improving the fitting capacity and expressive power of polynomial filters. Extensive experimental results on both synthetic and real-world datasets demonstrate the superiority of our method. The code is available at: https://github.com/Lukangkang123/EC-GNN

Weicong Liu, Zixuan Yang, Yibo Zhao et al.

Reviewer assignment is increasingly critical yet challenging in the LLM era, where rapid topic shifts render many pre-2023 benchmarks outdated and where proxy signals poorly reflect true reviewer familiarity. We address this evaluation bottleneck by introducing LR-bench, a high-fidelity, up-to-date benchmark curated from 2024-2025 AI/NLP manuscripts with five-level self-assessed familiarity ratings collected via a large-scale email survey, yielding 1055 expert-annotated paper-reviewer-score annotations. We further propose RATE, a reviewer-centric ranking framework that distills each reviewer's recent publications into compact keyword-based profiles and fine-tunes an embedding model with weak preference supervision constructed from heuristic retrieval signals, enabling matching each manuscript against a reviewer profile directly. Across LR-bench and the CMU gold-standard dataset, our approach consistently achieves state-of-the-art performance, outperforming strong embedding baselines by a clear margin. We release LR-bench at https://huggingface.co/datasets/Gnociew/LR-bench, and a GitHub repository at https://github.com/Gnociew/RATE-Reviewer-Assign.

Rao Fu, Zixuan Yang, Jiankun Zhang et al.

Scaling model performance typically requires increasing model size. Looped Transformer offers a compelling alternative by iteratively reusing the same Transformer blocks, trading additional computation for improved performance without increasing parameter count or context length. Because the number of loop iterations can be adjusted at inference, it also provides a natural mechanism for balancing performance and test-time compute. However, Looped Transformer still suffers from training instability when the number of loop iterations increases. Our analysis reveals that this instability stems from two sources: gradient oscillation and residual explosion. To address these two problems, we propose the Fully Looped Transformer, which introduces two parameter-free modifications: (1) Fully Looped Architecture, which distributes inter-loop signals across all layers to mitigate residual explosion; (2) Attention Injection, which reuses the existing attention block to suppress gradient oscillation. These modifications stabilize training dynamics, enabling the Fully Looped Transformer to be trained stably up to 12 loop iterations, whereas other baseline looped models collapse in this regime. In milder settings where Looped Transformer does not collapse, Fully Looped Transformer still improves average downstream-task performance by up to 13.2\%. Overall, our experiments demonstrate that Fully Looped Transformer improves training stability, enhances downstream performance, and provides preliminary adaptability under different test-time compute budgets by varying loop iterations at inference.

Yiqun Chen, Erhan Zhang, Tianyi Hu et al.

The evolution of Retrieval-Augmented Generation (RAG) has shifted from static retrieval pipelines to dynamic, agentic workflows where a central planner orchestrates multi-turn reasoning. However, existing paradigms face a critical dichotomy: they either optimize modules jointly within rigid, fixed-graph architectures, or empower dynamic planning while treating executors as frozen, black-box tools. We identify that this \textit{decoupled optimization} creates a ``strategic-operational mismatch,'' where sophisticated planning strategies fail to materialize due to unadapted local executors, often leading to negative performance gains despite increased system complexity. In this paper, we propose \textbf{JADE} (\textbf{J}oint \textbf{A}gentic \textbf{D}ynamic \textbf{E}xecution), a unified framework for the joint optimization of planning and execution within dynamic, multi-turn workflows. By modeling the system as a cooperative multi-agent team unified under a single shared backbone, JADE enables end-to-end learning driven by outcome-based rewards. This approach facilitates \textit{co-adaptation}: the planner learns to operate within the capability boundaries of the executors, while the executors evolve to align with high-level strategic intent. Empirical results demonstrate that JADE transforms disjoint modules into a synergistic system, yielding remarkable performance improvements via joint optimization and enabling a flexible balance between efficiency and effectiveness through dynamic workflow orchestration.

Zixuan Yang, Jiaqi Zheng, Guihai Chen

Ensuring verifiable and interpretable safety of deep reinforcement learning (DRL) is crucial for its deployment in real-world applications. Existing approaches like verification-in-the-loop training, however, face challenges such as difficulty in deployment, inefficient training, lack of interpretability, and suboptimal performance in property satisfaction and reward performance. In this work, we propose a novel verification-driven interpretation-in-the-loop framework Reintrainer to develop trustworthy DRL models, which are guaranteed to meet the expected constraint properties. Specifically, in each iteration, this framework measures the gap between the on-training model and predefined properties using formal verification, interprets the contribution of each input feature to the model's output, and then generates the training strategy derived from the on-the-fly measure results, until all predefined properties are proven. Additionally, the low reusability of existing verifiers and interpreters motivates us to develop Reinfier, a general and fundamental tool within Reintrainer for DRL verification and interpretation. Reinfier features breakpoints searching and verification-driven interpretation, associated with a concise constraint-encoding language DRLP. Evaluations demonstrate that Reintrainer outperforms the state-of-the-art on six public benchmarks in both performance and property guarantees. Our framework can be accessed at https://github.com/Kurayuri/Reinfier.

Yunli Wang, Zhen Zhang, Zixuan Yang et al.

The scaling law is a notable property of neural network models and has significantly propelled the development of large language models. Scaling laws hold great promise in guiding model design and resource allocation. Recent research increasingly shows that scaling laws are not limited to NLP tasks or Transformer architectures; they also apply to domains such as recommendation. However, there is still a lack of literature on scaling law research in online advertisement retrieval systems. This may be because 1) identifying the scaling law for resource cost and online revenue is often expensive in both time and training resources for industrial applications, and 2) varying settings for different systems prevent the scaling law from being applied across various scenarios. To address these issues, we propose a lightweight paradigm to identify online scaling laws of retrieval models, incorporating a novel offline metric and an offline simulation algorithm. We prove that under mild assumptions, the correlation between the novel metric and online revenue asymptotically approaches 1 and empirically validates its effectiveness. The simulation algorithm can estimate the machine cost offline. Based on the lightweight paradigm, we can identify online scaling laws for retrieval models almost exclusively through offline experiments, and quickly estimate machine costs and revenues for given model configurations. We further validate the existence of scaling laws across mainstream model architectures (e.g., Transformer, MLP, and DSSM) in our real-world advertising system. With the identified scaling laws, we demonstrate practical applications for ROI-constrained model designing and multi-scenario resource allocation in the online advertising system. To the best of our knowledge, this is the first work to study identification and application of online scaling laws for online advertisement retrieval.

Yunli Wang, Zhen Zhang, Zhiqiang Wang et al.

Cascade Ranking is a prevalent architecture in large-scale top-k selection systems like recommendation and advertising platforms. Traditional training methods focus on single-stage optimization, neglecting interactions between stages. Recent advances have introduced interaction-aware training paradigms, but still struggle to 1) align training objectives with the goal of the entire cascade ranking (i.e., end-to-end recall of ground-truth items) and 2) learn effective collaboration patterns for different stages. To address these challenges, we propose LCRON, which introduces a novel surrogate loss function derived from the lower bound probability that ground truth items are selected by cascade ranking, ensuring alignment with the overall objective of the system. According to the properties of the derived bound, we further design an auxiliary loss for each stage to drive the reduction of this bound, leading to a more robust and effective top-k selection. LCRON enables end-to-end training of the entire cascade ranking system as a unified network. Experimental results demonstrate that LCRON achieves significant improvement over existing methods on public benchmarks and industrial applications, addressing key limitations in cascade ranking training and significantly enhancing system performance.

Wenxuan Sun, Zixuan Yang, Yunli Wang et al.

Advertising systems often face the multi-domain challenge, where data distributions vary significantly across scenarios. Existing domain adaptation methods primarily focus on building domain-adaptive neural networks but often rely on hand-crafted domain information, e.g., advertising placement, which may be sub-optimal. We think that fine-grained "domain" patterns exist that are difficult to hand-craft in online advertisement. Thus, we propose Adaptive$^2$, a novel framework that first learns domains adaptively using a domain mining module by self-supervision and then employs a shared&specific network to model shared and conflicting information. As a practice, we use VQ-VAE as the domain mining module and conduct extensive experiments on public benchmarks. Results show that traditional domain adaptation methods with hand-crafted domains perform no better than single-domain models under fair FLOPS conditions, highlighting the importance of domain definition. In contrast, Adaptive$^2$ outperforms existing approaches, emphasizing the effectiveness of our method and the significance of domain mining. We also deployed Adaptive$^2$ in the live streaming scenario of Kuaishou Advertising System, demonstrating its commercial value and potential for automatic domain identification. To the best of our knowledge, Adaptive$^2$ is the first approach to automatically learn both domain identification and adaptation in online advertising, opening new research directions for this area.

Yue Lu, Yun Zhou, Zhuqing Jiang et al.

Convolutional neural networks (CNNs) have demonstrated superior performance in super-resolution (SR). However, most CNN-based SR methods neglect the different importance among feature channels or fail to take full advantage of the hierarchical features. To address these issues, this paper presents a novel recursive unit. Firstly, at the beginning of each unit, we adopt a compact channel attention mechanism to adaptively recalibrate the channel importance of input features. Then, the multi-level features, rather than only deep-level features, are extracted and fused. Additionally, we find that it will force our model to learn more details by using the learnable upsampling method (i.e., transposed convolution) only on residual branch (instead of using it both on residual branch and identity branch) while using the bicubic interpolation on the other branch. Analytic experiments show that our method achieves competitive results compared with the state-of-the-art methods and maintains faster speed as well.