arXiv subjects

Kewen Wang

Publications and source records attributed to Kewen Wang.

At least 19 recordsLinked to original sources

arXiv2026

Iterative Multimodal Retrieval-Augmented Generation for Medical Question Answering

Medical retrieval-augmented generation (RAG) systems typically operate on text chunks extracted from biomedical literature, discarding the rich visual content (tables, figures, structured layouts) of original document pages. We propose MED-VRAG, an iterative multimodal RAG framework that retrieves and reasons over PMC document page images instead of OCR'd text. The system pairs ColQwen2.5 patch-level page embeddings with a sharded MapReduce LLM filter, scaling to ~350K pages while keeping Stage-1 retrieval under 30 ms via an offline coarse-to-fine index (C=8 centroids per page, ANN over centroids, exact two-way scoring on the top-R shortlist). A vision-language model (VLM) then iteratively refines its query and accumulates evidence in a memory bank across up to 3 reasoning rounds, with a single iteration costing ~15.9 s and the full three-round pipeline ~47.8 s on 4xA100. Across four medical QA benchmarks (MedQA, MedMCQA, PubMedQA, MMLU-Med), MEDVRAG reaches 78.6% average accuracy. Under controlled comparison with the same Qwen2.5-VL-32B backbone, retrieval contributes a +5.8 point gain over the no-retrieval baseline; we also note a +1.8 point edge over MedRAG + GPT-4 (76.8%), with the caveat that this is a cross-paper rather than head-to-head comparison. Ablations isolate +1.0 from page-image vs text-chunk retrieval, +1.5 from iteration, and +1.0 from the memory bank.

cs.AI↗

arXiv2026

HIME: Mitigating Object Hallucinations in LVLMs via Hallucination Insensitivity Model Editing

Large Vision-Language Models (LVLMs) have demonstrated impressive multimodal understanding capabilities, yet they remain prone to object hallucination, where models describe non-existent objects or attribute incorrect factual information, raising serious concerns for reliable real-world deployment. While fine-tuning is a commonly adopted mitigation strategy, its high computational cost and practical difficulty motivate the need for training-free alternatives, among which model editing has recently emerged as a promising direction. However, indiscriminate editing risks disrupting the rich implicit knowledge encoded in pre-trained LVLMs, leading to a fundamental question: how much intervention is necessary at each layer to suppress hallucinations while preserving pre-trained knowledge? To address this question, we present a systematic analysis of LVLM decoders built on three widely used large language model backbones-Qwen, LLaMA, and Vicuna-revealing clear layer-wise differences in susceptibility to object hallucination. Building on these insights, we introduce the Hallucination Insensitivity Score (HIS), a principled metric that quantifies each layer's sensitivity to hallucination and provides guidance for targeted intervention. Leveraging HIS, we propose Hallucination Insensitivity Model Editing (HIME), a simple yet effective layer-adaptive weight editing approach that selectively modifies latent features to suppress hallucinations while preserving pre-trained knowledge. Extensive experiments demonstrate that HIME reduces hallucinations by an average of 61.8% across open-ended generation benchmarks, including CHAIR, MME, and GPT-4V-aided evaluation, without introducing additional parameters, inference-time latency, or computational overhead.

cs.CV↗

arXiv2026

Bridging the Gap Between Estimated and True Regret Towards Reliable Regret Estimation in Deep Learning based Mechanism Design

Recent advances, such as RegretNet, ALGnet, RegretFormer and CITransNet, use deep learning to approximate optimal multi item auctions by relaxing incentive compatibility (IC) and measuring its violation via ex post regret. However, the true accuracy of these regret estimates remains unclear. Computing exact regret is computationally intractable, and current models rely on gradient based optimizers whose outcomes depend heavily on hyperparameter choices. Through extensive experiments, we reveal that existing methods systematically underestimate actual regret (In some models, the true regret is several hundred times larger than the reported regret), leading to overstated claims of IC and revenue. To address this issue, we derive a lower bound on regret and introduce an efficient item wise regret approximation. Building on this, we propose a guided refinement procedure that substantially improves regret estimation accuracy while reducing computational cost. Our method provides a more reliable foundation for evaluating incentive compatibility in deep learning based auction mechanisms and highlights the need to reassess prior performance claims in this area.

cs.GT↗

arXiv2025

Inductive Learning for Possibilistic Logic Programs Under Stable Models

Possibilistic logic programs (poss-programs) under stable models are a major variant of answer set programming (ASP). While its semantics (possibilistic stable models) and properties have been well investigated, the problem of inductive reasoning has not been investigated yet. This paper presents an approach to extracting poss-programs from a background program and examples (parts of intended possibilistic stable models). To this end, the notion of induction tasks is first formally defined, its properties are investigated and two algorithms ilpsm and ilpsmmin for computing induction solutions are presented. An implementation of ilpsmmin is also provided and experimental results show that when inputs are ordinary logic programs, the prototype outperforms a major inductive learning system for normal logic programs from stable models on the datasets that are randomly generated.

cs.AI↗

arXiv2024

On the $\mathrm{Sp}_{2n}$-distinguished automorphic spectrum of $\mathrm{U}_{2n}$

Given a reductive group $G$ and a reductive subgroup $H$, both defined over a number field $F$, we introduce the notion of the $H$-distinguished automorphic spectrum of $G$ and analyze it for the pair $(\mathrm{U}_{2n},\mathrm{Sp}_{2n})$. We derived a formula for period integrals of pseudo-Eisenstein series of $\mathrm{U}_{2n}$ in analogy with the main result of Lapid and Offen in their work analyzing the pair $(\mathrm{Sp}_{4n},\mathrm{Sp}_{2n} \times \mathrm{Sp}_{2n})$. We give an upper bound of the distinguished spectrum with the formula. A non-trivial lower bound of the discrete distinguished spectrum is expected from the formula, given the previous work.

math.NT↗

arXiv2024

TPCL: Task Progressive Curriculum Learning for Robust Visual Question Answering

Visual Question Answering (VQA) systems are notoriously brittle under distribution shifts and data scarcity. While previous solutions-such as ensemble methods and data augmentation-can improve performance in isolation, they fail to generalise well across in-distribution (IID), out-of-distribution (OOD), and low-data settings simultaneously. We argue that this limitation stems from the suboptimal training strategies employed. Specifically, treating all training samples uniformly-without accounting for question difficulty or semantic structure-leaves the models vulnerable to dataset biases. Thus, they struggle to generalise beyond the training distribution. To address this issue, we introduce Task-Progressive Curriculum Learning (TPCL)-a simple, model-agnostic framework that progressively trains VQA models using a curriculum built by jointly considering question type and difficulty. Specifically, TPCL first groups questions based on their semantic type (e.g., yes/no, counting) and then orders them using a novel Optimal Transport-based difficulty measure. Without relying on data augmentation or explicit debiasing, TPCL improves generalisation across IID, OOD, and low-data regimes and achieves state-of-the-art performance on VQA-CP v2, VQA-CP v1, and VQA v2. It outperforms the most competitive robust VQA baselines by over 5% and 7% on VQA-CP v2 and v1, respectively, and boosts backbone performance by up to 28.5%.

cs.CV↗

arXiv2024

Differentiating Choices via Commonality for Multiple-Choice Question Answering

Multiple-choice question answering (MCQA) becomes particularly challenging when all choices are relevant to the question and are semantically similar. Yet this setting of MCQA can potentially provide valuable clues for choosing the right answer. Existing models often rank each choice separately, overlooking the context provided by other choices. Specifically, they fail to leverage the semantic commonalities and nuances among the choices for reasoning. In this paper, we propose a novel MCQA model by differentiating choices through identifying and eliminating their commonality, called DCQA. Our model captures token-level attention of each choice to the question, and separates tokens of the question attended to by all the choices (i.e., commonalities) from those by individual choices (i.e., nuances). Using the nuances as refined contexts for the choices, our model can effectively differentiate choices with subtle differences and provide justifications for choosing the correct answer. We conduct comprehensive experiments across five commonly used MCQA benchmarks, demonstrating that DCQA consistently outperforms baseline models. Furthermore, our case study illustrates the effectiveness of the approach in directing the attention of the model to more differentiating features.

cs.CL↗

arXiv2022

Maximising the Influence of Temporary Participants in Opinion Formation

DeGroot-style opinion formation presumes a continuous interaction among agents of a social network. Hence, it cannot handle agents external to the social network that interact only temporarily with the permanent ones. Many real-world organisations and individuals fall into such a category. For instance, a company tries to persuade as many as possible to buy its products and, due to various constraints, can only exert its influence for a limited amount of time. We propose a variant of the DeGroot model that allows an external agent to interact with the permanent ones for a preset period of time. We obtain several insights on maximising an external agent's influence in opinion formation by analysing and simulating the variant.

cs.MA↗

arXiv2022

Function-words Enhanced Attention Networks for Few-Shot Inverse Relation Classification

The relation classification is to identify semantic relations between two entities in a given text. While existing models perform well for classifying inverse relations with large datasets, their performance is significantly reduced for few-shot learning. In this paper, we propose a function words adaptively enhanced attention framework (FAEA) for few-shot inverse relation classification, in which a hybrid attention model is designed to attend class-related function words based on meta-learning. As the involvement of function words brings in significant intra-class redundancy, an adaptive message passing mechanism is introduced to capture and transfer inter-class differences.We mathematically analyze the negative impact of function words from dot-product measurement, which explains why message passing mechanism effectively reduces the impact. Our experimental results show that FAEA outperforms strong baselines, especially the inverse relation accuracy is improved by 14.33% under 1-shot setting in FewRel1.0.

cs.CL↗

arXiv2019

Transverse photon spin beyond interfaces

Photons possess spin degree of freedom, corresponding to clockwise and counter clockwise rotating direction of the fields. Photon spin plays an important role in various applications such as optical communications, information processing and sensing. In conventional isotropic media, photon spin is aligned with the propagation direction of light, obeying spin momentum locking. Interestingly, at certain interfaces, the surface waves decaying away from the interface possess a photon spin transverse to its propagation, opening exciting opportunities for observation of spin dependent unidirectional excitation in confined systems. Here we propose and realize transverse photon spin (T-spin) in the interior of a bulk medium, without relying on the presence of any interfaces. We show the complete mapping of the T-spin of surface modes to that of the bulk modes by introducing the coupling between electric and magnetic responses along orthogonal directions, i.e., the bianisotropy, into the medium. We further discover that an interface formed by two bianisotropic media of opposite orientations supports edge-dependent propagating modes with tunable cutoff frequencies. Our results provide a new platform for manipulating the spin orbit interaction of electromagnetic waves.

physics.optics↗

arXiv2018

Knowledge Compilation in Multi-Agent Epistemic Logics

Epistemic logics are a primary formalism for multi-agent systems but major reasoning tasks in such epistemic logics are intractable, which impedes applications of multi-agent epistemic logics in automatic planning. Knowledge compilation provides a promising way of resolving the intractability by identifying expressive fragments of epistemic logics that are tractable for important reasoning tasks such as satisfiability and forgetting. The property of logical separability allows to decompose a formula into some of its subformulas and thus modular algorithms for various reasoning tasks can be developed. In this paper, by employing logical separability, we propose an approach to knowledge compilation for the logic Kn by defining a normal form SDNF. Among several novel results, we show that every epistemic formula can be equivalently compiled into a formula in SDNF, major reasoning tasks in SDNF are tractable, and formulas in SDNF enjoy the logical separability. Our results shed some lights on modular approaches to knowledge compilation. Furthermore, we apply our results in the multi-agent epistemic planning. Finally, we extend the above result to the logic K45n that is Kn extended by introspection axioms 4 and 5.

cs.AI↗

arXiv2018

Giant Asymmetric Radiation from an Ultrathin Bianisotropic Metamaterial

Unidirectional radiation is of particular interest in high-power lasing and optics. Commonly, however, it is difficult to achieve a unidirectional profile in such a system without breaking reciprocity. Recently, assisted by metamaterials without structural symmetry, antennas that radiate asymmetrically have been developed, hence providing the possibility of achieving unidirectionality. Nevertheless, it has been challenging to achieve extremely high radiation asymmetry in such antennas. Here, we demonstrate that this radiation asymmetry is further enhanced when magnetic plasmons are present in the metamaterials. Experimentally, we show that a thin metamaterial with a thickness of approximately λ_0/8 can exhibit a forward-to-backward emission asymmetry of up to 1:32 without any optimization. Our work paves the way for manipulating asymmetric radiation by means of metamaterials and may have a variety of promising applications, such as directional optical and quantum emitters, lasers, and absorbers.

physics.app-ph↗

arXiv2017

Syntax-Preserving Belief Change Operators for Logic Programs

Recent methods have adapted the well-established AGM and belief base frameworks for belief change to cover belief revision in logic programs. In this study here, we present two new sets of belief change operators for logic programs. They focus on preserving the explicit relationships expressed in the rules of a program, a feature that is missing in purely semantic approaches that consider programs only in their entirety. In particular, operators of the latter class fail to satisfy preservation and support, two important properties for belief change in logic programs required to ensure intuitive results. We address this shortcoming of existing approaches by introducing partial meet and ensconcement constructions for logic program belief change, which allow us to define syntax-preserving operators that satisfy preservation and support. Our work is novel in that our constructions not only preserve more information from a logic program during a change operation than existing ones, but they also facilitate natural definitions of contraction operators, the first in the field to the best of our knowledge. In order to evaluate the rationality of our operators, we translate the revision and contraction postulates from the AGM and belief base frameworks to the logic programming setting. We show that our operators fully comply with the belief base framework and formally state the interdefinability between our operators. We further propose an algorithm that is based on modularising a logic program to reduce partial meet and ensconcement revisions or contractions to performing the operation only on the relevant modules of that program. Finally, we compare our approach to two state-of-the-art logic program revision methods and demonstrate that our operators address the shortcomings of one and generalise the other method.

cs.AI↗

arXiv2015

A Distance-based Paraconsistent Semantics for DL-Lite

DL-Lite is an important family of description logics. Recently, there is an increasing interest in handling inconsistency in DL-Lite as the constraint imposed by a TBox can be easily violated by assertions in ABox in DL-Lite. In this paper, we present a distance-based paraconsistent semantics based on the notion of feature in DL-Lite, which provides a novel way to rationally draw meaningful conclusions even from an inconsistent knowledge base. Finally, we investigate several important logical properties of this entailment relation based on the new semantics and show its promising advantages in non-monotonic reasoning for DL-Lite.

cs.AI↗

arXiv2015

Preferential Multi-Context Systems

Multi-context systems (MCS) presented by Brewka and Eiter can be considered as a promising way to interlink decentralized and heterogeneous knowledge contexts. In this paper, we propose preferential multi-context systems (PMCS), which provide a framework for incorporating a total preorder relation over contexts in a multi-context system. In a given PMCS, its contexts are divided into several parts according to the total preorder relation over them, moreover, only information flows from a context to ones of the same part or less preferred parts are allowed to occur. As such, the first $l$ preferred parts of an PMCS always fully capture the information exchange between contexts of these parts, and then compose another meaningful PMCS, termed the $l$-section of that PMCS. We generalize the equilibrium semantics for an MCS to the (maximal) $l_{\leq}$-equilibrium which represents belief states at least acceptable for the $l$-section of an PMCS. We also investigate inconsistency analysis in PMCS and related computational complexity issues.

cs.AI↗

arXiv2014

Random Logic Programs: Linear Model

This paper proposes a model, the linear model, for randomly generating logic programs with low density of rules and investigates statistical properties of such random logic programs. It is mathematically shown that the average number of answer sets for a random program converges to a constant when the number of atoms approaches infinity. Several experimental results are also reported, which justify the suitability of the linear model. It is also experimentally shown that, under this model, the size distribution of answer sets for random programs tends to a normal distribution when the number of atoms is sufficiently large.

cs.AI↗

arXiv2014

An Approach to Forgetting in Disjunctive Logic Programs that Preserves Strong Equivalence

In this paper we investigate forgetting in disjunctive logic programs, where forgetting an atom from a program amounts to a reduction in the signature of that program. The goal is to provide an approach that is syntax-independent, in that if two programs are strongly equivalent, then the results of forgetting an atom in each program should also be strongly equivalent. Our central definition of forgetting is impractical but satisfies this goal: Forgetting an atom is characterised by the set of SE consequences of the program that do not mention the atom to be forgotten. We then provide an equivalent, practical definition, wherein forgetting an atom $p$ is given by those rules in the program that don't mention $p$, together with rules obtained by a single inference step from rules that do mention $p$. Forgetting is shown to have appropriate properties; as well, the finite characterisation results in a modest (at worst quadratic) blowup. Finally we have also obtained a prototype implementation of this approach to forgetting.

cs.AI↗

arXiv2003

A semantic framework for preference handling in answer set programming

We provide a semantic framework for preference handling in answer set programming. To this end, we introduce preference preserving consequence operators. The resulting fixpoint characterizations provide us with a uniform semantic framework for characterizing preference handling in existing approaches. Although our approach is extensible to other semantics by means of an alternating fixpoint theory, we focus here on the elaboration of preferences under answer set semantics. Alternatively, we show how these approaches can be characterized by the concept of order preservation. These uniform semantic characterizations provide us with new insights about interrelationships and moreover about ways of implementation.

cs.AI↗