arXiv subjects

Zihe Liu

Publications and source records attributed to Zihe Liu.

18 recordsLinked to original sources

arXiv2026

PolyWorkBench: Benchmarking LLM Agents for Cross-Lingual Long-Horizon Workflows

While Large Language Model (LLM) agents excel at monolingual long-horizon planning and tool use, enterprise workflows inherently require processing multilingual resources across extended trajectories. The interaction between multilinguality and long-horizon execution, however, remains underexplored. We introduce PolyWorkBench, a benchmark designed to evaluate LLM agents on multilingual, long-horizon workplace workflows. PolyWorkBench features 67 tasks across five core domains: commerce, knowledge work, legal analysis, localization, and manufacturing. Tasks are authored by the paper's authors from real-world data seeds and independently verified through a second-author audit. Agents must integrate heterogeneous multilingual inputs, execute iterative tool-use trajectories, and produce structured domain artifacts. To rigorously assess performance, we adopt Grade, a task-specific structural scoring rubric, as our primary ranking metric, and complement it with Pytest for executable state verification and LLM-as-Judge for semantic quality diagnostics. Benchmark evaluations reveal that agent performance varies substantially across languages and drops sharply on the harder cross-lingual tasks, and our analysis shows that multilingual execution exposes systematic failure modes across planning, tool interaction, and decision-making in long-horizon agents.

cs.AI↗

arXiv2026

Collision and coalescence dynamics of bosonic quantum Hall droplets

Recently bosonic quantum Hall droplets have been observed in rapidly rotating two-dimensional Bose-Einstein condensates (BECs), which exhibit robust dynamical stability. Inspired by this, we systematically investigate the collision and coalescence dynamics of these droplets within the Gross-Pitaevskii framework. For two-droplet collisions, we find two distinct collision outcomes, namely merging and separation, that are controlled by the initial relative velocity. The critical velocity exhibits a universal scaling law with the interaction and the particle number as $v_c \propto (gN)^{1/4}$, which can be interpreted from a simplified analytical model, revealing the essential role of the collision time. It differs fundamentally from the mechanism governing the conventional Lee-Huang-Yang stabilized quantum droplets. Furthermore, while the collision can change the shape of the droplet significantly, the center of mass trajectory remains nearly unaffected, owing to the conservation of angular momentum. For overlapping stationary droplets, vortex arrays can emerge through Kelvin-Helmholtz instability driven by phase-induced shear flow. Although two droplets may merge into a larger one, extended states cannot be constructed from multiple overlapping droplets. Instead, the system dynamically reorganizes into new isolated droplets, revealing the localized property in the bulk region. Our results reveal the unique nonequilibrium dynamics of quantum Hall droplets and suggest new pathways for manipulating strongly correlated rotating quantum fluids.

cond-mat.quant-gas↗

arXiv2026

CAR-EnKF: A Covariance-Adaptive and Recalibrated Ensemble Kalman Filter Framework

The ensemble Kalman filter (EnKF) is widely used for nonlinear and high-dimensional state estimation because it replaces complex covariance propagation with simple ensemble statistics. However, conventional EnKF implementations can become overconfident in the presence of measurement nonlinearity. The commonly used covariance inflation technique only partially alleviates this issue. This paper proposes a covariance-adaptive and recalibrated ensemble Kalman filter (CAR-EnKF) framework for nonlinear state estimation. The framework introduces two improvements that are only active for nonlinear measurements and reduce to the conventional EnKF framework without covariance inflation in the linear case: (i) a recalibration mechanism that reassesses the effect of the chosen Kalman gain after updating the ensemble mean, and (ii) a positive semidefinite covariance compensation term that accounts for measurement nonlinearity. An adaptive update law based on the normalized innovation squared further tunes the compensation magnitude online. The framework is algorithmically general and is specialized here to the stochastic EnKF and the ensemble transform Kalman filter (ETKF). Experiments on feature-based SLAM and the Lorenz--96 system show that CAR-EnKF consistently reduces RMSE relative to conventional EnKF baselines, with especially large improvements at low measurement-noise levels. The related codes are available at \href{https://github.com/Shida-Jiang/CAR-EnKF-A-Covariance-Adaptive-and-Recalibrated-Ensemble-Kalman-Filter-Framework}

eess.SY↗

arXiv2026

Multi-Aspect Knowledge Distillation for Language Model with Low-rank Factorization

Knowledge distillation is an effective technique for pre-trained language model compression. However, existing methods only focus on the knowledge distribution among layers, which may cause the loss of fine-grained information in the alignment process. To address this issue, we introduce the Multi-aspect Knowledge Distillation (MaKD) method, which mimics the self-attention and feed-forward modules in greater depth to capture rich language knowledge information at different aspects. Experimental results demonstrate that MaKD can achieve competitive performance compared with various strong baselines with the same storage parameter budget. In addition, our method also performs well in distilling auto-regressive architecture models.

cs.CL↗

arXiv2026

AutoTour: Automatic Photo Tour Guide with Smartphones and LLMs

We present AutoTour, a system that enhances user exploration by automatically generating fine-grained landmark annotations and descriptive narratives for photos captured by users. The key idea of AutoTour is to fuse visual features extracted from photos with nearby geospatial features queried from open matching databases. Unlike existing tour applications that rely on pre-defined content or proprietary datasets, AutoTour leverages open and extensible data sources to provide scalable and context-aware photo-based guidance. To achieve this, we design a training-free pipeline that first extracts and filters relevant geospatial features around the user's GPS location. It then detects major landmarks in user photos through VLM-based feature detection and projects them into the horizontal spatial plane. A geometric matching algorithm aligns photo features with corresponding geospatial entities based on their estimated distance and direction. The matched features are subsequently grounded and annotated directly on the original photo, accompanied by large language model-generated textual and audio descriptions to provide an informative, tour-like experience. We demonstrate that AutoTour can deliver rich, interpretable annotations for both iconic and lesser-known landmarks, enabling a new form of interactive, context-aware exploration that bridges visual perception and geospatial understanding.

cs.HC↗

arXiv2026

Supervised and Unsupervised Neural Network Solver for First Order Hyperbolic Nonlinear PDEs

We present a neural network-based method for learning scalar hyperbolic conservation laws. Our method replaces the traditional numerical flux in finite volume schemes with a trainable neural network while preserving the conservative structure of the scheme. The model can be trained both in a supervised setting with efficiently generated synthetic data or in an unsupervised manner, leveraging the weak formulation of the partial differential equation. We provide theoretical results that our model can perform arbitrarily well, and provide associated upper bounds on neural network size. Extensive experiments demonstrate that our method often outperforms efficient schemes such as Godunov's scheme, WENO, and Discontinuous Galerkin for comparable computational budgets. Finally, we demonstrate the effectiveness of our method on a traffic prediction task, leveraging field experimental highway data from the Berkeley DeepDrive drone dataset.

math.NA↗

arXiv2025

Part II: ROLL Flash -- Accelerating RLVR and Agentic Training with Asynchrony

Synchronous Reinforcement Learning (RL) post-training has emerged as a crucial step for enhancing Large Language Models (LLMs) with diverse capabilities. However, many systems designed to accelerate RL post-training still suffer from low resource utilization and limited scalability. We present ROLL Flash, a system that extends ROLL with native support for asynchronous RL post-training. ROLL Flash is built upon two core design principles: fine-grained parallelism and rollout-train decoupling. Guided by these principles, ROLL Flash provides flexible programming interfaces that enable a fully asynchronous training architecture and support efficient rollout mechanisms, including queue scheduling and environment-level asynchronous execution. Through comprehensive theoretical analysis and extensive experiments, we demonstrate that ROLL Flash significantly improves resource utilization and scalability over synchronous RL post-training. ROLL Flash achieves up to 2.24x speedup on RLVR tasks and 2.72x on agentic tasks, using the same GPU budget as synchronous baselines. Furthermore, we implement several popular off-policy algorithms and verify that asynchronous training can achieve performance on par with synchronous training.

cs.LG↗

arXiv2025

Part I: Tricks or Traps? A Deep Dive into RL for LLM Reasoning

Reinforcement learning for LLM reasoning has rapidly emerged as a prominent research area, marked by a significant surge in related studies on both algorithmic innovations and practical applications. Despite this progress, several critical challenges remain, including the absence of standardized guidelines for employing RL techniques and a fragmented understanding of their underlying mechanisms. Additionally, inconsistent experimental settings, variations in training data, and differences in model initialization have led to conflicting conclusions, obscuring the key characteristics of these techniques and creating confusion among practitioners when selecting appropriate techniques. This paper systematically reviews widely adopted RL techniques through rigorous reproductions and isolated evaluations within a unified open-source framework. We analyze the internal mechanisms, applicable scenarios, and core principles of each technique through fine-grained experiments, including datasets of varying difficulty, model sizes, and architectures. Based on these insights, we present clear guidelines for selecting RL techniques tailored to specific setups, and provide a reliable roadmap for practitioners navigating the RL for the LLM domain. Finally, we reveal that a minimalist combination of two techniques can unlock the learning capability of critic-free policies using vanilla PPO loss. The results demonstrate that our simple combination consistently improves performance, surpassing strategies like GRPO and DAPO.

cs.LG↗

arXiv2025

A Robust Incomplete Multimodal Low-Rank Adaptation Approach for Emotion Recognition

Multimodal Emotion Recognition (MER) often encounters incomplete multimodality in practical applications due to sensor failures or privacy protection requirements. While existing methods attempt to address various incomplete multimodal scenarios by balancing the training of each modality combination through additional gradients, these approaches face a critical limitation: training gradients from different modality combinations conflict with each other, ultimately degrading the performance of the final prediction model. In this paper, we propose a unimodal decoupled dynamic low-rank adaptation method based on modality combinations, named MCULoRA, which is a novel framework for the parameter-efficient training of incomplete multimodal learning models. MCULoRA consists of two key modules, modality combination aware low-rank adaptation (MCLA) and dynamic parameter fine-tuning (DPFT). The MCLA module effectively decouples the shared information from the distinct characteristics of individual modality combinations. The DPFT module adjusts the training ratio of modality combinations based on the separability of each modality's representation space, optimizing the learning efficiency across different modality combinations. Our extensive experimental evaluation in multiple benchmark datasets demonstrates that MCULoRA substantially outperforms previous incomplete multimodal learning approaches in downstream task accuracy.

cs.CV↗

arXiv2025

Reinforcement Learning Optimization for Large-Scale Learning: An Efficient and User-Friendly Scaling Library

We introduce ROLL, an efficient, scalable, and user-friendly library designed for Reinforcement Learning Optimization for Large-scale Learning. ROLL caters to three primary user groups: tech pioneers aiming for cost-effective, fault-tolerant large-scale training, developers requiring flexible control over training workflows, and researchers seeking agile experimentation. ROLL is built upon several key modules to serve these user groups effectively. First, a single-controller architecture combined with an abstraction of the parallel worker simplifies the development of the training pipeline. Second, the parallel strategy and data transfer modules enable efficient and scalable training. Third, the rollout scheduler offers fine-grained management of each sample's lifecycle during the rollout stage. Fourth, the environment worker and reward worker support rapid and flexible experimentation with agentic RL algorithms and reward designs. Finally, AutoDeviceMapping allows users to assign resources to different models flexibly across various stages.

cs.LG↗

arXiv2024

Efficient Model Compression for Bayesian Neural Networks

Model Compression has drawn much attention within the deep learning community recently. Compressing a dense neural network offers many advantages including lower computation cost, deployability to devices of limited storage and memories, and resistance to adversarial attacks. This may be achieved via weight pruning or fully discarding certain input features. Here we demonstrate a novel strategy to emulate principles of Bayesian model selection in a deep learning setup. Given a fully connected Bayesian neural network with spike-and-slab priors trained via a variational algorithm, we obtain the posterior inclusion probability for every node that typically gets lost. We employ these probabilities for pruning and feature selection on a host of simulated and real-world benchmark data and find evidence of better generalizability of the pruned model in all our experiments.

cs.LG↗

arXiv2024

Bayesian Smoothing and Feature Selection Using variational Automatic Relevance Determination

This study introduces Variational Automatic Relevance Determination (VARD), a novel approach tailored for fitting sparse additive regression models in high-dimensional settings. VARD distinguishes itself by its ability to independently assess the smoothness of each feature while enabling precise determination of whether a feature's contribution to the response is zero, linear, or nonlinear. Further, an efficient coordinate descent algorithm is introduced to implement VARD. Empirical evaluations on simulated and real-world data underscore VARD's superiority over alternative variable selection methods for additive models.

stat.ME↗

arXiv2024

Textualized Agent-Style Reasoning for Complex Tasks by Multiple Round LLM Generation

Chain-of-thought prompting significantly boosts the reasoning ability of large language models but still faces three issues: hallucination problem, restricted interpretability, and uncontrollable generation. To address these challenges, we present AgentCOT, a llm-based autonomous agent framework, which can solve complex problems in an agent-style manner by multiple round LLM generation. At each step, AgentCOT selects an action and executes it to yield an intermediate result with supporting evidence. In addition, we integrate the step's index into the reasoning process to form a graph structure for complex inference logic. We introduce two new strategies to enhance the performance of AgentCOT.We conduct extensive experiments to verify the effectiveness of our method on six common benchmarks. Results exhibit that our method brings in substantial improvements over current competitive approaches.

cs.CL↗

arXiv2024

NGD converges to less degenerate solutions than SGD

The number of free parameters, or dimension, of a model is a straightforward way to measure its complexity: a model with more parameters can encode more information. However, this is not an accurate measure of complexity: models capable of memorizing their training data often generalize well despite their high dimension. Effective dimension aims to more directly capture the complexity of a model by counting only the number of parameters required to represent the functionality of the model. Singular learning theory (SLT) proposes the learning coefficient $ \lambda $ as a more accurate measure of effective dimension. By describing the rate of increase of the volume of the region of parameter space around a local minimum with respect to loss, $ \lambda $ incorporates information from higher-order terms. We compare $ \lambda $ of models trained using natural gradient descent (NGD) and stochastic gradient descent (SGD), and find that those trained with NGD consistently have a higher effective dimension for both of our methods: the Hessian trace $ \text{Tr}(\mathbf{H}) $, and the estimate of the local learning coefficient (LLC) $ \hat{\lambda}(w^*) $.

cs.LG↗

arXiv2024

A Behavior-Aware Approach for Deep Reinforcement Learning in Non-stationary Environments without Known Change Points

Deep reinforcement learning is used in various domains, but usually under the assumption that the environment has stationary conditions like transitions and state distributions. When this assumption is not met, performance suffers. For this reason, tracking continuous environmental changes and adapting to unpredictable conditions is challenging yet crucial because it ensures that systems remain reliable and flexible in practical scenarios. Our research introduces Behavior-Aware Detection and Adaptation (BADA), an innovative framework that merges environmental change detection with behavior adaptation. The key inspiration behind our method is that policies exhibit different global behaviors in changing environments. Specifically, environmental changes are identified by analyzing variations between behaviors using Wasserstein distances without manually set thresholds. The model adapts to the new environment through behavior regularization based on the extent of changes. The results of a series of experiments demonstrate better performance relative to several current algorithms. This research also indicates significant potential for tackling this long-standing challenge.

cs.LG↗

arXiv2022

Self-supervised Group Meiosis Contrastive Learning for EEG-Based Emotion Recognition

The progress of EEG-based emotion recognition has received widespread attention from the fields of human-machine interactions and cognitive science in recent years. However, how to recognize emotions with limited labels has become a new research and application bottleneck. To address the issue, this paper proposes a Self-supervised Group Meiosis Contrastive learning framework (SGMC) based on the stimuli consistent EEG signals in human being. In the SGMC, a novel genetics-inspired data augmentation method, named Meiosis, is developed. It takes advantage of the alignment of stimuli among the EEG samples in a group for generating augmented groups by pairing, cross exchanging, and separating. And the model adopts a group projector to extract group-level feature representations from group EEG samples triggered by the same emotion video stimuli. Then contrastive learning is employed to maximize the similarity of group-level representations of augmented groups with the same stimuli. The SGMC achieves the state-of-the-art emotion recognition results on the publicly available DEAP dataset with an accuracy of 94.72% and 95.68% in valence and arousal dimensions, and also reaches competitive performance on the public SEED dataset with an accuracy of 94.04%. It is worthy of noting that the SGMC shows significant performance even when using limited labels. Moreover, the results of feature visualization suggest that the model might have learned video-level emotion-related feature representations to improve emotion recognition. And the effects of group size are further evaluated in the hyper parametric analysis. Finally, a control experiment and ablation study are carried out to examine the rationality of architecture. The code is provided publicly online.

eess.SP↗

arXiv2021

Relation-aware Hierarchical Attention Framework for Video Question Answering

Video Question Answering (VideoQA) is a challenging video understanding task since it requires a deep understanding of both question and video. Previous studies mainly focus on extracting sophisticated visual and language embeddings, fusing them by delicate hand-crafted networks. However, the relevance of different frames, objects, and modalities to the question are varied along with the time, which is ignored in most of existing methods. Lacking understanding of the the dynamic relationships and interactions among objects brings a great challenge to VideoQA task. To address this problem, we propose a novel Relation-aware Hierarchical Attention (RHA) framework to learn both the static and dynamic relations of the objects in videos. In particular, videos and questions are embedded by pre-trained models firstly to obtain the visual and textual features. Then a graph-based relation encoder is utilized to extract the static relationship between visual objects. To capture the dynamic changes of multimodal objects in different video frames, we consider the temporal, spatial, and semantic relations, and fuse the multimodal features by hierarchical attention mechanism to predict the answer. We conduct extensive experiments on a large scale VideoQA dataset, and the experimental results demonstrate that our RHA outperforms the state-of-the-art methods.

cs.CV↗

arXiv2020

Frame Aggregation and Multi-Modal Fusion Framework for Video-Based Person Recognition

Video-based person recognition is challenging due to persons being blocked and blurred, and the variation of shooting angle. Previous research always focused on person recognition on still images, ignoring similarity and continuity between video frames. To tackle the challenges above, we propose a novel Frame Aggregation and Multi-Modal Fusion (FAMF) framework for video-based person recognition, which aggregates face features and incorporates them with multi-modal information to identify persons in videos. For frame aggregation, we propose a novel trainable layer based on NetVLAD (named AttentionVLAD), which takes arbitrary number of features as input and computes a fixed-length aggregation feature based on feature quality. We show that introducing an attention mechanism to NetVLAD can effectively decrease the impact of low-quality frames. For the multi-model information of videos, we propose a Multi-Layer Multi-Modal Attention (MLMA) module to learn the correlation of multi-modality by adaptively updating Gram matrix. Experimental results on iQIYI-VID-2019 dataset show that our framework outperforms other state-of-the-art methods.

cs.CV↗