Searcharxiv⌕ Search

arXiv subjects

Zeyi Wen

Publications and source records attributed to Zeyi Wen.

At least 19 recordsLinked to original sources

OfficePrism: Benchmarking Agents on Single-Domain Operations Across Enterprise Business Functions

Broad workplace scores can mix competence within one business function with the extra cost of coordinating across functions and systems, making failures hard to localize. OfficePrism evaluates this prerequisite through 420 single-domain enterprise tasks spanning seven business functions under shared, benchmark-defined L1--L3 task-structure levels. Each task runs in a fresh environment combining a shared synthetic organization with one function-specific domain and yped state access. Our pipeline uses grounded records, business rules, and tool schemas to jointly construct a Chinese-language ticket, a hidden executable reference witness, and an expected response, followed by automated filtering and human curation. The same task set is evaluated through full-schema and search-and-dispatch packages. Across 124 model--harness--package configurations, the per-case joint natural-language-and-database pass rate ranges from 22.6% to 76.4%, with a median of 67.9%. L1 is highest in all 124 configurations, and L2 exceeds L3 in 122. Among 50 configurations sharing the same harness and model labels across both packages, search scores higher in 30, ties in one, and scores lower in 19.

cs.CL↗

OOM: Offload-Optimized Full-Parameter MoE Training via Memory-Execution Co-Design

Offloading model state to host memory and streaming layer weights let a single multi-GPU node train large mixture-of-experts (MoE) models. GPU memory then becomes a performance budget, shared by resident expert state, retained activations, and execution worksets. The best allocation depends on the expert execution path, batch size, and sequence length. We present OOM (Offload-Optimized MoE), a full-parameter MoE training system that co-designs expert execution and memory allocation. For expert-parallel execution, OOM updates resident experts on their owning GPUs and jointly plans state residency and MoE activation retention according to pipeline dependencies. For local execution, OOM bounds GPU worksets through dependency-aware partitioning; in hybrid models, it also retains attention results on the host to avoid costly recomputation and selectively regenerates layer boundaries through cheaper layers to make room. Calibrated cost models guide joint configuration within GPU and host budgets to minimize step time; OOM selects among the configured paths through short executions. On four H100s, OOM trains with 1M-token contexts, whereas FSDP2, ZeRO-Offload, and MegaTrain exhaust GPU memory at 512K. It also trains a 122B-parameter MoE model, on which all three exhaust host memory. On four 24 GB RTX 4090s, it trains 512K-token contexts and 1,024 1K-token sequences per step without gradient accumulation. Across matched-global-batch comparisons, OOM delivers 1.58x and 4.08x geometric-mean (up to 2.14x and 7.77x) throughput over CPU-offloaded FSDP2 on the two platforms while using the least host memory among the evaluated systems -- as little as 1/3.0 of FSDP2's.

cs.DC↗

EHR-Bench: A Chinese Multi-Task Benchmark for Recruitment AI

Recruitment applications span diverse operations, from drafting job descriptions and interpreting queries to organizing profiles, assessing fit, and recommending candidates and jobs. Systematic LLM evaluation therefore requires task coverage that reflects this range of business needs. We introduce EHR-Bench, an open Chinese benchmark comprising twelve tasks across six capability families and five recruitment scenarios. The tasks cover generation, extraction, classification, retrieval, semantic similarity, and reciprocal ranking within one evaluation framework. An internal recruitment dataset guides task definitions, the occupation taxonomy and sampling targets, output requirements, and scoring protocols. Following these specifications, we construct 8,336 task instances from public materials and controlled augmentation, with source-aware expert review and documented provenance. We evaluate twelve LLMs through task-level, capability-level, and aggregate comparisons. Closely ranked models exhibit distinct task strengths, and performance varies across input sources, reference evidence, and candidate pools. Five shared models retain the same observed Overall ordering on a private operational reference, with variation in capability and task correspondence. EHR-Bench provides a reproducible foundation for systematic model assessment, task-specific model selection, and further research on LLMs for recruitment. The project repository is available at https://github.com/EhrBench/EHR-Bench.

cs.CL↗

SlideDP: Scaling Host-Resident LLM Fine-Tuning Across Multiple GPUs

Host-resident layer streaming enables full-parameter LLM fine-tuning beyond GPU memory, but data-parallel ranks compete for shared host resources. Replicated transfers amplify traffic, while strong scaling can expose host work as computation windows shrink. We present SlideDP, a synchronous data-parallel runtime for shared-host multi-GPU systems. It maintains one authoritative host state, decouples communication routes from state layout, and pipelines parameter delivery, gradient aggregation, and CPU updates across ranks and chunks. An analytical step-time model characterizes resource bottlenecks and pipeline exposure; runtime measurements guide communication, chunking, and activation policies under a GPU memory budget. In matched-batch sweeps, SlideDP achieves geometric-mean throughput ratios of 1.46-2.64$\times$ over SlideFormer, MegaTrain, and ZeRO-Offload. On four H100s, SlideDP approaches GPU-resident FSDP2 throughput for Qwen3-14B at a smaller batch size. With a larger batch, it processes over 1M tokens per step and exceeds FSDP2's measured peak throughput by 11.2%. Separately, it supports 256K-token sequences for the same model and fine-tunes Qwen2.5-72B on four RTX 4090 GPUs. Project page: https://github.com/RegiaYoung/SlideDP.

cs.DC↗

TRACE: Single-Pass Decoding-Trace Risk Localization for Generation Calibration

Reliable confidence estimation is essential for large language model deployment. However, answer-level calibration remains challenging because generation errors are often localized: a response may be fluent and high-probability overall while still failing at a critical number, entity, or factual claim. Existing estimators compress token probabilities, sequence likelihoods, entropy, or beam statistics into a global score, which can dilute such local risk signals. We propose TRACE, a single-pass, decoded-answer-preserving confidence estimator that treats decoding-time uncertainty as a trajectory through three steps: (i) recording token-level surprisal and predictive entropy during decoding, (ii) applying local risk operators to preserve uncertainty spikes, and (iii) converting localized trace risk into answer-level confidence. TRACE produces a label-free risk score, while TRACE+ calibrates trace-only features into probabilities using a held-out split, without extra generations or external verifiers. We evaluate four tasks against 19 calibration baselines, and TRACE+ reduces Brier from 0.149 to 0.137 and improves AUROC from 0.758 to 0.792 over the strongest likelihood baseline. Across seven LLMs, TRACE+ improves over the best non-TRACE baseline pool from 0.136 to 0.120 Brier and from 0.764 to 0.817 AUROC. Results show that localizing decoding-time risk provides a general approach to calibration.

cs.CL↗

Strong Drafts Need Compact Memories: Long-Context Speculative Decoding with Compressed KV Cache

Long-context LLM applications such as document summarization and multi-turn agents require generation from prefixes spanning tens of thousands of tokens, making decoding latency a major bottleneck. Speculative decoding (SD) reduces latency without changing model outputs, but its speedup depends on both accepted draft tokens and draft-step latency: Lightweight drafts are fast but lack the capacity to capture long-range dependencies, whereas strong independent drafts recover acceptance but incur growing KV-access cost at long prefixes. We introduce memory-augmented drafting for long-context SD, equipping a strong independent draft with compressed draft-side KV memory: A lightweight adaptor constructs and incrementally updates this memory to retain distant information and exact recent context. The target verifier retains its full KV cache and applies the standard accept/reject rule, preserving SD's lossless guarantee. Experiments on Llama~3.1-8B and 70B targets at prefix lengths up to 32K show that our method reduces draft-side memory by over 70%. It achieves speedups of up to 2.08x and 3.33x , respectively, over autoregressive decoding.

cs.LG↗

EntSQL: A Benchmark for Grounding Text-to-SQL in Long-Context Enterprise Knowledge

Text-to-SQL enables natural language access to databases, and recent LLMs have substantially advanced its capabilities. Existing benchmarks such as Spider, BIRD, and Spider~2.0 evaluate schema generalization, large-scale databases, and realistic workflows, but largely overlook enterprise scenarios where SQL generation depends on private business knowledge, such as internal metrics, reporting conventions, and organizational rules. We introduce EntSQL, an enterprise-oriented Text-to-SQL benchmark for evaluating long-context grounding over proprietary business documents. EntSQL contains 1,066 aligned Chinese-English semantic examples across five business domains, with most examples requiring domain knowledge beyond the question and schema and involving complex SQL structures. On English inputs, the best evaluated system reaches only 15.9\% when long-form documents are provided, highlighting the difficulty of grounding SQL generation in enterprise knowledge.

cs.CL↗

An Efficient Heterogeneous Co-Design for Fine-Tuning on a Single GPU

Fine-tuning Large Language Models (LLMs) has become essential for domain adaptation, but its memory-intensive property exceeds the capabilities of most GPUs. To address this challenge and democratize LLM fine-tuning, we present SlideFormer, a novel system designed for single-GPU environments. Our innovations are: (1) A lightweight asynchronous engine that treats the GPU as a sliding window and overlaps GPU computation with CPU updates and multi-tier I/O. (2) A highly efficient heterogeneous memory management scheme significantly reduces peak memory usage. (3) Optimized Triton kernels to solve key bottlenecks and integrated advanced I/O. This collaborative design enables fine-tuning of the latest 123B+ models on a single RTX 4090, supporting up to 8x larger batch sizes and 6x larger models. In evaluations, SlideFormer achieves 1.40x to 6.27x higher throughput while roughly halving CPU/GPU memory usage compared to baselines, sustaining >95% peak performance on both NVIDIA and AMD GPUs.The code is available at https://github.com/RegiaYoung/SlideFormer.

cs.DC↗

Preserving Data Privacy in Learning Causal Structure with Fully Homomorphic Encryption

Preserving data privacy is an important topic in structural data management and data mining. However, the issue of privacy leakage in distributed causal structure learning is a persistent challenge, especially in cases where data transmission and computation are required. In this paper, we propose a method based on fully homomorphic encryption (FHE) that performs calculations on ciphertexts, keeping data encrypted in transition and computation. Nevertheless, adopting FHE to causal structure learning is challenging due to the high computation cost and limited support on division as well as logarithm operations in FHE. To tackle this challenge, we propose a series of novel techniques including (i) circuit simplification for better efficiency, (ii) approximation of division and logarithm through Newton-Raphson Reciprocal and Taylor expansion, and (iii) a batching technique with SIMD-acceleration to enhance the whole learning process. Additionally, our method can be easily extended beyond FHE by demonstration of its portability to support differential privacy. Empirical results show that our method achieves high consistency and comparable causal structure with the plaintext version in the datasets tested. Last, our method is efficient and practical to complete learning causal structures in tens of minutes even under the privacy protection of FHE.

cs.CR↗

DeInfer: Efficient Parallel Inferencing for Decomposed Large Language Models

Existing works on large language model (LLM) decomposition mainly focus on improving performance on downstream tasks, but they ignore the poor parallel inference performance when trying to scale up the model size. To mitigate this important performance issue, this paper introduces DeInfer, a high-performance inference system dedicated to parallel inference of decomposed LLMs. It consists of multiple optimizations to maximize performance and be compatible with state-of-the-art optimization techniques. Extensive experiments are carried out to evaluate DeInfer's performance, where the results demonstrate its superiority, suggesting it can greatly facilitate the parallel inference of decomposed LLMs.

cs.CL↗

Efficient Hyperparameter Optimization for LLM Reinforcement Learning

Reinforcement learning (RL) for large language models (LLMs) is highly sensitive to hyperparameter configurations, making hyperparameter optimization (HPO) essential yet computationally expensive. Existing multi-fidelity HPO methods remain inefficient for LLM RL due to the massive model scale and resource-intensive training cycles. In this paper, we propose Joint Fidelity Hyperparameter Optimization (JF-HPO), which simultaneously adapts both model size and training budget as fidelity. JF-HPO is empowered by: (i) it leverages a small proxy model of the target LLM for efficient training and evaluation in each HPO trial; (ii) it integrates carefully designed early-stopping strategies based on training dynamics; (iii) it introduces an efficient checkpointing mechanism to eliminate redundant computations. Compared with existing HPO methods, JF-HPO significantly improves the computational efficiency of each trial (up to 14.9 times), while achieving better or competitive predictive accuracy under the same time budget. Notably, compared with utilizing hyperparameter configurations from the VeRL Recipe, JF-HPO delivers performance improvements ranging from 5.8% to 111.6%.

cs.LG↗

Taxon: Hierarchical Tax Code Prediction with Semantically Aligned LLM Expert Guidance

Tax code prediction is a crucial yet underexplored task in automating invoicing and compliance management for large-scale e-commerce platforms. Each product must be accurately mapped to a node within a multi-level taxonomic hierarchy defined by national standards, where errors lead to financial inconsistencies and regulatory risks. This paper presents Taxon, a semantically aligned and expert-guided framework for hierarchical tax code prediction. Taxon integrates (i) a feature-gating mixture-of-experts architecture that adaptively routes multi-modal features across taxonomy levels, and (ii) a semantic consistency model distilled from large language models acting as domain experts to verify alignment between product titles and official tax definitions. To address noisy supervision in real business records, we design a multi-source training pipeline that combines curated tax databases, invoice validation logs, and merchant registration data to provide both structural and semantic supervision. Extensive experiments on the proprietary TaxCode dataset and public benchmarks demonstrate that Taxon achieves state-of-the-art performance, outperforming strong baselines. Further, an additional full hierarchical paths reconstruction procedure significantly improves structural consistency, yielding the highest overall F1 scores. Taxon has been deployed in production within Alibaba's tax service system, handling an average of over 500,000 tax code queries per day and reaching peak volumes above five million requests during business event with improved accuracy, interpretability, and robustness.

cs.LG↗

Enhancing Online Recruitment with Category-Aware MoE and LLM-based Data Augmentation

Person-Job Fit (PJF) is a critical component for online recruitment. Existing approaches face several challenges, particularly in handling low-quality job descriptions and similar candidate-job pairs, which impair model performance. To address these challenges, this paper proposes a large language model (LLM) based method with two novel techniques: (1) LLM-based data augmentation, which polishes and rewrites low-quality job descriptions by leveraging chain-of-thought (COT) prompts, and (2) category-aware Mixture of Experts (MoE) that assists in identifying similar candidate-job pairs. This MoE module incorporates category embeddings to dynamically assign weights to the experts and learns more distinguishable patterns for similar candidate-job pairs. We perform offline evaluations and online A/B tests on our recruitment platform. Our method relatively surpasses existing methods by 2.40% in AUC and 7.46% in GAUC, and boosts click-through conversion rate (CTCVR) by 19.4% in online tests, saving millions of CNY in external headhunting expenses.

cs.AI↗

DOne: Decoupling Structure and Rendering for High-Fidelity Design-to-Code Generation

While Vision Language Models (VLMs) have shown promise in Design-to-Code generation, they suffer from a "holistic bottleneck-failing to reconcile high-level structural hierarchy with fine-grained visual details, often resulting in layout distortions or generic placeholders. To bridge this gap, we propose DOne, an end-to-end framework that decouples structure understanding from element rendering. DOne introduces (1) a learned layout segmentation module to decompose complex designs, avoiding the limitations of heuristic cropping; (2) a specialized hybrid element retriever to handle the extreme aspect ratios and densities of UI components; and (3) a schema-guided generation paradigm that bridges layout and code. To rigorously assess performance, we introduce HiFi2Code, a benchmark featuring significantly higher layout complexity than existing datasets. Extensive evaluations on the HiFi2Code demonstrate that DOne outperforms exiting methods in both high-level visual similarity (e.g., over 10% in GPT Score) and fine-grained element alignment. Human evaluations confirm a 3 times productivity gain with higher visual fidelity.

cs.CV↗

SoLA: Leveraging Soft Activation Sparsity and Low-Rank Decomposition for Large Language Model Compression

Large language models (LLMs) have demonstrated impressive capabilities across various tasks, but the billion-scale parameters pose deployment challenges. Although existing methods attempt to reduce the scale of LLMs, they require either special hardware support or expensive post-training to maintain model quality. To facilitate efficient and affordable model slimming, we propose a novel training-free compression method for LLMs, named "SoLA", which leverages \textbf{So}ft activation sparsity and \textbf{L}ow-r\textbf{A}nk decomposition. SoLA can identify and retain a minority of components significantly contributing to inference, while compressing the majority through low-rank decomposition, based on our analysis of the activation pattern in the feed-forward network (FFN) of modern LLMs. To alleviate the decomposition loss, SoLA is equipped with an adaptive component-wise low-rank allocation strategy to assign appropriate truncation positions for different weight matrices. We conduct extensive experiments on LLaMA-2-7B/13B/70B and Mistral-7B models across a variety of benchmarks. SoLA exhibits remarkable improvement in both language modeling and downstream task accuracy without post-training. For example, with a 30\% compression rate on the LLaMA-2-70B model, SoLA surpasses the state-of-the-art method by reducing perplexity from 6.95 to 4.44 and enhancing downstream task accuracy by 10\%.

cs.CL↗

Quasar: Quantized Self-Speculative Acceleration for Rapid Inference via Memory-Efficient Verification

Speculative Decoding (SD) has emerged as a premier technique for accelerating Large Language Model (LLM) inference by decoupling token generation into rapid drafting and parallel verification. While recent advancements in self-speculation and lookahead decoding have successfully minimized drafting overhead, they have shifted the primary performance bottleneck to the verification phase. Since verification requires a full forward pass of the target model, it remains strictly memory-bandwidth bound, fundamentally limiting the maximum achievable speedup.In this paper, we introduce \textbf{Quasar} (\textbf{Qua}ntized \textbf{S}elf-speculative \textbf{A}cceleration for \textbf{R}apid Inference), a novel, training-free framework designed to overcome this "memory wall" by employing low-bit quantization specifically for the verification stage. Our empirical analysis reveals that while aggressive structural pruning significantly degrades verification accuracy, quantization-based verification preserves the logit distribution with high fidelity while effectively halving memory traffic. Extensive experiments on state-of-the-art models (e.g., OpenPangu and Qwen3) demonstrate that Quasar maintains a speculative acceptance length comparable to full-precision methods while achieving a $1.28\times$ improvement in end-to-end throughput. Being orthogonal to existing drafting strategies, Quasar offers a generic and efficient pathway to accelerate the verification leg of speculative execution. Code is available at https://github.com/Tom-HG/Quasar.

cs.DC↗

Enhancing Talent Search Ranking with Role-Aware Expert Mixtures and LLM-based Fine-Grained Job Descriptions

Talent search is a cornerstone of modern recruitment systems, yet existing approaches often struggle to capture nuanced job-specific preferences, model recruiter behavior at a fine-grained level, and mitigate noise from subjective human judgments. We present a novel framework that enhances talent search effectiveness and delivers substantial business value through two key innovations: (i) leveraging LLMs to extract fine-grained recruitment signals from job descriptions and historical hiring data, and (ii) employing a role-aware multi-gate MoE network to capture behavioral differences across recruiter roles. To further reduce noise, we introduce a multi-task learning module that jointly optimizes click-through rate (CTR), conversion rate (CVR), and resume matching relevance. Experiments on real-world recruitment data and online A/B testing show relative AUC gains of 1.70% (CTR) and 5.97% (CVR), and a 17.29% lift in click-through conversion rate. These improvements reduce dependence on external sourcing channels, enabling an estimated annual cost saving of millions of CNY.

cs.IR↗

SEAL: Structure and Element Aware Learning to Improve Long Structured Document Retrieval

In long structured document retrieval, existing methods typically fine-tune pre-trained language models (PLMs) using contrastive learning on datasets lacking explicit structural information. This practice suffers from two critical issues: 1) current methods fail to leverage structural features and element-level semantics effectively, and 2) the lack of datasets containing structural metadata. To bridge these gaps, we propose \our, a novel contrastive learning framework. It leverages structure-aware learning to preserve semantic hierarchies and masked element alignment for fine-grained semantic discrimination. Furthermore, we release \dataset, a long structured document retrieval dataset with rich structural annotations. Extensive experiments on both released and industrial datasets across various modern PLMs, along with online A/B testing, demonstrate consistent performance improvements, boosting NDCG@10 from 73.96\% to 77.84\% on BGE-M3. The resources are available at https://github.com/xinhaoH/SEAL.

cs.IR↗