Selected work

Representative Papers

Application-Driven Innovation in Machine Learning

Mar 26, 2024International Conference on Machine Learning

Application-driven machine learning (ML) research has been systematically undervalued in academia, leading to a growing disconnect between algorithmic innovation and real-world needs; this marginalization is reinforced by structural biases in peer review, faculty hiring, and pedagogy. Method: This paper introduces, for the first time, a formally defined “application-driven ML research paradigm,” elucidating its complementary relationship with the dominant methodology-driven paradigm. Drawing on interdisciplinary frameworks from education theory, research governance, and ML practice—and substantiated by empirical case studies and institutional critique—it diagnoses three systemic barriers hindering such research. Contribution/Results: The core contribution is a set of actionable, process-level interventions to reform academic evaluation systems, grounded in both theoretical analysis and pragmatic implementation pathways. These proposals have already catalyzed curricular reforms in AI education and adjustments to national funding review criteria across multiple universities, fostering cross-domain collaboration and methodological feedback loops between application domains and core ML research.

22 citations2 influentialRead paper

Mixture of Experts Softens the Curse of Dimensionality in Operator Learning

Apr 13, 2024

To address the computational and memory bottlenecks imposed by the curse of dimensionality in high-dimensional operator learning, this paper proposes the Mixture-of-Experts Neural Operator (MoNO): a framework that decomposes a global nonlinear operator into multiple lightweight expert sub-operators, routed input-adaptively via a learnable decision-tree mechanism. Theoretically, we establish the first distributed universal approximation theorem, proving that MoNO uniformly approximates any Lipschitz-continuous nonlinear operator in Sobolev spaces; each expert’s depth, width, and rank scale as O(ε⁻¹), ensuring controllable memory footprint compatible with standard hardware. We further derive the first quantitative approximation rate for classical neural operators. Experiments and theory jointly demonstrate that MoNO achieves ε-accuracy with significantly reduced complexity, overcoming both expressive and deployability limitations inherent to monolithic neural operators.

20 citationsRead paper

In-and-Out: Algorithmic Diffusion for Sampling Convex Bodies

May 02, 2024Neural Information Processing Systems

This work addresses efficient uniform sampling from high-dimensional convex bodies, providing strong convergence guarantees under Rényi divergence—including total variation (TV), Wasserstein-2 ($mathcal{W}_2$), Kullback–Leibler (KL), and $chi^2$ divergences. We propose a novel stochastic walk algorithm that models sampling through the lens of stochastic diffusion—a first in this context—and characterizes convergence rates via the functional isoperimetric constant of the target distribution, thereby departing from conventional polynomial mixing-time analysis. Theoretically, our algorithm achieves the optimal time complexity $O^*(n^2 R^2)$, where $n$ denotes dimensionality and $R$ the body’s diameter. Moreover, we derive unified, tight convergence bounds across the entire Rényi divergence family. To our knowledge, this is the first uniform sampling scheme attaining simultaneous optimality under multiple probability metrics.

12 citations2 influentialRead paper

Machine Unlearning Fails to Remove Data Poisoning Attacks

Jun 25, 2024arXiv.org

This work exposes a systemic failure of existing approximate machine unlearning methods against diverse data poisoning attacks—including indiscriminate, targeted, and a newly proposed Gaussian poisoning attack—demonstrating that these methods fail to meaningfully mitigate poisoning effects in both image classifiers and large language models, performing no better than full retraining and thus posing a false sense of security. To address this, the authors introduce a unified evaluation framework covering multiple attack types and model architectures, propose a novel poisoning-aware unlearning effectiveness metric, and empirically benchmark mainstream approaches (e.g., gradient updates, influence function approximation, subset retraining). Results reveal that current methods lack theoretical guarantees and exhibit unreliable real-world behavior. The paper advocates for a more rigorous, scenario-driven unlearning evaluation paradigm, establishing foundational benchmarks and research directions for trustworthy machine unlearning.

6 citations1 influentialRead paper

When Single-Agent with Skills Replace Multi-Agent Systems and When They Fail

Jan 08, 2026arXiv.org

This study investigates whether a single agent can internalize multi-agent collaborative behaviors as a skill selection mechanism to replace conventional multi-agent systems and reduce communication overhead. The authors propose a large language model (LLM)-based skill selection framework augmented with a hierarchical routing mechanism and, drawing inspiration from cognitive science, analyze scalability bottlenecks under expanding skill libraries. They reveal for the first time that LLMs exhibit human-like cognitive capacity limits in skill selection: as the skill library grows, accuracy undergoes a phase-transition-like decline primarily due to semantic interference. Experimental results demonstrate that the proposed approach significantly reduces token consumption and latency while preserving reasoning accuracy, and that hierarchical organization effectively mitigates the performance degradation associated with large-scale skill sets.

4 citationsRead paper

Recent publications

Latest Papers

RoboSynChallenge: Mastering Real-World Dexterity via Generalizing Synthesized Manipulation Skills

Aug 12, 2026

This work addresses the limited generalization of real-world robotic manipulation policies, which stems from the scarcity and insufficient diversity of real data. To systematically advance research on generalizable manipulation across tasks and environments, the paper introduces the first unified benchmark that integrates large-scale synthetic data training with standardized real-robot evaluation. Methodologically, it pioneers the combination of synthetic skill learning with rigorous real-world validation, leveraging state-of-the-art architectures—including Transformers, diffusion models, and vision-language-action frameworks—to construct manipulation policies. The project provides multiple baseline implementations and establishes a reproducible, comparable evaluation framework for generalization, thereby facilitating the development of data-efficient and adaptable general-purpose robotic manipulation systems.

0 citationsRead paper

DYSANOS Generative Dynamic Smooth Arbitrage-free Non-parametric Option Surfaces

Aug 12, 2026

This work proposes SANOS, the first generative market model capable of producing smooth, static-arbitrage-free option surfaces spanning the full range of strikes and maturities, while supporting multi-year simulation of joint dynamics for both underlying assets and option prices. Built upon an AR(1) latent-state framework, the method integrates a nonparametric surface representation with explicit static and dynamic arbitrage constraints, accompanied by a complete data pipeline and training protocol. Empirical evaluation on S&P 500 index options from 2020 to 2025 demonstrates that SANOS substantially outperforms conventional implied volatility PCA benchmarks, achieving full-dimensional surface consistency, numerical stability, and strict adherence to no-arbitrage conditions across both time and state space.

0 citationsRead paper

FiGuRO: Intrinsic Dimension Estimation for Multi-Modal Data

Aug 11, 2026

This work addresses the challenge of accurately estimating the intrinsic dimensions of shared and private information in multimodal data, a task inadequately handled by existing methods that are often static or only implicitly model shared structures. To overcome this limitation, we propose FiGuRO, a fidelity-guided rank optimization framework that dynamically estimates the intrinsic dimensions of both unimodal and multimodal representations. FiGuRO adaptively adjusts low-rank projection dimensions under constraints of model capacity and hyperparameters, enabling the disentanglement of shared and private components to emerge naturally from the optimization process—without requiring complex auxiliary losses. The framework also supports efficient post-hoc disentanglement of pre-trained unimodal models. Experiments on both synthetic and real-world datasets demonstrate that FiGuRO consistently outperforms current approaches, robustly capturing multiscale intrinsic dimensions and subspace proportions while effectively separating shared and private information.

0 citationsRead paper

Solver-Guided Reasoning for Mixed-Equilibrium Strategies

Aug 06, 2026

This work addresses the suboptimal performance of large language models (LLMs) in game-theoretic equilibrium reasoning, which stems from their reliance on human behavioral data that often deviates from mixed-strategy equilibria. To overcome this limitation, the authors propose replacing human data with equilibrium strategies generated by game solvers to construct an interpretable Mixed-strategy Decision Tree (MDT). This framework translates implicit optimal strategies into sparse, human- and model-interpretable rules. The MDT is trained using shadow-based contrastive learning and extensive state querying—encompassing 250 million decisions. Evaluated across eight LLMs, the approach reduces the ℓ₁ distance between model policies and theoretical equilibria by 52.6% on average, demonstrating high policy fidelity and strong generalization across states in both river-end poker scenarios and Liar’s Dice tasks.

0 citationsRead paper

Empirical Analysis of Evasion and Poisoning Against Malware Data Drift Detection

Aug 04, 2026

This study addresses the challenge posed by concept drift induced by malware evolution to classification systems, particularly under adversarial conditions where evasion and poisoning attacks target the classifier. While data drift detectors are commonly deployed to maintain model reliability, their resilience against such classifier-focused attacks remains unclear. This work presents the first systematic evaluation of how these adversarial attacks impact the joint system comprising a drift detector and a malware classifier. Through empirical analysis integrating representative attack strategies with machine learning–based drift detection techniques, the study reveals significant disparities in attack effectiveness across system components—demonstrating that attacks optimized for classifiers do not necessarily compromise drift detectors. Furthermore, it elucidates how intrinsic properties of drift detectors modulate adversarial success, offering critical insights and design principles for building robust malware detection systems.

0 citationsRead paper