Selected work

Representative Papers

Fastest Mixing Reversible Markov Chain: Clique Lifted Graphs and Subgraphs

Jan 06, 2020IEEE Transactions on Signal and Information Processing over Networks

This paper addresses the design of fastest-mixing reversible Markov chains: given a target stationary distribution, how to construct a reversible chain with minimal mixing time. Focusing on two fundamental graph structures—clique-lifted graphs and arbitrary subgraphs—the work makes two key contributions: (1) It establishes, for the first time, a strict reduction of the fastest-mixing problem on clique-lifted graphs to that on their base graphs, preserving the optimal mixing time; (2) It proves that the optimal transition probabilities on a subgraph can be determined independently of the ambient graph, ensuring consistency between local optimization and global optimality. Methodologically, the approach integrates semidefinite programming (SDP), graph-theoretic techniques (clique lifting and subgraph embedding), and spectral analysis of reversible chains. Experiments demonstrate that the framework yields analytically optimal solutions across diverse topologies, combining theoretical rigor with computational scalability.

9 citationsRead paper

Fastest mixing reversible Markov chain on friendship graph: Trade-off between transition probabilities among friends and convergence rate

Aug 01, 2019Systems & control letters (Print)

This work addresses the fastest-mixing reversible Markov chain problem on friendship graphs, balancing information propagation efficiency (i.e., convergence rate) against structural stability (i.e., constraints on transition probabilities). Methodologically, it integrates spectral graph theory, convex optimization, and symmetry reduction to derive, for the first time, an analytical solution for the fastest-mixing reversible chain on this graph class; via Lagrangian duality, it obtains a closed-form expression for the optimal transition matrix and rigorously characterizes the tight trade-off between spectral gap and transition probabilities. Theoretical contributions include a proven lower bound on mixing time. Empirically, the proposed chain accelerates convergence by over 40% compared to the uniform random walk on typical friendship graphs. The results provide a tractable, verifiable theoretical framework and constructive design methodology for accelerating consensus in structurally constrained networks.

7 citationsRead paper

HACK++: Towards More Effective Head-Aware Key-Value Compression for Efficient Visual Autoregressive Modeling

Jun 06, 2026

This work addresses the high computational complexity and memory overhead in visual autoregressive models caused by cross-scale accumulation of key-value caches. It reveals, for the first time, that attention heads can be categorized into two distinct types—semantic-consistent and spatial-coherent—and leverages this insight to propose a training-free, head-aware cache compression framework. The method employs offline calibration to identify head types and applies pattern-specific compression strategies coupled with dependency-aware adaptive budget allocation for efficient cache management during inference. Experimental results demonstrate that the approach achieves near-lossless generation using only 30% of the attention budget and 10% of the cache budget, while maintaining robust performance even under an extremely constrained cache budget as low as 1%.

0 citationsRead paper

Covariance-Aware Simplex Projection for Cardinality-Constrained Portfolio Optimization

Dec 22, 2025

In cardinality-constrained portfolio optimization, standard Euclidean projection onto the probability simplex ignores asset covariance structure, leading to insufficient diversification. To address this, we propose a covariance-aware simplex projection method. Our approach is the first to geometrically embed tracking-error risk into the repair operator, constructing a covariance-driven distance metric. It adopts a two-stage design: volatility-normalized screening followed by covariance-weighted projection—ensuring theoretical rigor while maintaining engineering plug-and-play compatibility. Evaluated on S&P 500 data from 2020–2024, the method significantly reduces portfolio variance (p < 0.01) without requiring expected return inputs. Ablation studies confirm synergistic benefits between the two stages. Empirically, it improves the Sharpe ratio and demonstrates strong out-of-sample robustness.

0 citationsRead paper

Rakuten Data Release: A Large-Scale and Long-Term Reviews Corpus for Hotel Domain

Dec 17, 2025

The hotel domain has long lacked large-scale, highly structured, multi-year review corpora. Method: We construct a fine-grained corpus spanning 2009–2024 (16 years), comprising 7.3 million user reviews and corresponding merchant responses—featuring unprecedented temporal breadth, comprehensive metadata (e.g., room type, traveler group, multi-dimensional ratings), and bidirectional user–merchant interaction. We propose a data distribution modeling framework coupled with time-series contrastive analysis to systematically identify key drivers of data drift between 2019–2024 (e.g., pandemic impacts, platform policy shifts). Contribution/Results: All data undergo rigorous anonymization and metadata standardization before open release. The corpus significantly advances NLP model training and evaluation, while uncovering stage-wise evolutionary patterns in user review behavior and platform response strategies.

0 citationsRead paper

Recent publications

Latest Papers

HACK++: Towards More Effective Head-Aware Key-Value Compression for Efficient Visual Autoregressive Modeling

Jun 06, 2026

This work addresses the high computational complexity and memory overhead in visual autoregressive models caused by cross-scale accumulation of key-value caches. It reveals, for the first time, that attention heads can be categorized into two distinct types—semantic-consistent and spatial-coherent—and leverages this insight to propose a training-free, head-aware cache compression framework. The method employs offline calibration to identify head types and applies pattern-specific compression strategies coupled with dependency-aware adaptive budget allocation for efficient cache management during inference. Experimental results demonstrate that the approach achieves near-lossless generation using only 30% of the attention budget and 10% of the cache budget, while maintaining robust performance even under an extremely constrained cache budget as low as 1%.

0 citationsRead paper

Covariance-Aware Simplex Projection for Cardinality-Constrained Portfolio Optimization

Dec 22, 2025

In cardinality-constrained portfolio optimization, standard Euclidean projection onto the probability simplex ignores asset covariance structure, leading to insufficient diversification. To address this, we propose a covariance-aware simplex projection method. Our approach is the first to geometrically embed tracking-error risk into the repair operator, constructing a covariance-driven distance metric. It adopts a two-stage design: volatility-normalized screening followed by covariance-weighted projection—ensuring theoretical rigor while maintaining engineering plug-and-play compatibility. Evaluated on S&P 500 data from 2020–2024, the method significantly reduces portfolio variance (p < 0.01) without requiring expected return inputs. Ablation studies confirm synergistic benefits between the two stages. Empirically, it improves the Sharpe ratio and demonstrates strong out-of-sample robustness.

0 citationsRead paper

Rakuten Data Release: A Large-Scale and Long-Term Reviews Corpus for Hotel Domain

Dec 17, 2025

The hotel domain has long lacked large-scale, highly structured, multi-year review corpora. Method: We construct a fine-grained corpus spanning 2009–2024 (16 years), comprising 7.3 million user reviews and corresponding merchant responses—featuring unprecedented temporal breadth, comprehensive metadata (e.g., room type, traveler group, multi-dimensional ratings), and bidirectional user–merchant interaction. We propose a data distribution modeling framework coupled with time-series contrastive analysis to systematically identify key drivers of data drift between 2019–2024 (e.g., pandemic impacts, platform policy shifts). Contribution/Results: All data undergo rigorous anonymization and metadata standardization before open release. The corpus significantly advances NLP model training and evaluation, while uncovering stage-wise evolutionary patterns in user review behavior and platform response strategies.

0 citationsRead paper

Autoregressive Image Generation Needs Only a Few Lines of Cached Tokens

Dec 04, 2025

Autoregressive image generation suffers from memory explosion and throughput bottlenecks due to linear growth of KV caches with decoding length. This paper proposes LineAR—a training-free, progressive KV cache compression method based on a 2D row-wise structure. Leveraging spatial locality in visual attention and inter-row dependencies, LineAR dynamically identifies and prunes low-information tokens, retaining only critical row-level caches. Its core innovation is a cross-row attention-guided, training-free compression strategy, fully compatible with diverse autoregressive image generation architectures. Evaluated on multiple state-of-the-art models, LineAR reduces GPU memory consumption by up to 67.61% and accelerates inference by up to 7.57×, while simultaneously achieving new SOTA performance on ImageNet and COCO generation benchmarks. Remarkably, it attains superior generation quality using only 1/6–1/8 of the original cache footprint.

0 citationsRead paper

Test-Time Temporal Sampling for Efficient MLLM Video Understanding

Nov 22, 2025

To address the quadratic computational complexity and inefficiency of self-attention in multimodal large language models (MLLMs) when processing long videos, this paper proposes T3S—a training-free, plug-and-play test-time inference framework. T3S leverages spatiotemporal redundancy in videos as a computational advantage: it dynamically subsamples the input video into multiple short, complementary clips, which are then encoded and fused in parallel within a single forward pass. The method requires no architectural modification or model fine-tuning. Consequently, computational complexity is reduced from O(L²) to O(∑αᵢ²L²), where αᵢ denotes clip length ratios. Evaluated on multiple long-video understanding benchmarks, T3S achieves up to 3.1% average accuracy gain, reduces first-token latency by 2.04×, and maintains full compatibility with diverse pre-trained MLLMs at negligible integration cost.

0 citationsRead paper