arXiv subjects

Runsheng Yu

Publications and source records attributed to Runsheng Yu.

At least 19 recordsLinked to original sources

arXiv2026

Structural Dual Adaptation for Safe Reinforcement Learning

Primal-dual methods enforce safety constraints in reinforcement learning through Lagrangian relaxation. For a fixed policy, the exact bounded multiplier response exhibits bang-bang switching at the constraint boundary. Practical gradient updates can oscillate through coupling with policy optimization. We propose Structural DuaL Adaptation (SDLA), which replaces this bang-bang response with a unique continuous response through current-cost centering, trajectory weighting, and quadratic regularization. We characterize activation thresholds and response sensitivity. Ternary search computes multipliers without a dual gradient stepsize. For entropy-regularized tabular policies, we bound equilibrium constraint violations alongside reward gaps. We establish convergence of delayed updates under mild assumptions. Neural PPO/SAC experiments on OmniSafe and HASARD demonstrate competitive rewards with improved constraint tracking. On HASARD, PPO-SDLA achieves higher reported mean reward than the human baseline in six of eighteen settings while meeting the respective mean-cost budgets, including four with lower mean cost. To the best of our knowledge, this is the first report of higher mean reward at lower mean cost than the human baseline across multiple HASARD settings.

cs.LG↗

arXiv2026

AlphaIdeaBench: Benchmarking Automated Research Idea Generation under Feasibility Constraints with Peer Review Alignment

Research ideas are always proposed under resource constraints: a plan must fit the available compute, API credit, and time. Yet evaluations of research-ideation agents mostly score unconstrained novelty or overlap with a published paper, and their LLM judges are rarely checked against peer review. We introduce AlphaIdeaBench, a benchmark of 100 ideation tasks (50 workshop calls for papers and 50 open problems stated in surveys), each paired with an explicit resource budget and a shared proposal schema. An LLM judge scores each proposal on 12 dimensions covering validity, novelty, and feasibility. To validate the judge, we compare its scores on four novelty axes (problem, method, scenario, and insight) with scores derived from reviewer comments on 600 papers from ICLR 2026, NeurIPS 2025, and ICML 2025. For problem and method novelty, Gemini 3.7 Flash is within one point of the reviewer-derived score on 94-97% of papers; agreement on scenario novelty is relatively lower. We evaluate eight published harnesses on a shared backbone and, within one harness, eight backbone LLMs. CoI-Agent obtains the highest overall score, narrowly ahead of SciMON, which leads on method and insight novelty. All harnesses score low on scenario novelty and rarely change the problem setting. The backbone largely determines which open problem a proposal targets, whereas the harness determines whether a proposal addresses one problem or combines several. Weaker backbones such as Llama 4 Scout fail to carry out that harness's strategy and score low on budget feasibility. Finally, when only the budget is varied on three topics, a larger budget changes the experimental setting on compute-intensive topics but only adds tests on an inexpensive one. We release the tasks, rubrics, and scores to foster standardized, reproducible benchmarking of future research-ideation agents under constraints.

cs.AI↗

arXiv2026

DGA-Muon: Decoupled Geometry-Aligned Adaptive Scaling for Muon

While NorMuon has achieved strong empirical performance in pretraining, its underlying adaptive mechanism remains largely heuristic and poorly understood. In this work, we provide the first systematic theoretical analysis of NorMuon's adaptivity, revealing that it primarily arises from orthogonalization-induced geometry rather than genuine optimization dynamics, serving to offset the resulting geometric non-uniformity. Under exact orthogonalization, the adaptive scaling factors degenerate into a single global scalar for square and wide matrices, while for tall matrices their variation results from unevenly distributed row energy after orthogonalization. Under approximate orthogonalization, the orthogonality residuals introduce additional variation into the scaling, giving rise to a counterintuitive Orthogonalization--Adaptivity Paradox: more accurate orthogonalization weakens adaptivity. We further show that NorMuon's rigid row-wise scaling is geometrically misaligned with the one-sided orthogonal structure of tall matrices by distorting column orthogonality. Motivated by these limitations, we propose two core design principles that a desirable adaptive mechanism for Muon should satisfy. First, adaptive scaling should be decoupled from orthogonalization, with the scaling factors computed directly from raw gradients. Second, adaptive scaling should be aligned with the shape-dependent orthogonal structure of the polar factor, using row-wise scaling for wide matrices and column-wise scaling for tall matrices. We prove that this geometry-aligned scaling preserves the orthogonal structure of the update. By incorporating several other techniques, we obtain Decoupled Geometry-Aligned Muon (DGA-Muon). We establish convergence guarantees for DGA-Muon and empirically validate both our theoretical characterization of NorMuon's scaling degeneration and the superiority of DGA-Muon.

cs.AI↗

arXiv2026

Matrix AdaGrad: Row-wise and Column-wise Adaptive Subgradient Methods

Adaptive optimization methods such as AdaGrad and Adam are widely used in modern deep neural network training, but their adaptive scaling is primarily designed for vector-valued parameters and does not explicitly exploit matrix structure. Recent matrix-aware optimizers demonstrate the benefits of structured optimization, yet a general theoretical framework for deriving matrix-aware adaptivity comparable to that of AdaGrad remains lacking. In this work, we develop an Online Mirror Descent framework with adaptive proximal functions for matrix-valued parameters, providing a principled methodology for deriving matrix-aware adaptive optimization through online regret minimization. By introducing row-wise and column-wise matrix proximal functions, our framework explicitly reveals the trade-off governing adaptive scaling: increasing the scaling factors reduces the gradient-dependent dual norm term while increasing the cost of evolving the proximal geometry. In the row-wise setting, this trade-off becomes separable under diagonal parameterization, allowing the adaptive scaling for each row to be derived independently by minimizing its corresponding row-wise regret bound. The column-wise counterpart follows directly by applying the row-wise construction to the transposed matrix. This framework yields Row-wise Matrix AdaGrad and Column-wise Matrix AdaGrad as concrete instantiations, with regret guarantees that are strictly tighter than those of entry-wise AdaGrad under row-sparse or column-sparse gradient structures. Experiments on matrix factorization and stacked deep MLP training further demonstrate the benefits of matrix-aware adaptive scaling, yielding improved optimization performance in both settings and enhanced optimization stability and trainability at larger learning rates and greater network depths in the latter.

cs.LG↗

arXiv2026

Model Evolution Under Zeroth-Order Optimization: A Neural Tangent Kernel Perspective

Zeroth-order (ZO) optimization enables memory-efficient training of neural networks by estimating gradients via forward passes only, eliminating the need for backpropagation. However, the stochastic nature of gradient estimation significantly obscures the training dynamics, in contrast to the well-characterized behavior of first-order methods under Neural Tangent Kernel (NTK) theory. To address this, we introduce the Neural Zeroth-order Kernel (NZK) to describe model evolution in function space under ZO updates. For linear models, we prove that the expected NZK remains constant throughout training and depends explicitly on the first and second moments of the random perturbation directions. This invariance yields a closed-form expression for model evolution under squared loss. We further extend the analysis to linearized neural networks. Interpreting ZO updates as kernel gradient descent via NZK provides a novel perspective for potentially accelerating convergence. Extensive experiments across synthetic and real-world datasets (including MNIST, CIFAR-10, and Tiny ImageNet) validate our theoretical results and demonstrate acceleration when using a single shared random vector.

cs.LG↗

arXiv2026

How Much Information Can a Vision Token Hold? A Scaling Law for Recognition Limits in VLMs

Recent vision-centric approaches have made significant strides in long-context modeling. Represented by DeepSeek-OCR, these models encode rendered text into continuous vision tokens, achieving high compression rates without sacrificing recognition precision. However, viewing the vision encoder as a lossy channel with finite representational capacity raises a fundamental question: what is the information upper bound of visual tokens? To investigate this limit, we conduct controlled stress tests by progressively increasing the information quantity (character count) within an image. We observe a distinct phase-transition phenomenon characterized by three regimes: a near-perfect Stable Phase, an Instability Phase marked by increased error variance, and a total Collapse Phase. We analyze the mechanical origins of these transitions and identify key factors. Furthermore, we formulate a probabilistic scaling law that unifies average vision token load and visual density into a latent difficulty metric. Extensive experiments across various Vision-Language Models demonstrate the universality of this scaling law, providing critical empirical guidance for optimizing the efficiency-accuracy trade-off in visual context compression.

cs.LG↗

arXiv2025

Fighter: Unveiling the Graph Convolutional Nature of Transformers in Time Series Modeling

Transformers have achieved remarkable success in time series modeling, yet their internal mechanisms remain opaque. This work demystifies the Transformer encoder by establishing its fundamental equivalence to a Graph Convolutional Network (GCN). We show that in the forward pass, the attention distribution matrix serves as a dynamic adjacency matrix, and its composition with subsequent transformations performs computations analogous to graph convolution. Moreover, we demonstrate that in the backward pass, the update dynamics of value and feed-forward projections mirror those of GCN parameters. Building on this unified theoretical reinterpretation, we propose \textbf{Fighter} (Flexible Graph Convolutional Transformer), a streamlined architecture that removes redundant linear projections and incorporates multi-hop graph aggregation. This perspective yields an explicit and interpretable representation of temporal dependencies across different scales, naturally expressed as graph edges. Experiments on standard forecasting benchmarks confirm that Fighter achieves competitive performance while providing clearer mechanistic interpretability of its predictions.

cs.LG↗

arXiv2025

Particle Dynamics for Latent-Variable Energy-Based Models

Latent-variable energy-based models (LVEBMs) assign a single normalized energy to joint pairs of observed data and latent variables, offering expressive generative modeling while capturing hidden structure. We recast maximum-likelihood training as a saddle problem over distributions on the latent and joint manifolds and view the inner updates as coupled Wasserstein gradient flows. The resulting algorithm alternates overdamped Langevin updates for a joint negative pool and for conditional latent particles with stochastic parameter ascent, requiring no discriminator or auxiliary networks. We prove existence and convergence under standard smoothness and dissipativity assumptions, with decay rates in KL divergence and Wasserstein-2 distance. The saddle-point view further yields an ELBO strictly tighter than bounds obtained with restricted amortized posteriors. Our method is evaluated on numerical approximations of physical systems and performs competitively against comparable approaches.

cs.LG↗

arXiv2025

Seed-X: Building Strong Multilingual Translation LLM with 7B Parameters

Multilingual translation stands as a challenging task for large language models (LLMs) to handle intricate language patterns and stilted translations that arise in automated translations. In this paper, we introduce Seed-X, a family of open-source LLMs comprising instruct and reasoning models, pushing the limits of translation capability with 7B parameter size. The base model is pre-trained on a diverse, high-quality dataset encompassing both monolingual and bilingual content across 28 languages, harnessing the full potential of multilingual data. The instruct model is then finetuned to translate by Chain-of-Thought (CoT) reasoning and further enhanced through reinforcement learning (RL) to achieve better generalization across diverse language pairs. Seed-X achieves performance comparable to leading closed-source models, including Gemini-2.5 and GPT-4o, across 28 languages, and significantly outperforms larger open-source models in both automatic metrics and human evaluations. We share the best practices through our optimization process, and make the parameter public available for advancing translation research and applications.

cs.CL↗

arXiv2025

Seed LiveInterpret 2.0: End-to-end Simultaneous Speech-to-speech Translation with Your Voice

Simultaneous Interpretation (SI) represents one of the most daunting frontiers in the translation industry, with product-level automatic systems long plagued by intractable challenges: subpar transcription and translation quality, lack of real-time speech generation, multi-speaker confusion, and translated speech inflation, especially in long-form discourses. In this study, we introduce Seed-LiveInterpret 2.0, an end-to-end SI model that delivers high-fidelity, ultra-low-latency speech-to-speech generation with voice cloning capabilities. As a fully operational product-level solution, Seed-LiveInterpret 2.0 tackles these challenges head-on through our novel duplex speech-to-speech understanding-generating framework. Experimental results demonstrate that through large-scale pretraining and reinforcement learning, the model achieves a significantly better balance between translation accuracy and latency, validated by human interpreters to exceed 70% correctness in complex scenarios. Notably, Seed-LiveInterpret 2.0 outperforms commercial SI solutions by significant margins in translation quality, while slashing the average latency of cloned speech from nearly 10 seconds to a near-real-time 3 seconds, which is around a near 70% reduction that drastically enhances practical usability.

cs.CL↗

arXiv2025

BEYOND DIALOGUE: A Profile-Dialogue Alignment Framework Towards General Role-Playing Language Model

The rapid advancement of large language models (LLMs) has revolutionized role-playing, enabling the development of general role-playing models. However, current role-playing training has two significant issues: (I) Using a predefined role profile to prompt dialogue training for specific scenarios usually leads to inconsistencies and even conflicts between the dialogue and the profile, resulting in training biases. (II) The model learns to imitate the role based solely on the profile, neglecting profile-dialogue alignment at the sentence level. In this work, we propose a simple yet effective framework called BEYOND DIALOGUE, designed to overcome these hurdles. This framework innovatively introduces "beyond dialogue" tasks to align dialogue with profile traits based on each specific scenario, thereby eliminating biases during training. Furthermore, by adopting an innovative prompting mechanism that generates reasoning outcomes for training, the framework allows the model to achieve fine-grained alignment between profile and dialogue at the sentence level. The aforementioned methods are fully automated and low-cost. Additionally, the integration of automated dialogue and objective evaluation methods forms a comprehensive framework, paving the way for general role-playing. Experimental results demonstrate that our model excels in adhering to and reflecting various dimensions of role profiles, outperforming most proprietary general and specialized role-playing baselines. All code and datasets are available at https://github.com/yuyouyu32/BeyondDialogue.

cs.CL↗

arXiv2024

Tailed Low-Rank Matrix Factorization for Similarity Matrix Completion

Similarity matrix serves as a fundamental tool at the core of numerous downstream machine-learning tasks. However, missing data is inevitable and often results in an inaccurate similarity matrix. To address this issue, Similarity Matrix Completion (SMC) methods have been proposed, but they suffer from high computation complexity due to the Singular Value Decomposition (SVD) operation. To reduce the computation complexity, Matrix Factorization (MF) techniques are more explicit and frequently applied to provide a low-rank solution, but the exact low-rank optimal solution can not be guaranteed since it suffers from a non-convex structure. In this paper, we introduce a novel SMC framework that offers a more reliable and efficient solution. Specifically, beyond simply utilizing the unique Positive Semi-definiteness (PSD) property to guide the completion process, our approach further complements a carefully designed rank-minimization regularizer, aiming to achieve an optimal and low-rank solution. Based on the key insights that the underlying PSD property and Low-Rank property improve the SMC performance, we present two novel, scalable, and effective algorithms, SMCNN and SMCNmF, which investigate the PSD property to guide the estimation process and incorporate nonconvex low-rank regularizer to ensure the low-rank solution. Theoretical analysis ensures better estimation performance and convergence speed. Empirical results on real-world datasets demonstrate the superiority and efficiency of our proposed methods compared to various baseline methods.

cs.LG↗

arXiv2024

Direct Alignment of Language Models via Quality-Aware Self-Refinement

Reinforcement Learning from Human Feedback (RLHF) has been commonly used to align the behaviors of Large Language Models (LLMs) with human preferences. Recently, a popular alternative is Direct Policy Optimization (DPO), which replaces an LLM-based reward model with the policy itself, thus obviating the need for extra memory and training time to learn the reward model. However, DPO does not consider the relative qualities of the positive and negative responses, and can lead to sub-optimal training outcomes. To alleviate this problem, we investigate the use of intrinsic knowledge within the on-the-fly fine-tuning LLM to obtain relative qualities and help to refine the loss function. Specifically, we leverage the knowledge of the LLM to design a refinement function to estimate the quality of both the positive and negative responses. We show that the constructed refinement function can help self-refine the loss function under mild assumptions. The refinement function is integrated into DPO and its variant Identity Policy Optimization (IPO). Experiments across various evaluators indicate that they can improve the performance of the fine-tuned models over DPO and IPO.

cs.CL↗

arXiv2022

Image-Adaptive YOLO for Object Detection in Adverse Weather Conditions

Though deep learning-based object detection methods have achieved promising results on the conventional datasets, it is still challenging to locate objects from the low-quality images captured in adverse weather conditions. The existing methods either have difficulties in balancing the tasks of image enhancement and object detection, or often ignore the latent information beneficial for detection. To alleviate this problem, we propose a novel Image-Adaptive YOLO (IA-YOLO) framework, where each image can be adaptively enhanced for better detection performance. Specifically, a differentiable image processing (DIP) module is presented to take into account the adverse weather conditions for YOLO detector, whose parameters are predicted by a small convolutional neural net-work (CNN-PP). We learn CNN-PP and YOLOv3 jointly in an end-to-end fashion, which ensures that CNN-PP can learn an appropriate DIP to enhance the image for detection in a weakly supervised manner. Our proposed IA-YOLO approach can adaptively process images in both normal and adverse weather conditions. The experimental results are very encouraging, demonstrating the effectiveness of our proposed IA-YOLO method in both foggy and low-light scenarios.

cs.CV↗

arXiv2022

DO-GAN: A Double Oracle Framework for Generative Adversarial Networks

In this paper, we propose a new approach to train Generative Adversarial Networks (GANs) where we deploy a double-oracle framework using the generator and discriminator oracles. GAN is essentially a two-player zero-sum game between the generator and the discriminator. Training GANs is challenging as a pure Nash equilibrium may not exist and even finding the mixed Nash equilibrium is difficult as GANs have a large-scale strategy space. In DO-GAN, we extend the double oracle framework to GANs. We first generalize the players' strategies as the trained models of generator and discriminator from the best response oracles. We then compute the meta-strategies using a linear program. For scalability of the framework where multiple generators and discriminator best responses are stored in the memory, we propose two solutions: 1) pruning the weakly-dominated players' strategies to keep the oracles from becoming intractable; 2) applying continual learning to retain the previous knowledge of the networks. We apply our framework to established GAN architectures such as vanilla GAN, Deep Convolutional GAN, Spectral Normalization GAN and Stacked GAN. Finally, we conduct experiments on MNIST, CIFAR-10 and CelebA datasets and show that DO-GAN variants have significant improvements in both subjective qualitative evaluation and quantitative metrics, compared with their respective GAN architectures.

cs.LG↗

arXiv2021

RMIX: Learning Risk-Sensitive Policies for Cooperative Reinforcement Learning Agents

Current value-based multi-agent reinforcement learning methods optimize individual Q values to guide individuals' behaviours via centralized training with decentralized execution (CTDE). However, such expected, i.e., risk-neutral, Q value is not sufficient even with CTDE due to the randomness of rewards and the uncertainty in environments, which causes the failure of these methods to train coordinating agents in complex environments. To address these issues, we propose RMIX, a novel cooperative MARL method with the Conditional Value at Risk (CVaR) measure over the learned distributions of individuals' Q values. Specifically, we first learn the return distributions of individuals to analytically calculate CVaR for decentralized execution. Then, to handle the temporal nature of the stochastic outcomes during executions, we propose a dynamic risk level predictor for risk level tuning. Finally, we optimize the CVaR policies with CVaR values used to estimate the target in TD error during centralized training and the CVaR values are used as auxiliary local rewards to update the local distribution via Quantile Regression loss. Empirically, we show that our method significantly outperforms state-of-the-art methods on challenging StarCraft II tasks, demonstrating enhanced coordination and improved sample efficiency.

cs.LG↗

arXiv2020

Personalized Adaptive Meta Learning for Cold-start User Preference Prediction

A common challenge in personalized user preference prediction is the cold-start problem. Due to the lack of user-item interactions, directly learning from the new users' log data causes serious over-fitting problem. Recently, many existing studies regard the cold-start personalized preference prediction as a few-shot learning problem, where each user is the task and recommended items are the classes, and the gradient-based meta learning method (MAML) is leveraged to address this challenge. However, in real-world application, the users are not uniformly distributed (i.e., different users may have different browsing history, recommended items, and user profiles. We define the major users as the users in the groups with large numbers of users sharing similar user information, and other users are the minor users), existing MAML approaches tend to fit the major users and ignore the minor users. To address this cold-start task-overfitting problem, we propose a novel personalized adaptive meta learning approach to consider both the major and the minor users with three key contributions: 1) We are the first to present a personalized adaptive learning rate meta-learning approach to improve the performance of MAML by focusing on both the major and minor users. 2) To provide better personalized learning rates for each user, we introduce a similarity-based method to find similar users as a reference and a tree-based method to store users' features for fast search. 3) To reduce the memory usage, we design a memory agnostic regularizer to further reduce the space complexity to constant while maintain the performance. Experiments on MovieLens, BookCrossing, and real-world production datasets reveal that our method outperforms the state-of-the-art methods dramatically for both the minor and major users.

cs.IR↗

arXiv2020

Learning to Collaborate in Multi-Module Recommendation via Multi-Agent Reinforcement Learning without Communication

With the rise of online e-commerce platforms, more and more customers prefer to shop online. To sell more products, online platforms introduce various modules to recommend items with different properties such as huge discounts. A web page often consists of different independent modules. The ranking policies of these modules are decided by different teams and optimized individually without cooperation, which might result in competition between modules. Thus, the global policy of the whole page could be sub-optimal. In this paper, we propose a novel multi-agent cooperative reinforcement learning approach with the restriction that different modules cannot communicate. Our contributions are three-fold. Firstly, inspired by a solution concept in game theory named correlated equilibrium, we design a signal network to promote cooperation of all modules by generating signals (vectors) for different modules. Secondly, an entropy-regularized version of the signal network is proposed to coordinate agents' exploration of the optimal global policy. Furthermore, experiments based on real-world e-commerce data demonstrate that our algorithm obtains superior performance over baselines.

cs.LG↗