arXiv subjects

Yifan Li

Publications and source records attributed to Yifan Li.

At least 127 records · Page 7Linked to original sources

arXiv2025

GThinker: Towards General Multimodal Reasoning via Cue-Guided Rethinking

Despite notable advancements in multimodal reasoning, leading Multimodal Large Language Models (MLLMs) still underperform on vision-centric multimodal reasoning tasks in general scenarios. This shortfall stems from their predominant reliance on logic- and knowledge-based slow thinking strategies, while effective for domains like math and science, fail to integrate visual information effectively during reasoning. Consequently, these models often fail to adequately ground visual cues, resulting in suboptimal performance in tasks that require multiple plausible visual interpretations and inferences. To address this, we present GThinker (General Thinker), a novel reasoning MLLM excelling in multimodal reasoning across general scenarios, mathematics, and science. GThinker introduces Cue-Rethinking, a flexible reasoning pattern that grounds inferences in visual cues and iteratively reinterprets these cues to resolve inconsistencies. Building on this pattern, we further propose a two-stage training pipeline, including pattern-guided cold start and incentive reinforcement learning, designed to enable multimodal reasoning capabilities across domains. Furthermore, to support the training, we construct GThinker-11K, comprising 7K high-quality, iteratively-annotated reasoning paths and 4K curated reinforcement learning samples, filling the data gap toward general multimodal reasoning. Extensive experiments demonstrate that GThinker achieves 81.5% on the challenging comprehensive multimodal reasoning benchmark M$^3$CoT, surpassing the latest O4-mini model. It also shows an average improvement of 2.1% on general scenario multimodal reasoning benchmarks, while maintaining on-par performance in mathematical reasoning compared to counterpart advanced reasoning models. The code, model, and data will be released soon at https://github.com/jefferyZhan/GThinker.

cs.CV↗

arXiv2025

IndustryEQA: Pushing the Frontiers of Embodied Question Answering in Industrial Scenarios

Existing Embodied Question Answering (EQA) benchmarks primarily focus on household environments, often overlooking safety-critical aspects and reasoning processes pertinent to industrial settings. This drawback limits the evaluation of agent readiness for real-world industrial applications. To bridge this, we introduce IndustryEQA, the first benchmark dedicated to evaluating embodied agent capabilities within safety-critical warehouse scenarios. Built upon the NVIDIA Isaac Sim platform, IndustryEQA provides high-fidelity episodic memory videos featuring diverse industrial assets, dynamic human agents, and carefully designed hazardous situations inspired by real-world safety guidelines. The benchmark includes rich annotations covering six categories: equipment safety, human safety, object recognition, attribute recognition, temporal understanding, and spatial understanding. Besides, it also provides extra reasoning evaluation based on these categories. Specifically, it comprises 971 question-answer pairs generated from small warehouse and 373 pairs from large ones, incorporating scenarios with and without human. We further propose a comprehensive evaluation framework, including various baseline models, to assess their general perception and reasoning abilities in industrial environments. IndustryEQA aims to steer EQA research towards developing more robust, safety-aware, and practically applicable embodied agents for complex industrial environments. Benchmark and codes are available.

cs.CV↗

arXiv2025

A portable diagnosis model for Keratoconus using a smartphone

Keratoconus (KC) is a corneal disorder that results in blurry and distorted vision. Traditional diagnostic tools, while effective, are often bulky, costly, and require professional operation. In this paper, we present a portable and innovative methodology for diagnosing. Our proposed approach first captures the image reflected on the eye's cornea when a smartphone screen-generated Placido disc sheds its light on an eye, then utilizes a two-stage diagnosis for identifying the KC cornea and pinpointing the location of the KC on the cornea. The first stage estimates the height and width of the Placido disc extracted from the captured image to identify whether it has KC. In this KC identification, k-means clustering is implemented to discern statistical characteristics, such as height and width values of extracted Placido discs, from non-KC (control) and KC-affected groups. The second stage involves the creation of a distance matrix, providing a precise localization of KC on the cornea, which is critical for efficient treatment planning. The analysis of these distance matrices, paired with a logistic regression model and robust statistical analysis, reveals a clear distinction between control and KC groups. The logistic regression model, which classifies small areas on the cornea as either control or KC-affected based on the corresponding inter-disc distances in the distance matrix, reported a classification accuracy of 96.94%, which indicates that we can effectively pinpoint the protrusion caused by KC. This comprehensive, smartphone-based method is expected to detect KC and streamline timely treatment.

eess.IV↗

arXiv2025

A Deep Learning-Driven Inhalation Injury Grading Assistant Using Bronchoscopy Images

Inhalation injuries present a challenge in clinical diagnosis and grading due to Conventional grading methods such as the Abbreviated Injury Score (AIS) being subjective and lacking robust correlation with clinical parameters like mechanical ventilation duration and patient mortality. This study introduces a novel deep learning-based diagnosis assistant tool for grading inhalation injuries using bronchoscopy images to overcome subjective variability and enhance consistency in severity assessment. Our approach leverages data augmentation techniques, including graphic transformations, Contrastive Unpaired Translation (CUT), and CycleGAN, to address the scarcity of medical imaging data. We evaluate the classification performance of two deep learning models, GoogLeNet and Vision Transformer (ViT), across a dataset significantly expanded through these augmentation methods. The results demonstrate GoogLeNet combined with CUT as the most effective configuration for grading inhalation injuries through bronchoscopy images and achieves a classification accuracy of 97.8%. The histograms and frequency analysis evaluations reveal variations caused by the augmentation CUT with distribution changes in the histogram and texture details of the frequency spectrum. PCA visualizations underscore the CUT substantially enhances class separability in the feature space. Moreover, Grad-CAM analyses provide insight into the decision-making process; mean intensity for CUT heatmaps is 119.6, which significantly exceeds 98.8 of the original datasets. Our proposed tool leverages mechanical ventilation periods as a novel grading standard, providing comprehensive diagnostic support.

cs.CV↗

arXiv2025

Cardinalities in Height 1

In this article, we give an introduction to the notion of ambidexterity and norm map, and construct inductively the canonical norm map for $m$-truncated maps for some $m\geq-1$, on which the definitions of integration and cardinality are built. We then use several propositions to justify the properties of cardinality and integration and their compatibility with monoidal structure. We give a brief introduction of the definition and behaviors of semiadditive height. Focusing on stable monoidal $p$-local $\infty$-categories of height 1, for any finite group $G$, with the help of Möbius function and Burnside ring, we give an explicit decomposition of the cardinality of $BG$ into an expression of the cardinality of $BC_p$. Eventually, we generalize the result and conclude with a formula of the cardinality of any $π$-finite space $A$.

math.AT↗

arXiv2025

Localized plasmonic meron-antimeron pairs in doubly degenerate orbitals

Topological defects are pivotal in elucidating kaleidoscopic topological phenomena in different physical systems. Meron-antimeron pairs are a type of topological defects firstly found as soliton solutions to SU(2) Yang-Mills equations in gauge theory, and then identified in condensed matter physics as a type of magnetic quasiparticles created in the context of topological charge conservation. Here, we show that isolated meron-antimeron pairs constitute a new form of optical topological quasiparticles that naturally emerge in doubly degenerate orbitals of plasmonic systems, including fundamental and higher-order ones, and their target-type counterparts. We demonstrate that their topological charges are strictly imposed by orbital indices from the doubly degenerate irreducible representations (irreps) of groups consisting of rotational symmetries, and thus are upper-bounded by the orbital indices imposed by group theory. In addition, we find that there exist highly-localized isolated (anti)merons in plasmonic spin textures, which were previously observed mostly in the form of lattices or clusters. We further demonstrate a locking effect between the chirality of the (anti)merons and the parity of the irreps. Then, the topological origins of the revealed topological quasiparticles, i.e., phase, V-point and L-line singularities in plasmonic fields, are investigated. Finally, a complete symmetry classification of the topological quasiparticles is provided. Generalizing the meron-antimeron pairs to photonic systems provides various possibilities for the applications in optical vectorial imaging, deep-subwavelength sensing and metrology.

cond-mat.mes-hall↗

arXiv2025

Are We Merely Justifying Results ex Post Facto? Quantifying Explanatory Inversion in Post-Hoc Model Explanations

Post-hoc explanation methods provide interpretation by attributing predictions to input features. Natural explanations are expected to interpret how the inputs lead to the predictions. Thus, a fundamental question arises: Do these explanations unintentionally reverse the natural relationship between inputs and outputs? Specifically, are the explanations rationalizing predictions from the output rather than reflecting the true decision process? To investigate such explanatory inversion, we propose Inversion Quantification (IQ), a framework that quantifies the degree to which explanations rely on outputs and deviate from faithful input-output relationships. Using the framework, we demonstrate on synthetic datasets that widely used methods such as LIME and SHAP are prone to such inversion, particularly in the presence of spurious correlations, across tabular, image, and text domains. Finally, we propose Reproduce-by-Poking (RBP), a simple and model-agnostic enhancement to post-hoc explanation methods that integrates forward perturbation checks. We further show that under the IQ framework, RBP theoretically guarantees the mitigation of explanatory inversion. Empirically, for example, on the synthesized data, RBP can reduce the inversion by 1.8% on average across iconic post-hoc explanation approaches and domains.

cs.LG↗

arXiv2025

Machine Learning-based Near-field Emitter Location Sensing via Grouped Hybrid Analog and Digital XL-MIMO Receive Array

As a green MIMO structure, the partially-connected hybrid analog and digital (PC-HAD) structure has been widely used in the far-field (FF) scenario for it can significantly reduce the hardware cost and complexity of large-scale or extremely large-scale MIMO (XL-MIMO) array. Recently, near-field (NF) emitter localization including direction-of-arrival (DOA) and range estimations has drawn a lot of attention, but is rarely explored via PC-HAD structure. In this paper, we first analyze the impact of PC-HAD structure on the NF emitter localization and observe that the phase ambiguity (PA) problem caused by PC-HAD structure can be removed inherently with low-latency in the NF scenario. To obtain the exact NF DOA estimation results, we propose a grouped PC-HAD structure, which is capable of dividing the NF DOA estimation problem into multiple FF DOA estimation problems via partitioning the large-scale PC-HAD array into small-scale groups. An angle calibration method is developed to address the inconsistency among these FF DOA estimation problems. Then, to eliminate PA and improve the NF emitter localization performance, we develop three machine learning (ML)-based methods, i.e., two low-complexity data-driven clustering-based methods and one model-driven regression method, namely RegNet. Furthermore, the Cramer-Rao lower bound (CRLB) of NF emitter localization for the proposed grouped PC-HAD structure is derived and reveals that localization performance will decrease with the increasing of the number of groups. The simulation results show that the proposed methods can achieve CRLB at different SNR regions, the RegNet has great performance advantages at low SNR regions and the clustering-based methods have much lower computation complexity.

eess.SP↗

arXiv2025

Spectral Similarity Masks Structural Diversity at Hydrophobic Water Interfaces

The air-water and graphene-water interfaces represent quintessential examples of the liquid-gas and liquid-solid boundaries, respectively. While the sum-frequency generation (SFG) spectra of these interfaces exhibit certain similarities, a consensus on their signals and interpretations has yet to be reached. Leveraging deep learning, we accessed fully first-principles SFG spectra for both systems, addressing recent experimental discrepancies. Despite both interfaces exhibiting microscopically hydrophobic characteristics, our findings reveal that similarities in SFG signals do not translate into comparable interfacial microscopic properties. Instead, graphene-water and air-water interfaces exhibit fundamental differences in SFG-active thicknesses, hydrogen-bonding networks, and dynamic diffusion behavior. These distinctions underscore the stronger confinements imposed by the solid-liquid interface compared with the weaker constraints of the gas-liquid interface.

physics.comp-ph↗

arXiv2025

Monte-Carlo based model for the extraction of oil from oil-water mixtures using wetting and surface acoustic waves

This work presents a Monte Carlo (MC) based microscopic model for simulating the extraction of oil from oil-in-water emulsions under the influence of surface acoustic waves (SAWs). The proposed model is a two-dimensional Ising-lattice gas model that employs Kawasaki dynamics to mimic the interactions between oil, water, and air, as well as external forces such as gravity and acoustic stress. By incorporating both acoustic streaming and acoustic radiation pressure, the model captures key experimental observations, including selective oil extraction and droplet motion under SAW excitation. The results highlight the critical role of acoustic radiation pressure in enabling oil film formation and detachment, governed by the balance between capillary and acoustic stresses. The study provides qualitative agreement with experimental findings and offers insights into the essential mechanisms driving acoustowetting-induced phase separation, demonstrating the utility of discrete modeling for complex fluid dynamics problems.

physics.flu-dyn↗

arXiv2025

Window Token Concatenation for Efficient Visual Large Language Models

To effectively reduce the visual tokens in Visual Large Language Models (VLLMs), we propose a novel approach called Window Token Concatenation (WiCo). Specifically, we employ a sliding window to concatenate spatially adjacent visual tokens. However, directly concatenating these tokens may group diverse tokens into one, and thus obscure some fine details. To address this challenge, we propose fine-tuning the last few layers of the vision encoder to adaptively adjust the visual tokens, encouraging that those within the same window exhibit similar features. To further enhance the performance on fine-grained visual understanding tasks, we introduce WiCo+, which decomposes the visual tokens in later layers of the LLM. Such a design enjoys the merits of the large perception field of the LLM for fine-grained visual understanding while keeping a small number of visual tokens for efficient inference. We perform extensive experiments on both coarse- and fine-grained visual understanding tasks based on LLaVA-1.5 and Shikra, showing better performance compared with existing token reduction projectors. The code is available: https://github.com/JackYFL/WiCo.

cs.CV↗

arXiv2025

Language-based Image Colorization: A Benchmark and Beyond

Image colorization aims to bring colors back to grayscale images. Automatic image colorization methods, which requires no additional guidance, struggle to generate high-quality images due to color ambiguity, and provides limited user controllability. Thanks to the emergency of cross-modality datasets and models, language-based colorization methods are proposed to fully utilize the efficiency and flexibly of text descriptions to guide colorization. In view of the lack of a comprehensive review of language-based colorization literature, we conduct a thorough analysis and benchmarking. We first briefly summarize existing automatic colorization methods. Then, we focus on language-based methods and point out their core challenge on cross-modal alignment. We further divide these methods into two categories: one attempts to train a cross-modality network from scratch, while the other utilizes the pre-trained cross-modality model to establish the textual-visual correspondence. Based on the analyzed limitations of existing language-based methods, we propose a simple yet effective method based on distilled diffusion model. Extensive experiments demonstrate that our simple baseline can produces better results than previous complex methods with 14 times speed up. To the best of our knowledge, this is the first comprehensive review and benchmark on language-based image colorization field, providing meaningful insights for the community. The code is available at https://github.com/lyf1212/Color-Turbo.

cs.CV↗

arXiv2025

Demo of picoRing mouse: an ultra-low-powered wireless mouse ring with ring-to-wristband coil-based impedance sensing

Wireless mouse rings offer subtle, reliable pointing interactions for wearable computing platforms, but the small battery below 27 mAh in the miniature rings restricts the ring's continuous lifespan to just 1-2 hours due to the power consumption of current low-powered wireless communication like BLE. However, the picoRing mouse addresses this by enabling continuous ring-based mouse interaction with ultra-low-powered ring-to-wristband wireless communication through a coil-based impedance sensing method called semi-passive inductive telemetry. This allows a wristband coil to capture a unique frequency response of a nearby ring coil via sensitive inductive coupling, converting the user's mouse input into the unique frequency response via an 820 uW mouse-driven modulation module. Thus, the continuous use of picoRing mouse can potentially last over 92 hours on a single charge of a 20 mAh battery while supporting subtle scrolling and pressing interactions.

cs.HC↗

arXiv2025

DeePMD-kit v3: A Multiple-Backend Framework for Machine Learning Potentials

In recent years, machine learning potentials (MLPs) have become indispensable tools in physics, chemistry, and materials science, driving the development of software packages for molecular dynamics (MD) simulations and related applications. These packages, typically built on specific machine learning frameworks such as TensorFlow, PyTorch, or JAX, face integration challenges when advanced applications demand communication across different frameworks. The previous TensorFlow-based implementation of DeePMD-kit exemplified these limitations. In this work, we introduce DeePMD-kit version 3, a significant update featuring a multi-backend framework that supports TensorFlow, PyTorch, JAX, and PaddlePaddle backends, and demonstrate the versatility of this architecture through the integration of other MLPs packages and of Differentiable Molecular Force Field. This architecture allows seamless backend switching with minimal modifications, enabling users and developers to integrate DeePMD-kit with other packages using different machine learning frameworks. This innovation facilitates the development of more complex and interoperable workflows, paving the way for broader applications of MLPs in scientific research.

physics.chem-ph↗

arXiv2025

Virgo: A Preliminary Exploration on Reproducing o1-like MLLM

Recently, slow-thinking reasoning systems, built upon large language models (LLMs), have garnered widespread attention by scaling the thinking time during inference. There is also growing interest in adapting this capability to multimodal large language models (MLLMs). Given that MLLMs handle more complex data semantics across different modalities, it is intuitively more challenging to implement multimodal slow-thinking systems. To address this issue, in this paper, we explore a straightforward approach by fine-tuning a capable MLLM with a small amount of textual long-form thought data, resulting in a multimodal slow-thinking system, Virgo (Visual reasoning with long thought). We find that these long-form reasoning processes, expressed in natural language, can be effectively transferred to MLLMs. Moreover, it seems that such textual reasoning data can be even more effective than visual reasoning data in eliciting the slow-thinking capacities of MLLMs. While this work is preliminary, it demonstrates that slow-thinking capacities are fundamentally associated with the language model component, which can be transferred across modalities or domains. This finding can be leveraged to guide the development of more powerful slow-thinking reasoning systems. We release our resources at https://github.com/RUCAIBox/Virgo.

cs.CV↗

arXiv2025

Global Public Sentiment on Decentralized Finance: A Spatiotemporal Analysis of Geo-tagged Tweets from 150 Countries

Blockchain technology and decentralized finance (DeFi) are reshaping global financial systems. Despite their impact, the spatial distribution of public sentiment and its economic and geopolitical determinants are often overlooked. This study analyzes over 150 million geo-tagged, DeFi-related tweets from 2012 to 2022, sourced from a larger dataset of 7.4 billion tweets. Using sentiment scores from a BERT-based multilingual classification model, we integrated these tweets with economic and geopolitical data to create a multimodal dataset. Employing techniques like sentiment analysis, spatial econometrics, clustering, and topic modeling, we uncovered significant global variations in DeFi engagement and sentiment. Our findings indicate that economic development significantly influences DeFi engagement, particularly after 2015. Geographically weighted regression analysis revealed GDP per capita as a key predictor of DeFi tweet proportions, with its impact growing following major increases in cryptocurrency values such as bitcoin. While wealthier nations are more actively engaged in DeFi discourse, the lowest-income countries often discuss DeFi in terms of financial security and sudden wealth. Conversely, middle-income countries relate DeFi to social and religious themes, whereas high-income countries view it mainly as a speculative instrument or entertainment. This research advances interdisciplinary studies in computational social science and finance and supports open science by making our dataset and code available on GitHub, and providing a non-code workflow on the KNIME platform. These contributions enable a broad range of scholars to explore DeFi adoption and sentiment, aiding policymakers, regulators, and developers in promoting financial inclusion and responsible DeFi engagement globally.

econ.GN↗

arXiv2025

Images are Achilles' Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Models

In this paper, we study the harmlessness alignment problem of multimodal large language models (MLLMs). We conduct a systematic empirical analysis of the harmlessness performance of representative MLLMs and reveal that the image input poses the alignment vulnerability of MLLMs. Inspired by this, we propose a novel jailbreak method named HADES, which hides and amplifies the harmfulness of the malicious intent within the text input, using meticulously crafted images. Experimental results show that HADES can effectively jailbreak existing MLLMs, which achieves an average Attack Success Rate (ASR) of 90.26% for LLaVA-1.5 and 71.60% for Gemini Pro Vision. Our code and data are available at https://github.com/RUCAIBox/HADES.

cs.CV↗

arXiv2025

DNN-based Enhanced DOA Sensing via Massive MIMO Receiver with Switches-based Hybrid Architecture

Switches-based hybrid architecture has attracted much attention, especially in directional-of-arrival (DOA) sensing, due to its ability of significantly reducing the hardware cost by compressing massive multiple-input multiple-output (MIMO) arrays with switching networks. However, this structure will lead to a degradation in the degrees of freedom (DOF) and accuracy of DOA estimation. To address these two issues, we first propose a switches-based sparse hybrid array (SW-SHA). In this method, we design a dynamic switching network to form a synthesized sparse array, i.e., SW-SHA, that can enlarge the virtual aperture obtained by the difference co-array, thereby significantly enhancing the DOF. Second, in order to improve the DOA estimation accuracy of switches-based hybrid arrays, a deep neural network (DNN)-based method called ASN-DNN is proposed. It includes an antenna selection network (ASN) for optimizing the switch connections based on the criterion of minimizing the Cramer-Rao lower bound (CRLB) under the peak sidelobe level (PSL) constraint and a DNN for DOA estimation. Then by integrating ASN and DNN into an iterative process, the ASN-DNN is obtained. Furthermore, the closed-form expression of CRLB for DOA estimation is derived to evaluate the performance lower bound of switches-based hybrid arrays and provide a benchmark for ASN-DNN. The simulation results show the proposed ASN-DNN can achieve a greater performance than traditional methods, especially in the low signal-to-noise ratio (SNR) regions.

eess.SP↗