arXiv subjects

Chentao Li

Publications and source records attributed to Chentao Li.

9 recordsLinked to original sources

arXiv2026

Hasse Diagrams for Attention: A Partial Order Framework for Designing Transformer Masks

During the training of large Transformer models, attention masks regulate the scope and direction of information flow across a sequence. Numerous mask variants exist, and operators such as FlexAttention already support arbitrary attention masks. Nevertheless, a systematic formal analysis of the information-flow structure induced by arbitrary masks has been missing. This paper develops a complete theoretical framework. We prove that, with sufficient depth, the information flow of a multi-layer Transformer converges to a Hasse diagram -- a directed acyclic graph representing a partial order. Building on this, we recast the design of parallel training tasks as the problem of finding a minimal common supergraph of Hasse diagrams, and we establish a criterion for the minimal common supergraph. This yields a constructive method to derive attention masks directly from a family of tasks. Applying the framework, we design two novel masks: a block-generation attention mask that ensures training-inference consistency (Block Two-Stream Attention), and a fully supervised bidirectional attention mask (Butterfly Attention). These results demonstrate the framework's capacity to discover new structures.

cs.LG↗

arXiv2026

Do MLLMs Understand Pointing? Benchmarking and Enhancing Referential Reasoning in Egocentric Vision

Egocentric AI agents, such as smart glasses, rely on pointing gestures to resolve referential ambiguities in natural language commands. However, despite advancements in Multimodal Large Language Models (MLLMs), current systems often fail to precisely ground the spatial semantics of pointing. Instead, they rely on spurious correlations with visual proximity or object saliency, a phenomenon we term "Referential Hallucination." To address this gap, we introduce EgoPoint-Bench, a comprehensive question-answering benchmark designed to evaluate and enhance multimodal pointing reasoning in egocentric views. Comprising over 11k high-fidelity simulated and real-world samples, the benchmark spans five evaluation dimensions and three levels of referential complexity. Extensive experiments demonstrate that while state-of-the-art proprietary and open-source models struggle with egocentric pointing, models fine-tuned on our synthetic data achieve significant performance gains and robust sim-to-real generalization. This work highlights the importance of spatially aware supervision and offers a scalable path toward precise egocentric AI assistants. Project page: https://guyyyug.github.io/EgoPoint-Bench/

cs.CV↗

arXiv2026

Prototype Instance-semantic Disentanglement with Low-rank Regularized Subspace Clustering for WSIs Explainable Recognition

The tumor region plays a key role in pathological diagnosis. Tumor tissues are highly similar to precancerous lesions and non tumor instances often greatly exceed tumor instances in whole slide images (WSIs). These issues cause instance-semantic entanglement in multi-instance learning frameworks, degrading both model representation capability and interpretability. To address this, we propose an end-to-end prototype instance semantic disentanglement framework with low-rank regularized subspace clustering, PID-LRSC, in two aspects. First, we use secondary instance subspace learning to construct low-rank regularized subspace clustering (LRSC), addressing instance entanglement caused by an excessive proportion of non tumor instances. Second, we employ enhanced contrastive learning to design prototype instance semantic disentanglement (PID), resolving semantic entanglement caused by the high similarity between tumor and precancerous tissues. We conduct extensive experiments on multicentre pathology datasets, implying that PID-LRSC outperforms other SOTA methods. Overall, PID-LRSC provides clearer instance semantics during decision-making and significantly enhances the reliability of auxiliary diagnostic outcomes.

cs.CV↗

arXiv2026

MacNet: An End-to-End Manifold-Constrained Adaptive Clustering Network for Interpretable Whole Slide Image Classification

Whole slide images (WSIs) are the gold standard for pathological diagnosis and sub-typing. Current main-stream two-step frameworks employ offline feature encoders trained without domain-specific knowledge. Among them, attention-based multiple instance learning (MIL) methods are outcome-oriented and offer limited interpretability. Clustering-based approaches can provide explainable decision-making process but suffer from high dimension features and semantically ambiguous centroids. To this end, we propose an end-to-end MIL framework that integrates Grassmann re-embedding and manifold adaptive clustering, where the manifold geometric structure facilitates robust clustering results. Furthermore, we design a prior knowledge guiding proxy instance labeling and aggregation strategy to approximate patch labels and focus on pathologically relevant tumor regions. Experiments on multicentre WSI datasets demonstrate that: 1) our cluster-incorporated model achieves superior performance in both grading accuracy and interpretability; 2) end-to-end learning refines better feature representations and it requires acceptable computation resources.

cs.CV↗

arXiv2025

Positive Semi-definite Latent Factor Grouping-Boosted Cluster-reasoning Instance Disentangled Learning for WSI Representation

Multiple instance learning (MIL) has been widely used for representing whole-slide pathology images. However, spatial, semantic, and decision entanglements among instances limit its representation and interpretability. To address these challenges, we propose a latent factor grouping-boosted cluster-reasoning instance disentangled learning framework for whole-slide image (WSI) interpretable representation in three phases. First, we introduce a novel positive semi-definite latent factor grouping that maps instances into a latent subspace, effectively mitigating spatial entanglement in MIL. To alleviate semantic entanglement, we employs instance probability counterfactual inference and optimization via cluster-reasoning instance disentangling. Finally, we employ a generalized linear weighted decision via instance effect re-weighting to address decision entanglement. Extensive experiments on multicentre datasets demonstrate that our model outperforms all state-of-the-art models. Moreover, it attains pathologist-aligned interpretability through disentangled representations and a transparent decision-making process.

cs.CV↗

arXiv2023

Charge and Energy Transfer Dynamics of Hybridized Exciton-Polaritons in 2D Halide Perovskites

Excitons, bound electron-hole pairs, in Two-Dimensional Hybrid Organic Inorganic Perovskites (2D HOIPs) are capable of forming hybrid light-matter states known as exciton-polaritons (E-Ps) when the excitonic medium is confined in an optical cavity. In the case of 2D HOIPs, they can self-hybridize into E-Ps at specific thicknesses of the HOIP crystals that form a resonant optical cavity with the excitons. However, the fundamental properties of these self-hybridized E-Ps in 2D HOIPs, including their role in ultrafast energy and/or charge transfer at interfaces, remain unclear. Here, we demonstrate that > 0.5 um thick 2D HOIP crystals on Au substrates are capable of supporting multiple-orders of self-hybridized E-P modes. These E-Ps have high Q factors (> 100) and modulate the optical dispersion for the crystal to enhance sub-gap absorption and emission. Through varying excitation energy and ultrafast measurements, we also confirm energy transfer from higher energy upper E-Ps to lower energy, lower E-Ps. Finally, we also demonstrate that E-Ps are capable of charge transport and transfer at interfaces. Our findings provide new insights into charge and energy transfer in E-Ps opening new opportunities towards their manipulation for polaritonic devices.

cond-mat.mes-hall↗

arXiv2022

Invertible Optical Nonlinearity in Epsilon-near-zero Materials

Epsilon-near-zero (ENZ) materials such as indium tin oxide (ITO), have recently emerged as a new platform to enhance optical nonlinearities. Here we report a theoretical and experimental study on the origin of nonlinearities in ITO thin films that are dominated by two mechanisms based on intraband and interband transitions. We show that there are two competing factors that jointly contribute to a spectrally-invertible nonlinearity of ITO near its ENZ region i.e. the non-parabolicity of the band structure that results in a larger effective mass in the intraband transition and the Fermi energy shift, which determines the free carrier density. Our work reveals the relationship between the large nonlinearity and the intrinsic material properties of the ITO films.

cond-mat.mtrl-sci↗

arXiv2020

Second Harmonic Generation from a Single Plasmonic Nanorod Strongly Coupled to a WSe2 Monolayer

Monolayer transition metal dichalcogenides, coupled to metal plasmonic nanocavities, have recently emerged as new platforms for strong light-matter interactions. These systems are expected to have nonlinear optical properties that will enable them to be used as entangled photon sources, compact wave-mixing devices, and other elements for classical and quantum photonic technologies. Here we report the first experimental investigation of the nonlinear properties of these strongly coupled systems, by observing second harmonic generation from a WSe2 monolayer strongly coupled to a single gold nanorod. The pump frequency dependence of the second harmonic signal displays a pronounced splitting that can be explained by a coupled oscillator model with second-order nonlinearities. Rigorous numerical simulations utilizing a nonperturbative nonlinear hydrodynamic model of conduction electrons support this interpretation and reproduce experimental results. Our study thus lays the groundwork for understanding the nonlinear properties of strongly coupled nanoscale systems.

physics.optics↗

arXiv2019

Enhancing second harmonic generation using dipolar-parity modes in non-planar plasmonic nanocavities

There is an active demand to develop efficient nanoscale nonlinear sources for applications in photonic circuitry, quantum optics and biosensing. To this end, plasmonic systems have been utilized to boost the nonlinear signal generation, however high efficiencies of frequency conversion for realistic applications remain a challenge. Metal-insulator-metal (MIM) nanocavities are good candidates for strongly concentrating the fields at the nanoscale to enhance the optical nonlinearities, however they typically suffer from the requirement to have a quadrupolar resonance at the emission wavelength. Here, we introduce nonplanar MIM nanocavities with a nonlinear spacer that can strongly enhance the second harmonic generation (SHG) despite of having fundamental and emission modes of the same parity. Our experimental and numerical results indicate that the enhancement is due to the non-planar design of the cavities and the bulk nonlinearity of the spacer layer.

physics.optics↗