Selected work

Representative Papers

Zero-shot Generalization in Inventory Management: Train, then Estimate and Decide

Nov 01, 2024arXiv.org

In real-world inventory management, dynamic uncertainty in demand and lead-time distributions severely limits the generalization capability of existing deep reinforcement learning (DRL) methods. To address this, we propose a novel “Train → Estimate → Decide” three-stage framework and introduce GC-LSN—the first zero-shot generalizable inventory agent for lost-sales settings with periodic demand and stochastic lead times. Our approach innovatively integrates a Super-MDP formulation with the Time-Evolving Distribution (TED) framework, combining nonparametric Kaplan–Meier distribution estimation, online parametric identification, and policy self-adaptation. Experiments demonstrate that GC-LSN significantly outperforms classical heuristics under known parameters and surpasses state-of-the-art online learning methods with worst-case guarantees under unknown demand and lead-time distributions. Crucially, it enables real-time, robust decision-making across unseen distributional shifts—achieving true zero-shot generalization in complex, non-stationary inventory environments.

1 citationsRead paper

EASE-TTT: Evidence-Aligned Selective Test-Time Training for Long-Context Question Answering

Jun 05, 2026

This work addresses the challenge that small-scale language models struggle to effectively leverage dispersed evidence in long-context question answering, as existing approaches either neglect query-side attention modulation or fail to accurately localize supporting context. The authors propose a novel framework integrating in-context retrieval with test-time training: key evidence spans are first retrieved and then transformed into soft attention supervision signals to dynamically refine query-side attention parameters, enabling precise alignment with relevant context while preserving the full input for generation. This approach is the first to incorporate soft attention targets for evidence alignment into test-time training, achieving state-of-the-art macro-average performance across six LongBench tasks and three small language models, significantly outperforming baselines such as full-context inference, pure retrieval, and query-side test-time training (qTTT).

0 citationsRead paper

From Replacement to Orchestration: A Socio-Technical Architecture for Agentic AI in Corporate R&D

May 23, 2026

This study addresses the “productivity paradox” in corporate R&D—where escalating investments in research and knowledge accumulation fail to yield commensurate innovation outputs—attributing it primarily to researchers’ cognitive load being consumed by low-value tasks such as coordination, documentation, and data governance. Employing a design science approach, the work proposes HARMONY, a human-AI collaborative R&D operations model featuring a novel four-dimensional socio-technical architecture comprising ResOps, a control tower, ethical fabric, and a talent studio. It introduces the Sciencepreneur role archetype and the Orchestration Leverage performance metric. Through expert interviews, 2040 scenario forecasting, and agent deployment case studies—validated via pattern matching and triangulation—the research emphasizes cognitive load reallocation and the design of bounded autonomy. The resulting framework offers actionable pathways to overcome R&D efficiency bottlenecks and establishes a new evaluation paradigm for hybrid human-AI research productivity.

0 citationsRead paper

Lens-descriptor guided evolutionary algorithm for optimization of complex optical systems with glass choice

Jan 29, 2026

This work addresses the challenge of obtaining diverse, high-performance feasible solutions in high-dimensional, heavily constrained optical system design, where conventional optimization methods often converge to a single local optimum. The authors propose LDG-EA, a two-stage framework: first, the design space is partitioned using a behavior descriptor based on curvature signs and material indices, with a probabilistic model guiding global exploration; second, within each partition, an evolutionary algorithm integrating covariance matrix adaptation and a Hill-Valley mechanism efficiently locates multiple local optima. This approach pioneers the use of behavior descriptors in multimodal optical design, substantially enhancing solution diversity and coverage. Applied to a six-element Double-Gauss configuration, LDG-EA generates an average of 14,500 candidate solutions spanning 636 unique descriptors—nearly an order of magnitude more than CMA-ES—within under one hour, achieving solution quality comparable to reference designs.

0 citationsRead paper

Balancing Fairness and Performance in Multi-User Spark Workloads with Dynamic Scheduling (extended version)

Oct 17, 2025

To address the challenge of simultaneously ensuring fairness and minimizing response time in multi-tenant, long-running Spark clusters, this paper proposes a dynamic scheduling framework. The method introduces three key innovations: (1) User-Weighted Fair Queuing (UWFQ), which enforces fairness at the user level—not the job level—by allocating resources proportionally to user weights; (2) a runtime partitioning mechanism grounded in a virtual fair queue model, integrating job completion time estimation and dynamic task granularity adjustment to mitigate data skew and priority inversion; and (3) fine-grained, adaptive scheduling decisions. Experimental evaluation demonstrates that, compared to Spark’s native scheduler and state-of-the-art fair schedulers, the framework reduces average response time for short jobs by up to 74%, while significantly improving overall system performance under stringent fairness guarantees.

0 citationsRead paper

Recent publications

Latest Papers

EASE-TTT: Evidence-Aligned Selective Test-Time Training for Long-Context Question Answering

Jun 05, 2026

This work addresses the challenge that small-scale language models struggle to effectively leverage dispersed evidence in long-context question answering, as existing approaches either neglect query-side attention modulation or fail to accurately localize supporting context. The authors propose a novel framework integrating in-context retrieval with test-time training: key evidence spans are first retrieved and then transformed into soft attention supervision signals to dynamically refine query-side attention parameters, enabling precise alignment with relevant context while preserving the full input for generation. This approach is the first to incorporate soft attention targets for evidence alignment into test-time training, achieving state-of-the-art macro-average performance across six LongBench tasks and three small language models, significantly outperforming baselines such as full-context inference, pure retrieval, and query-side test-time training (qTTT).

0 citationsRead paper

From Replacement to Orchestration: A Socio-Technical Architecture for Agentic AI in Corporate R&D

May 23, 2026

This study addresses the “productivity paradox” in corporate R&D—where escalating investments in research and knowledge accumulation fail to yield commensurate innovation outputs—attributing it primarily to researchers’ cognitive load being consumed by low-value tasks such as coordination, documentation, and data governance. Employing a design science approach, the work proposes HARMONY, a human-AI collaborative R&D operations model featuring a novel four-dimensional socio-technical architecture comprising ResOps, a control tower, ethical fabric, and a talent studio. It introduces the Sciencepreneur role archetype and the Orchestration Leverage performance metric. Through expert interviews, 2040 scenario forecasting, and agent deployment case studies—validated via pattern matching and triangulation—the research emphasizes cognitive load reallocation and the design of bounded autonomy. The resulting framework offers actionable pathways to overcome R&D efficiency bottlenecks and establishes a new evaluation paradigm for hybrid human-AI research productivity.

0 citationsRead paper

Lens-descriptor guided evolutionary algorithm for optimization of complex optical systems with glass choice

Jan 29, 2026

This work addresses the challenge of obtaining diverse, high-performance feasible solutions in high-dimensional, heavily constrained optical system design, where conventional optimization methods often converge to a single local optimum. The authors propose LDG-EA, a two-stage framework: first, the design space is partitioned using a behavior descriptor based on curvature signs and material indices, with a probabilistic model guiding global exploration; second, within each partition, an evolutionary algorithm integrating covariance matrix adaptation and a Hill-Valley mechanism efficiently locates multiple local optima. This approach pioneers the use of behavior descriptors in multimodal optical design, substantially enhancing solution diversity and coverage. Applied to a six-element Double-Gauss configuration, LDG-EA generates an average of 14,500 candidate solutions spanning 636 unique descriptors—nearly an order of magnitude more than CMA-ES—within under one hour, achieving solution quality comparable to reference designs.

0 citationsRead paper

Balancing Fairness and Performance in Multi-User Spark Workloads with Dynamic Scheduling (extended version)

Oct 17, 2025

To address the challenge of simultaneously ensuring fairness and minimizing response time in multi-tenant, long-running Spark clusters, this paper proposes a dynamic scheduling framework. The method introduces three key innovations: (1) User-Weighted Fair Queuing (UWFQ), which enforces fairness at the user level—not the job level—by allocating resources proportionally to user weights; (2) a runtime partitioning mechanism grounded in a virtual fair queue model, integrating job completion time estimation and dynamic task granularity adjustment to mitigate data skew and priority inversion; and (3) fine-grained, adaptive scheduling decisions. Experimental evaluation demonstrates that, compared to Spark’s native scheduler and state-of-the-art fair schedulers, the framework reduces average response time for short jobs by up to 74%, while significantly improving overall system performance under stringent fairness guarantees.

0 citationsRead paper

Learning to Recognize Correctly Completed Procedure Steps in Egocentric Assembly Videos through Spatio-Temporal Modeling

Oct 14, 2025

Existing program step recognition (PSR) models rely solely on single-frame spatial features, neglecting temporal dynamics and thus exhibiting insufficient robustness and accuracy under partial occlusion. To address this, we propose STORM-PSR, a dual-stream spatiotemporal modeling framework: its spatial branch leverages weakly supervised pretraining to enhance local feature representation, while its temporal branch employs a Transformer architecture to explicitly model inter-frame state evolution and capture fine-grained spatiotemporal changes. The two streams jointly enable precise completion-state discrimination and sequential reasoning of assembly steps. Evaluated on MECCANO and IndustReal, STORM-PSR reduces prediction latency by 11.2% and 26.1%, respectively, and significantly improves robustness under occlusion. This work pioneers the deep integration of strong temporal modeling with weakly supervised spatial learning, establishing a novel paradigm for first-person assembly understanding.

0 citationsRead paper