Selected work

Representative Papers

PaSTel: Anchoring Histology in Spatial Transcriptomics via Multi-Scale Hierarchical Bio-Prior Contrastive Pretraining

Aug 14, 2026

This study addresses the limitations of weak gene discriminability and neglected spatial dependencies in spatial transcriptomics pre-training by proposing PaSTel, a hierarchical multimodal framework. This method introduces a novel three-scale biologically informed contrastive learning mechanism spanning point, functional, and regional levels. By integrating TF-IDF reweighting, KEGG pathway anchoring, and spatial clustering, PaSTel achieves deep alignment between histology and gene expression. Experimental results demonstrate that PaSTel consistently outperforms existing vision and omics encoders across multiple downstream tasks. Crucially, this approach effectively bridges the gap between global semantics and spatial structure, significantly enhancing both representation informativeness and transferability for spatial transcriptomic analysis.

0 citationsRead paper

EFSPI CMCSNE SIG position on the 'Expected f2'

Jul 18, 2026

This study systematically evaluates and challenges the scientific validity of the “expected f2” approach for comparing dissolution profiles, particularly its suitability as a replacement for the conventional f2 metric under conditions of high variability. Through comprehensive literature review, statistical analysis, and expert consultation, the EFSPI CMCSNE SIG working group reveals fundamental flaws in the method, including the absence of original theoretical justification, mathematical bias, low statistical power, and ambiguous definition. The findings strongly advise against incorporating “expected f2” into regulatory guidance, thereby providing critical scientific evidence to inform policy decisions and addressing a significant gap in the systematic critique of this methodology.

0 citationsRead paper

EVOCHAMBER: Test-Time Co-evolution of Multi-Agent System at Individual, Team, and Population Scales

May 11, 2026

Existing test-time multi-agent evolution methods struggle to balance cross-agent learning with collaborative specialization. This work proposes a training-free, multi-scale co-evolution framework that dynamically constructs specialized structures through failure-driven collaborative reflection and asymmetric knowledge transfer across individual, team, and population levels. It is the first approach to enable test-time multi-scale co-evolution, allowing specialized agents to emerge spontaneously while preserving collaborative diversity. Built upon the CODREAM protocol, online team assembly, and population lifecycle operations—including forking, merging, pruning, and seeding—the framework implements heterogeneous task pipelines on Qwen3-8B, achieving accuracies of 63.9%, 75.7%, and 87.1% on competition mathematics, code generation, and multi-domain reasoning tasks, respectively. This represents a 32% relative improvement in mathematical performance and consistently yields 4–5 specialized agents.

0 citationsRead paper

Ground-Level Near Real-Time Modeling for PM2.5 Pollution Prediction

Apr 20, 2026

Existing PM2.5 prediction models struggle to simultaneously achieve near real-time performance and high spatiotemporal resolution, limiting their utility in public health decision-making. This study proposes a lightweight, grid-free deep learning architecture that integrates readily available multisource data—including topography, meteorology, and land use—through a spatially randomized sampling training strategy to enable high-accuracy PM2.5 interpolation between sparse monitoring stations. The method supports rapid querying at arbitrary locations, near real-time updates, and flexible deployment across multiple spatial scales, significantly enhancing model generalizability in both dense and sparse monitoring regions. By providing an efficient and practical exposure assessment tool, this approach advances support for epidemiological research and public health emergency response.

0 citationsRead paper

Infeasibility Aware Large Language Models for Combinatorial Optimization

Apr 01, 2026

This work addresses the challenge that current large language models (LLMs) struggle to explicitly identify infeasible instances in combinatorial optimization. The authors propose a unified framework integrating feasible solution generation with infeasibility detection: first, provably correct infeasibility labels are derived from exact mathematical programming to construct high-quality supervised data; then, an 8B-parameter LLM is fine-tuned using this data, and its outputs serve as warm starts for local search. Experimental results demonstrate that the proposed approach achieves up to a 30% higher accuracy compared to GPT-5.2 and accelerates downstream search by a factor of two through LLM-guided warm starts, substantially enhancing both solution efficiency and reliability.

0 citationsRead paper

Recent publications

Latest Papers

PaSTel: Anchoring Histology in Spatial Transcriptomics via Multi-Scale Hierarchical Bio-Prior Contrastive Pretraining

Aug 14, 2026

This study addresses the limitations of weak gene discriminability and neglected spatial dependencies in spatial transcriptomics pre-training by proposing PaSTel, a hierarchical multimodal framework. This method introduces a novel three-scale biologically informed contrastive learning mechanism spanning point, functional, and regional levels. By integrating TF-IDF reweighting, KEGG pathway anchoring, and spatial clustering, PaSTel achieves deep alignment between histology and gene expression. Experimental results demonstrate that PaSTel consistently outperforms existing vision and omics encoders across multiple downstream tasks. Crucially, this approach effectively bridges the gap between global semantics and spatial structure, significantly enhancing both representation informativeness and transferability for spatial transcriptomic analysis.

0 citationsRead paper

EFSPI CMCSNE SIG position on the 'Expected f2'

Jul 18, 2026

This study systematically evaluates and challenges the scientific validity of the “expected f2” approach for comparing dissolution profiles, particularly its suitability as a replacement for the conventional f2 metric under conditions of high variability. Through comprehensive literature review, statistical analysis, and expert consultation, the EFSPI CMCSNE SIG working group reveals fundamental flaws in the method, including the absence of original theoretical justification, mathematical bias, low statistical power, and ambiguous definition. The findings strongly advise against incorporating “expected f2” into regulatory guidance, thereby providing critical scientific evidence to inform policy decisions and addressing a significant gap in the systematic critique of this methodology.

0 citationsRead paper

EVOCHAMBER: Test-Time Co-evolution of Multi-Agent System at Individual, Team, and Population Scales

May 11, 2026

Existing test-time multi-agent evolution methods struggle to balance cross-agent learning with collaborative specialization. This work proposes a training-free, multi-scale co-evolution framework that dynamically constructs specialized structures through failure-driven collaborative reflection and asymmetric knowledge transfer across individual, team, and population levels. It is the first approach to enable test-time multi-scale co-evolution, allowing specialized agents to emerge spontaneously while preserving collaborative diversity. Built upon the CODREAM protocol, online team assembly, and population lifecycle operations—including forking, merging, pruning, and seeding—the framework implements heterogeneous task pipelines on Qwen3-8B, achieving accuracies of 63.9%, 75.7%, and 87.1% on competition mathematics, code generation, and multi-domain reasoning tasks, respectively. This represents a 32% relative improvement in mathematical performance and consistently yields 4–5 specialized agents.

0 citationsRead paper

Ground-Level Near Real-Time Modeling for PM2.5 Pollution Prediction

Apr 20, 2026

Existing PM2.5 prediction models struggle to simultaneously achieve near real-time performance and high spatiotemporal resolution, limiting their utility in public health decision-making. This study proposes a lightweight, grid-free deep learning architecture that integrates readily available multisource data—including topography, meteorology, and land use—through a spatially randomized sampling training strategy to enable high-accuracy PM2.5 interpolation between sparse monitoring stations. The method supports rapid querying at arbitrary locations, near real-time updates, and flexible deployment across multiple spatial scales, significantly enhancing model generalizability in both dense and sparse monitoring regions. By providing an efficient and practical exposure assessment tool, this approach advances support for epidemiological research and public health emergency response.

0 citationsRead paper

Infeasibility Aware Large Language Models for Combinatorial Optimization

Apr 01, 2026

This work addresses the challenge that current large language models (LLMs) struggle to explicitly identify infeasible instances in combinatorial optimization. The authors propose a unified framework integrating feasible solution generation with infeasibility detection: first, provably correct infeasibility labels are derived from exact mathematical programming to construct high-quality supervised data; then, an 8B-parameter LLM is fine-tuned using this data, and its outputs serve as warm starts for local search. Experimental results demonstrate that the proposed approach achieves up to a 30% higher accuracy compared to GPT-5.2 and accelerates downstream search by a factor of two through LLM-guided warm starts, substantially enhancing both solution efficiency and reliability.

0 citationsRead paper