Searcharxiv⌕ Search

arXiv subjects

Yi Su

Publications and source records attributed to Yi Su.

At least 19 recordsLinked to original sources
arXiv2026

Bole: Efficient Tree Speculation for Hybrid-Attention Language Models

Hybrid-attention large language models combine full attention with recurrent linear attention to reduce long-context inference costs, yet their autoregressive decoding remains memory-bound. Tree speculative decoding offers an attractive acceleration path, but existing tree-speculation systems are designed around the key--value caches of full-attention models. On hybrid models, they traverse recurrent layers branch by branch and materialize a full state for every proposal node, causing verification latency and transient memory to scale poorly with tree and batch sizes. We present Bole, a kernel--runtime co-design that enables efficient tree speculation for hybrid-attention LLMs. Bole transforms the linear-attention recurrence into a tree-structured closed form and realizes it with a resource-efficient GPU kernel, verifying all proposal nodes in parallel and accelerating linear-attention tree verification by 3.4--7.7$\times$. It losslessly encodes speculative state updates as token-level factors and reconstructs only the state selected after sampling, reducing transient state memory by 82--99$\times$ and freeing GPU capacity for KV caches. Its integration into SGLang, a widely deployed production LLM serving engine, couples efficient state management with a batch-wide verification budget calibrated to the complete hybrid forward. Across four models, two GPU platforms, and diverse datasets, Bole delivers up to $4.72\times$ the offline decode throughput of autoregressive decoding and up to $2.03\times$ that of the strongest tree-speculative baseline. Under online agent workloads, it reduces TTFT and TPOT by up to $67.6%$ and $49.9%$, respectively, over the strongest tree-speculative baseline.

cs.DC↗
arXiv2026

PIVOT: Efficient Query-Group Indexing for Token-Level Sparse Attention

Token-level sparse attention, as implemented by DeepSeek Sparse Attention (DSA) in production systems, makes the downstream attention efficient but shifts the bottleneck to the indexer that feeds it. To select the top-k tokens for each query, the indexer must still score every preceding token, incurring a cost of O(L^2) per layer for a sequence of length L. We observe that this per-query scan is largely redundant: nearby queries select highly overlapping top-k tokens, and the indexer scores are long-tailed along the key axis. We exploit these properties in PIVOT, Proxy Indexing Via One full-prefix Traversal, a training-free, drop-in replacement for the DSA indexer that shares one prefix scan across a group of nearby queries. PIVOT aggregates a group into a single proxy query, performs one shared full-prefix scan to obtain a candidate set, and then selects a top-k for each query from that set. Two variants trade speed for fidelity: PIVOT-Reuse shares the proxy top-k across the group for maximum speed, whereas PIVOT-Refine re-scores the candidate set with the indexer of each query and then selects an individual top-k, matching the dense indexer at a small additional cost. A single algorithm covers both inference phases, differing only in how groups are formed: fixed-size groups of consecutive queries in prefill, and the queries decoded together in one multi-token prediction (MTP) step in decode. On DeepSeek-V3.2 and GLM-5.1 across LongBench and RULER, PIVOT matches the accuracy of the dense DSA indexer while accelerating it by up to 4x and reducing end-to-end latency by up to 1.6x at long context.

cs.CL↗
arXiv2026

AudioDER: A Deduplication-Enhanced Reasoning Dataset for Post-Training Large Audio-Language Models

Recent advances in pretrained large audio-language models (LALMs) have demonstrated strong capabilities across speech, sound, and music. To adapt these models to downstream tasks without the cost of pretraining from scratch, post-training has become a widely adopted paradigm. However, the effectiveness of post-training depends critically on the quality of the training corpus. We observe that existing post-training corpora, often constructed by aggregating public audio datasets, suffer from substantial acoustic redundancy, as many of these datasets are sourced from overlapping media platforms. Such redundancy leads to repeated exposure to similar acoustic patterns, causing diminishing returns in performance despite increased data volume. address this issue, we propose a three-stage data construction pipeline that performs acoustic redundancy filtering, converts retained samples into a unified multiple-choice question-answering format with chain-of-thought generation, and finally applies quality verification and filtering. Using this pipeline, we construct AudioRE, a post-training dataset of approximately 286k instances spanning sound, speech, and music. Supervised fine-tuning on AudioRE consistently improves the performance of Qwen2-Audio-7B-Instruct across diverse audio understanding and reasoning benchmarks, outperforming models trained on the unfiltered raw corpus with substantially more instances. These results validate the effectiveness of our redundancy-aware data construction pipeline and the resulting AudioRE dataset, and further highlight the importance of minimizing acoustic redundancy in audio-language post-training. To facilitate future research, we will release both the AudioRE and the fine-tuned Qwen2-AudioRE checkpoint.

cs.SD↗
arXiv2026

Scalable Network-Aware Experiment Design for Two-Sided Marketplaces

Measuring causal effects in networked two-sided marketplaces is challenging due to treatment interference between market participants on different sides. When treatment is applied to one side (e.g., job seekers), their interactions with the other side (e.g., job posters) introduce spillover effects that violate the Stable Unit Treatment Value Assumption (SUTVA) and bias causal estimates. While cluster-based randomization mitigates this problem, prior approaches struggle with a fundamental trade-off: reducing spillover requires isolated clusters that will reduce the number of qualifying clusters, which decreases statistical power. This paper introduces EgoCluster V3, an iterative clustering algorithm that reduces spillover by 3x compared to prior versions while preserving node coverage and doubling test power. We further introduce MultiEgoCluster, which extends V3 through a two-stage procedure that first groups highly connected egos into multi-ego clusters before applying the iterative clustering algorithm. This achieves an additional ~56% spillover reduction and ~38% increase in sample size. Both methods are deployed in production at LinkedIn and have systematically enabled high-impact two-sided marketplace experiments. Since residual bias cannot be fully eliminated through clustering alone, we derive a theoretical bias correction method for average treatment effect (ATE) estimation based on graph structure and propose an approach to generalize results to the general population.

stat.AP↗
arXiv2026

Kwai Summary Attention Technical Report

Long-context ability, has become one of the most important iteration direction of next-generation Large Language Models, particularly in semantic understanding/reasoning, code agentic intelligence and recommendation system. However, the standard softmax attention exhibits quadratic time complexity with respect to sequence length. As the sequence length increases, this incurs substantial overhead in long-context settings, leading the training and inference costs of extremely long sequences deteriorate rapidly. Existing solutions mitigate this issue through two technique routings: i) Reducing the KV cache per layer, such as from the head-level compression GQA, and the embedding dimension-level compression MLA, but the KV cache remains linearly dependent on the sequence length at a 1:1 ratio. ii) Interleaving with KV Cache friendly architecture, such as local attention SWA, linear kernel GDN, but often involve trade-offs among KV Cache and long-context modeling effectiveness. Besides the two technique routings, we argue that there exists an intermediate path not well explored: {Maintaining a linear relationship between the KV cache and sequence length, but performing semantic-level compression through a specific ratio $k$}. This $O(n/k)$ path does not pursue a ``minimum KV cache'', but rather trades acceptable memory costs for complete, referential, and interpretable retention of long distant dependency. Motivated by this, we propose Kwai Summary Attention (KSA), a novel attention mechanism that reduces sequence modeling cost by compressing historical contexts into learnable summary tokens.

cs.CL↗
arXiv2026

Hierarchical Mesh Transformers with Topology-Guided Pretraining for Morphometric Analysis of Brain Structures

Representation learning on large-scale unstructured volumetric and surface meshes poses significant challenges in neuroimaging, especially when models must incorporate diverse vertex-level morphometric descriptors, such as cortical thickness, curvature, sulcal depth, and myelin content, which carry subtle disease-related signals. Current approaches either ignore these clinically informative features or support only a single mesh topology, restricting their use across imaging pipelines. We introduce a hierarchical transformer framework designed for heterogeneous mesh analysis that operates on spatially adaptive tree partitions constructed from simplicial complexes of arbitrary order. This design accommodates both volumetric and surface discretizations within a single architecture, enabling efficient multi-scale attention without topology-specific modifications. A feature projection module maps variable-length per-vertex clinical descriptors into the spatial hierarchy, separating geometric structure from feature dimensionality and allowing seamless integration of different neuroimaging feature sets. Self-supervised pretraining via masked reconstruction of both coordinates and morphometric channels on large unlabeled cohorts yields a transferable encoder backbone applicable to diverse downstream tasks and mesh modalities. We validate our approach on Alzheimer's disease classification and amyloid burden prediction using volumetric brain meshes from ADNI, as well as focal cortical dysplasia detection on cortical surface meshes from the MELD dataset, achieving state-of-the-art results across all benchmarks.

cs.CV↗
arXiv2026

SeGPruner: Semantic-Geometric Visual Token Pruner for 3D Question Answering

Vision-language models (VLMs) have been widely adopted for 3D question answering (3D QA). In typical pipelines, visual tokens extracted from multiple viewpoints are concatenated with language tokens and jointly processed by a large language model (LLM) for inference. However, aggregating multi-view observations inevitably introduces severe token redundancy, leading to an overly large visual token set that significantly hinders inference efficiency under constrained token budgets. Visual token pruning has emerged as a prevalent strategy to address this issue. Nevertheless, most existing pruners are primarily tailored to 2D inputs or rely on indirect geometric cues, which limits their ability to explicitly retain semantically critical objects and maintain sufficient spatial coverage for robust 3D reasoning. In this paper, we propose SeGPruner, a semantic-aware and geometry-guided token reduction framework for efficient 3D QA with multi-view images. Specifically, SeGPruner first preserves semantically salient tokens through an attention-based importance module (Saliency-aware Token Selector), ensuring that object-critical evidence is retained. It then complements these tokens with spatially diverse ones via a geometry-guided selector (Geometry-aware Token Diversifier), which jointly considers semantic relevance and 3D geometric distance. This cooperation between saliency preservation and geometry-guided diversification balances object-level evidence and global scene coverage under aggressive token reduction. Extensive experiments on ScanQA and OpenEQA demonstrate that SeGPruner substantially improves inference efficiency, reducing the visual token budget by 91% and inference latency by 86%, while maintaining competitive performance in 3D reasoning tasks.

cs.CV↗
arXiv2026

Quantized Inference for OneRec-V2

Quantized inference has demonstrated substantial system-level benefits in large language models while preserving model quality. In contrast, reliably applying low-precision quantization to recommender systems remains challenging in industrial settings. This difficulty arises from differences in training paradigms, architectural patterns, and computational characteristics, which lead to distinct numerical behaviors in weights and activations. Traditional recommender models often exhibit high-magnitude and high-variance weights and activations, making them more sensitive to quantization-induced perturbations. In addition, recommendation workloads frequently suffer from limited hardware utilization, limiting the practical gains of low-precision computation. In this work, we revisit low-precision inference in the context of generative recommendation. Through empirical distribution analysis, we show that the weight and activation statistics of OneRec-V2 are significantly more controlled and closer to those of large language models than traditional recommendation models. Moreover, OneRec-V2 exhibits a more compute-intensive inference pattern with substantially higher hardware utilization, enabling more end-to-end throughput gains with low-precision computation. Leveraging this property, we develop a FP8 post training quantization framework and integrate it into an optimized inference infrastructure. The proposed joint optimization achieves a 49\% reduction in end-to-end inference latency and a 92\% increase in throughput. Extensive online A/B testing further confirms that FP8 inference introduces no degradation in core metrics. These results suggest that as recommender systems evolve toward the paradigms of large language models, algorithm-level and system-level optimization techniques established in the LLM domain can be effectively adapted to large-scale recommendation workloads.

cs.IR↗
arXiv2026

LongFlow: Efficient KV Cache Compression for Reasoning Models

Recent reasoning models such as OpenAI-o1 and DeepSeek-R1 have shown strong performance on complex tasks including mathematical reasoning and code generation. However, this performance gain comes with substantially longer output sequences, leading to significantly increased deployment costs. In particular, long outputs require large KV caches, resulting in high memory consumption and severe bandwidth pressure during attention computation. Most existing KV cache optimization methods are designed for long-input, short-output scenarios and are ineffective for the long-output setting of reasoning models. Moreover, importance estimation in prior work is computationally expensive and becomes prohibitive when continuous re-evaluation is required during long generation. To address these challenges, we propose LongFlow, a KV cache compression method with an efficient importance estimation metric derived from an intermediate result of attention computation using only the current query. This design introduces negligible computational overhead and requires no auxiliary storage. We further develop a custom kernel that fuses FlashAttention, importance estimation, and token eviction into a single optimized operator, improving system-level efficiency. Experiments show that LongFlow achieves up to an 11.8 times throughput improvement with 80% KV cache compression with minimal impact on model accuracy.

cs.LG↗
arXiv2026

Where Matters More Than What: Decoding-aligned KV Cache Compression via Position-aware Pseudo Queries

The Key-Value (KV) cache is crucial for efficient Large Language Models (LLMs) inference, but excessively long contexts drastically increase KV cache memory footprint. Existing KV cache compression methods typically rely on input-side attention patterns within a prompt observation window to estimate token importance during the prefill stage. They fail to preserve critical tokens for future generation since these assessments are not derived from the decoding process. Intuitively, an effective observation window should mirror the decoding-stage queries to accurately reflect which tokens the generation process will attend to. However, ground-truth decoding queries are inherently unavailable during inference. For constructing pseudo queries to approximate them, we find that positional information plays a more critical role than semantic content. Motivated by this insight, we propose decoding-aligned KV cache compression via position-aware pseudo queries (DapQ), a novel and lightweight eviction framework that leverages position-aware pseudo queries to simulate the output tokens, thereby establishing an effective observation window for importance assessment. It aligns closely with the actual generation context and enables precise token eviction. Extensive evaluations across multiple benchmarks and LLMs demonstrate that DapQ achieves superior performance, particularly under strict memory constraints (e.g., up to nearly lossless performance 99.5% on NIAH with 3% KV cache budgets).

cs.CL↗
arXiv2025

OpenOneRec Technical Report

While the OneRec series has successfully unified the fragmented recommendation pipeline into an end-to-end generative framework, a significant gap remains between recommendation systems and general intelligence. Constrained by isolated data, they operate as domain specialists-proficient in pattern matching but lacking world knowledge, reasoning capabilities, and instruction following. This limitation is further compounded by the lack of a holistic benchmark to evaluate such integrated capabilities. To address this, our contributions are: 1) RecIF Bench & Open Data: We propose RecIF-Bench, a holistic benchmark covering 8 diverse tasks that thoroughly evaluate capabilities from fundamental prediction to complex reasoning. Concurrently, we release a massive training dataset comprising 96 million interactions from 160,000 users to facilitate reproducible research. 2) Framework & Scaling: To ensure full reproducibility, we open-source our comprehensive training pipeline, encompassing data processing, co-pretraining, and post-training. Leveraging this framework, we demonstrate that recommendation capabilities can scale predictably while mitigating catastrophic forgetting of general knowledge. 3) OneRec-Foundation: We release OneRec Foundation (1.7B and 8B), a family of models establishing new state-of-the-art (SOTA) results across all tasks in RecIF-Bench. Furthermore, when transferred to the Amazon benchmark, our models surpass the strongest baselines with an average 26.8% improvement in Recall@10 across 10 diverse datasets (Figure 1). This work marks a step towards building truly intelligent recommender systems. Nonetheless, realizing this vision presents significant technical and theoretical challenges, highlighting the need for broader research engagement in this promising direction.

cs.IR↗
arXiv2025

$A^3$: Attention-Aware Accurate KV Cache Fusion for Fast Large Language Model Serving

Large language models (LLMs) have demonstrated strong capabilities in processing long contexts, enabling them to tackle tasks involving long textual inputs such as multi-turn conversations, legal documents, or retrieved documents in Retrieval-Augmented Generation (RAG) systems. However, despite their ability to handle long sequences, the resulting decoding latency and memory overhead remain substantial, posing challenges for real-world deployment. Recent advances in KV Cache reuse have shown potential to mitigate these costs, but still suffer from notable performance degradation. To address this issue, we conduct an in-depth investigation of recomputation-based reuse methods and observe that the recomputed tokens often fail to align with the context segments most relevant to the question. This misalignment hinders proper updates to the critical contextual representations. Therefore, we propose the $\textbf{A}$ttention-$\textbf{A}$ware $\textbf{A}$ccurate KV Cache Fusion algorithm ($A^3$), which precomputes and selectively fuses the KV Cache of text chunks based on their relevance to the question, achieving accurate integration with minimal computational overhead. Extensive experiments on various benchmarks and LLMs demonstrate that $A^3$ achieves the best task performance compared to four baselines while reducing the time-to-first-token (TTFT) by 2$\times$.

cs.CL↗
arXiv2025

Highly efficient wideband and polarization-insensitive SMF-ARF coupling strategy with low back-reflection

We propose a lensed-fiber based coupling strategy for low-loss interconnection between single-mode fibers and anti-resonant fibers. By optimizing structural and geometric parameters, the design simultaneously achieves high coupling efficiency and suppressed back-reflection. Experimental results demonstrate an insertion loss of 1.2 dB and back-reflection of -36.22 dB at 1550 nm, with excellent spectral stability (below 0.72 dB variation across 1500-1600 nm) and polarization insensitivity (below 0.4 dB polarization dependent loss). The compact structure not only facilitates the fabrication process, but also enables seamless ARF integration into existing optical networks, thereby addressing critical demands for high-capacity data transmission.

physics.app-ph↗
arXiv2025

OneRec-Think: In-Text Reasoning for Generative Recommendation

The powerful generative capacity of Large Language Models (LLMs) has instigated a paradigm shift in recommendation. However, existing generative models (e.g., OneRec) operate as implicit predictors, critically lacking the capacity for explicit and controllable reasoning-a key advantage of LLMs. To bridge this gap, we propose OneRec-Think, a unified framework that seamlessly integrates dialogue, reasoning, and personalized recommendation. OneRec-Think incorporates: (1) Itemic Alignment: cross-modal Item-Textual Alignment for semantic grounding; (2) Reasoning Activation: Reasoning Scaffolding to activate LLM reasoning within the recommendation context; and (3) Reasoning Enhancement, where we design a recommendation-specific reward function that accounts for the multi-validity nature of user preferences. Experiments across public benchmarks show state-of-the-art performance. Moreover, our proposed "Think-Ahead" architecture enables effective industrial deployment on Kuaishou, achieving a 0.159\% gain in APP Stay Time and validating the practical efficacy of the model's explicit reasoning capability.

cs.IR↗
arXiv2025

TauGenNet: Plasma-Driven Tau PET Image Synthesis via Text-Guided 3D Diffusion Models

Accurate quantification of tau pathology via tau positron emission tomography (PET) scan is crucial for diagnosing and monitoring Alzheimer's disease (AD). However, the high cost and limited availability of tau PET restrict its widespread use. In contrast, structural magnetic resonance imaging (MRI) and plasma-based biomarkers provide non-invasive and widely available complementary information related to brain anatomy and disease progression. In this work, we propose a text-guided 3D diffusion model for 3D tau PET image synthesis, leveraging multimodal conditions from both structural MRI and plasma measurement. Specifically, the textual prompt is from the plasma p-tau217 measurement, which is a key indicator of AD progression, while MRI provides anatomical structure constraints. The proposed framework is trained and evaluated using clinical AV1451 tau PET data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Experimental results demonstrate that our approach can generate realistic, clinically meaningful 3D tau PET across a range of disease stages. The proposed framework can help perform tau PET data augmentation under different settings, provide a non-invasive, cost-effective alternative for visualizing tau pathology, and support the simulation of disease progression under varying plasma biomarker levels and cognitive conditions.

cs.CV↗
arXiv2025

OneRec-V2 Technical Report

Recent breakthroughs in generative AI have transformed recommender systems through end-to-end generation. OneRec reformulates recommendation as an autoregressive generation task, achieving high Model FLOPs Utilization. While OneRec-V1 has shown significant empirical success in real-world deployment, two critical challenges hinder its scalability and performance: (1) inefficient computational allocation where 97.66% of resources are consumed by sequence encoding rather than generation, and (2) limitations in reinforcement learning relying solely on reward models. To address these challenges, we propose OneRec-V2, featuring: (1) Lazy Decoder-Only Architecture: Eliminates encoder bottlenecks, reducing total computation by 94% and training resources by 90%, enabling successful scaling to 8B parameters. (2) Preference Alignment with Real-World User Interactions: Incorporates Duration-Aware Reward Shaping and Adaptive Ratio Clipping to better align with user preferences using real-world feedback. Extensive A/B tests on Kuaishou demonstrate OneRec-V2's effectiveness, improving App Stay Time by 0.467%/0.741% while balancing multi-objective recommendations. This work advances generative recommendation scalability and alignment with real-world feedback, representing a step forward in the development of end-to-end recommender systems.

cs.IR↗
arXiv2025

BrainPath: A Biologically-Informed AI Framework for Individualized Aging Brain Generation

The global population is aging rapidly, and aging is a major risk factor for various diseases. It is an important task to predict how each individual's brain will age, as the brain supports many human functions. This capability can greatly facilitate healthcare automation by enabling personalized, proactive intervention and efficient healthcare resource allocation. However, this task is extremely challenging because of the brain's complex 3D anatomy. While there have been successes in natural image generation and brain MRI synthesis, existing methods fall short in generating individualized, anatomically faithful aging brain trajectories. To address these gaps, we propose BrainPath, a novel AI model that, given a single structural MRI of an individual, generates synthetic longitudinal MRIs that represent that individual's expected brain anatomy as they age. BrainPath introduces three architectural innovations: an age-aware encoder with biologically grounded supervision, a differential age conditioned decoder for anatomically faithful MRI synthesis, and a swap-learning strategy that implicitly separates stable subject-specific anatomy from aging effects. We further design biologically informed loss functions, including an age calibration loss and an age and structural perceptual loss, to complement the conventional reconstruction loss. This enables the model to capture subtle, temporally meaningful anatomical changes associated with aging. We apply BrainPath to two of the largest public aging datasets and conduct a comprehensive, multifaceted evaluation. Our results demonstrate BrainPath's superior performance in generation accuracy, anatomical fidelity, and cross-dataset generalizability, outperforming competing methods.

q-bio.NC↗
arXiv2025

AudioSet-R: A Refined AudioSet with Multi-Stage LLM Label Reannotation

AudioSet is a widely used benchmark in the audio research community and has significantly advanced various audio-related tasks. However, persistent issues with label accuracy and completeness remain critical bottlenecks that limit performance in downstream applications.To address the aforementioned challenges, we propose a three-stage reannotation framework that harnesses general-purpose audio-language foundation models to systematically improve the label quality of AudioSet. The framework employs a cross-modal prompting strategy, inspired by the concept of prompt chaining, wherein prompts are sequentially composed to execute subtasks (audio comprehension, label synthesis, and semantic alignment). Leveraging this framework, we construct a high-quality, structured relabeled version of AudioSet-R. Extensive experiments conducted on representative audio classification models--including AST, PANNs, SSAST, and AudioMAE--consistently demonstrate substantial performance improvements, thereby validating the generalizability and effectiveness of the proposed approach in enhancing label reliability.The code is publicly available at: https://github.com/colaudiolab/AudioSet-R.

cs.SD↗