arXiv subjects

Pinjun Dong

Publications and source records attributed to Pinjun Dong.

4 recordsLinked to original sources

arXiv2026

A hybrid model for zero-inflated proportion data

This paper proposes a hybrid model for analyzing zero-inflated proportion data by distinguishing two types of zeros. The first type arises from sampling variability, whereas the second reflects intrinsically unsuitable covariates. Accordingly, the hybrid model combines a regression component for suitable observations with an estimator of the suitability region that identifies unsuitable covariates. Using a debiasing technique, we establish the consistency of both the regression and classification components, together with the asymptotic normality of the regression coefficient and a convergence rate for the support estimator. Comprehensive simulation studies verify our theoretical findings. We further apply the model to two real datasets, where it outperforms most competing methods.

stat.ME↗

arXiv2026

Do Tabular Foundation Models Still Need Feature Engineering?

Feature engineering has long been a cornerstone of tabular machine learning. Tabular foundation models (TFMs) are pretrained on a wide range of tabular datasets and applied via in-context learning. Their rise raises a natural question: does manual feature construction still matter as these models become more capable? To answer this, we perform a controlled study across several versions of two major TFM families, testing a wide range of existing feature engineering techniques on benchmark datasets from TabArena. We find a consistent pattern: feature engineering gains are concentrated in earlier model generations and become negligible for the strongest models. These results suggest that stronger TFMs depend less on explicitly engineered input representations. In a complementary experiment, however, adding in-context information from related datasets still improves performance. Our findings indicate a shift in the source of performance gains for stronger TFMs: re-representing existing inputs becomes less effective, while providing additional task-relevant context remains beneficial.

cs.LG↗

arXiv2026

Tools Are Not Islands: Set-Level Tool Retrieval for LLM Agents via Query-Conditioned Hyperedge Prediction

Large language model (LLM) agents increasingly rely on invoking external tools to complete real-world tasks. Tool retrieval, which selects a small task-relevant subset from a library of thousands of tools before the agent acts, has therefore become a critical component of LLM agent pipelines. However, existing retrievers either score each tool in isolation or assemble the tool set sequentially, so the joint utility of a candidate set is never evaluated as a whole. In this paper, we propose HYSET, short for HYperedge-based SEt-level Tool retrieval. Our contributions are threefold: (i) we formulate tool retrieval as query-conditioned hyperedge prediction on a tool co-invocation hypergraph, under which the tool set itself becomes the unit of scoring and most existing retrieval paradigms reduce to restricted instances; (ii) we capture size-dependent tool compatibility through cardinality-specific interactions; and (iii) we design HYSET as a pre-selection module requiring no modification to the downstream agent. Experiments on ToolBench demonstrate that HYSET consistently outperforms state-of-the-art baselines in both tool retrieval performance and end-to-end task success. Beyond the in-domain setting, HYSET further supports zero-shot/few-shot transfer, generalizing to held-out tools/categories and unseen domains with minimal supervision.

cs.LG↗

arXiv2025

Statistical ranking with dynamic covariates

We introduce a general covariate-assisted statistical ranking model within the Plackett--Luce framework. Unlike previous studies focusing on individual effects with fixed covariates, our model allows covariates to vary across comparisons. This added flexibility enhances model fitting yet brings significant challenges in analysis. This paper addresses these challenges in the context of maximum likelihood estimation (MLE). We first provide sufficient and necessary conditions for both model identifiability and the unique existence of the MLE. Then, we develop an efficient alternating maximization algorithm to compute the MLE. Under suitable assumptions on the design of comparison graphs and covariates, we establish a uniform consistency result for the MLE, with convergence rates determined by the asymptotic graph connectivity. We also construct random designs where the proposed assumptions hold almost surely. Numerical studies are conducted to support our findings and demonstrate the model's application to real-world datasets, including horse racing and tennis competitions.

stat.ME↗