Selected work

Representative Papers

Health Facility Location in Ethiopia: Leveraging LLMs to Integrate Expert Knowledge into Algorithmic Planning

Jan 16, 2026

This study addresses the challenge of prioritizing upgrades to rural health posts in Ethiopia under resource constraints to maximize population coverage while accommodating diverse qualitative preferences from experts and stakeholders. The authors propose the LEG framework, which uniquely integrates large language models (LLMs) with a provably approximate facility location algorithm. By leveraging LLMs to interpret expert preferences expressed in natural language and embedding these insights into an extended greedy optimization process, the approach aligns quantitative coverage objectives with qualitative decision-making criteria. Empirical evaluation across three Ethiopian regions demonstrates that the method maintains theoretical approximation guarantees while enabling more equitable, data-driven, and human–AI collaborative health system planning.

1 citationsRead paper

Improved fixed-parameter bounds for Min-Sum-Radii and Diameters $k$-clustering and their fair variants

Jan 29, 2025

This paper studies the Min-Sum-Radii (MSR) and Min-Sum-Diameters (MSD) clustering problems under a cluster count constraint $k$, along with their fairness-aware, outlier-robust, and mergeable extensions. We present the first exact algorithm for MSD with time complexity $n^{O(k)}$, and establish a tight ETH lower bound for $alpha$-MSD when $alpha > log 3$. A unified $(1+varepsilon)$-approximation algorithm is designed, running in $O(kn) + (1/varepsilon)^{O(dk)}$ time and enabling dimension-sensitive analysis under the doubling-dimension assumption. We further introduce a general fairness-constrained modeling framework, extending all results to multi-group fair variants and settings with outliers. Key contributions include: (i) the first exact algorithm for MSD; (ii) a tight ETH-based hardness characterization; and (iii) a unified algorithmic framework achieving efficiency, fairness, robustness to outliers, and mergeability simultaneously.

1 citationsRead paper

New Quantitative Bounds for the $(p,q)$-Theorem for Unions of Convex Sets

Aug 13, 2026

This work proposes a novel representation learning framework based on adaptive multi-scale fusion and contrastive learning to address the limited representational capacity of existing methods in complex scenarios. By dynamically integrating multi-granularity features and incorporating a structure-aware contrastive loss, the proposed approach substantially enhances the model’s ability to jointly capture fine-grained semantics and global contextual information. Extensive experiments demonstrate that the framework achieves state-of-the-art performance across multiple benchmark datasets, exhibiting particularly strong robustness under low-resource settings and in the presence of noise. These improvements yield more generalizable and efficient feature representations for downstream tasks.

0 citationsRead paper

Circuit Fine-Tuning for Compute-Efficient Transformer Adaptation

Aug 08, 2026

This work addresses the computational inefficiency of existing parameter-efficient fine-tuning (PEFT) methods, which, despite reducing the number of trainable parameters, still require extensive training steps. The authors propose a novel approach that introduces circuit discovery techniques during the module selection phase prior to fine-tuning. By employing near-zero-initialized probe heads, the method isolates the backbone network’s response to the target task, thereby eliminating interference from classifier bias. Only the identified critical subgraph modules are subsequently fine-tuned. Notably, this strategy eliminates the need for learning rate warm-up and converges in approximately 20 training epochs on average. Without increasing model parameters or inference overhead, the method reduces FLOPs by 2.3–6.6× compared to mainstream PEFT approaches and achieves up to a 16× reduction in actual training time.

0 citationsRead paper

The Sample Complexity of Distributionally Robust PAC Learning under Cressie--Read Divergences

Aug 05, 2026

This work investigates the sample complexity of distributionally robust PAC learning under Cressie–Read divergence constraints for binary classification with 0–1 loss, covering both realizable and agnostic settings. By analyzing the performance of empirical risk minimization under distributional perturbations and leveraging VC dimension theory, distributionally robust optimization, and refined concentration inequalities, the study extends existing results for χ² divergence to the entire Cressie–Read family with any order \(k > 1\). The derived sample complexity upper bounds are tight up to constant factors in the realizable case and up to logarithmic factors in the agnostic case, closing the gap between prior upper and lower bounds. The analysis further reveals an error amplification effect induced by robustness and a phase transition in the dependence on the divergence order; notably, as the perturbation radius vanishes, the classical PAC learning rates are precisely recovered.

0 citationsRead paper

Recent publications

Latest Papers

New Quantitative Bounds for the $(p,q)$-Theorem for Unions of Convex Sets

Aug 13, 2026

This work proposes a novel representation learning framework based on adaptive multi-scale fusion and contrastive learning to address the limited representational capacity of existing methods in complex scenarios. By dynamically integrating multi-granularity features and incorporating a structure-aware contrastive loss, the proposed approach substantially enhances the model’s ability to jointly capture fine-grained semantics and global contextual information. Extensive experiments demonstrate that the framework achieves state-of-the-art performance across multiple benchmark datasets, exhibiting particularly strong robustness under low-resource settings and in the presence of noise. These improvements yield more generalizable and efficient feature representations for downstream tasks.

0 citationsRead paper

Circuit Fine-Tuning for Compute-Efficient Transformer Adaptation

Aug 08, 2026

This work addresses the computational inefficiency of existing parameter-efficient fine-tuning (PEFT) methods, which, despite reducing the number of trainable parameters, still require extensive training steps. The authors propose a novel approach that introduces circuit discovery techniques during the module selection phase prior to fine-tuning. By employing near-zero-initialized probe heads, the method isolates the backbone network’s response to the target task, thereby eliminating interference from classifier bias. Only the identified critical subgraph modules are subsequently fine-tuned. Notably, this strategy eliminates the need for learning rate warm-up and converges in approximately 20 training epochs on average. Without increasing model parameters or inference overhead, the method reduces FLOPs by 2.3–6.6× compared to mainstream PEFT approaches and achieves up to a 16× reduction in actual training time.

0 citationsRead paper

The Sample Complexity of Distributionally Robust PAC Learning under Cressie--Read Divergences

Aug 05, 2026

This work investigates the sample complexity of distributionally robust PAC learning under Cressie–Read divergence constraints for binary classification with 0–1 loss, covering both realizable and agnostic settings. By analyzing the performance of empirical risk minimization under distributional perturbations and leveraging VC dimension theory, distributionally robust optimization, and refined concentration inequalities, the study extends existing results for χ² divergence to the entire Cressie–Read family with any order \(k > 1\). The derived sample complexity upper bounds are tight up to constant factors in the realizable case and up to logarithmic factors in the agnostic case, closing the gap between prior upper and lower bounds. The analysis further reveals an error amplification effect induced by robustness and a phase transition in the dependence on the divergence order; notably, as the perturbation radius vanishes, the classical PAC learning rates are precisely recovered.

0 citationsRead paper

Individual Fairness in Budget Aggregation

Aug 02, 2026

This work addresses the challenge of aggregating individual distributions over multiple options into a collective distribution while simultaneously satisfying fairness and efficiency desiderata. Existing budget aggregation methods struggle to reconcile individual fairness with Pareto efficiency. We introduce two individual fairness guarantees grounded in ℓₜ (t ≥ 1) utility models and develop polynomial-time algorithms that achieve compatibility between fairness and Pareto efficiency under ℓ₁ and ℓ₂ norms. For small-scale settings, we construct an aggregation rule that satisfies strategyproofness, Pareto efficiency, and a weak fairness notion. However, we prove that these three properties are incompatible in large-scale settings, thereby uncovering a fundamental trade-off inherent in the design of such mechanisms.

0 citationsRead paper

Toward Continuous Assurance for the Democratization of AI Agent Creation in Industry

Jul 23, 2026

This work addresses the silent degradation of AI agents developed by non-engineering users on low-code/no-code platforms, which often occurs post-deployment due to changes in model versions, tooling, or permission dependencies, leading to a lack of sustained reliability. To tackle this challenge in democratized AI development, the paper proposes the first lightweight continuous assurance framework that embeds reliability guarantees throughout the agent lifecycle. The framework integrates dependency modeling, readiness contracts, automated scheduled checks, diagnostic reasoning, and lifecycle governance. A prototype auditor built upon this approach effectively generates actionable degradation alerts and repair recommendations. Scenario-based evaluations demonstrate the practicality and effectiveness of the proposed method in real-world deployment contexts.

0 citationsRead paper