arXiv subjects

Minsu Park

Publications and source records attributed to Minsu Park.

At least 19 recordsLinked to original sources

arXiv2026

On-Screen Inertia: Persistent Racial and Gender Disparities in Hollywood Film (1900-2024)

Hollywood has diversified its casts. Whether this has translated into structural change in how those actors are positioned within narratives remains largely unexamined. Drawing on 76,815 U.S. English-language films (1900-2024) and over 3.1 million cast and crew entries, we move beyond headcounts to examine long-term inclusion trends through network centrality, occupational stereotypes, crew-to-cast diversity pathways, and financial outcomes. We find evidence of what we term on-screen inertia. While the raw inclusion of women and racial minorities has increased modestly, White actors have become more overrepresented relative to the U.S. Census in recent decades, not less. Within the visibility layer, women face a consistent longevity penalty with significantly shorter careers than men, and visual depictions framing men as dominant and women as sensual have remained stable since the 1950s. Structurally, White actors retain disproportionate network centrality; women achieve parity in centrality and lead billing yet cluster in secondary co-lead roles; and occupational stereotypes anchoring racial and gender groups to specific labor categories persist largely unchanged across the pre- and post-2000 periods. Crew diversity associates with cast inclusion only along matching demographic lines (i.e., racial with racial, gender with gender) and does not extend to narrative centrality, revealing a structural ceiling on hiring-based interventions. Critically, we find no consistent market penalty for diversity across decades of box office returns and audience ratings, eliminating the primary rationalization for these practices. Together, these findings demonstrate that Hollywood's representational inequalities are not a rational market response, but are an institutionally sustained choice.

cs.CY↗

arXiv2026

Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models

A central aspiration of mechanistic interpretability is controllability: if we know where a behavior is represented in a model's activations, we should be able to modify it. This rests on a hidden premise -- that the direction which detects a behavior and the direction which controls it are the same, or close. We test this geometrically: what is the angle between the direction that best detects a behavior and the one that best causes it? If detection implies control the cosine is near 1; otherwise it quantifies a detection-intervention gap. On Gemma 2-2B-it, output format (clean JSON vs markdown fencing) collapses both roles onto one axis. Hallucination does not: the model detects fake entities with perfect linear separability (AUC = 1.000 from layer 5), yet that direction sits at cos = 0.12 (about 83 degrees) from the direction producing a refusal -- a small, reproducible alignment, far from the cos = 1 that "detection is control" would require. A detector built from activations, with no chosen tokens, likewise fails to align (cos = -0.06). The gap generalizes: across four models from three families and two scales (1B-9B), cos stays in [0.12, 0.20], identical before and after instruction tuning (0.1197 vs 0.1200), placing its origin in pretraining. A 15-degree rotation toward the refusal direction partially bridges it -- 73% and 60% refusal on two held-out fake-entity categories at 1.8% false positives. We then ask whether this cosine predicts steerability, and it does not: detection is a high-dimensional class, not a single direction, and what separates the steerable case is functional, not readable from a static angle. The cosine is a weight-computable signature of the dissociation between knowing and steering, not a predictor of it.

cs.CL↗

arXiv2026

Do Melody and Rhythm Coevolve?

Music comprises two core structural components, melody and rhythm, that vary widely across cultures. Whether these components coevolve in a coupled way or follow independent trajectories remains unclear. We introduce a novel computational pipeline to extract vocal melodic pitch-interval and percussive inter-onset timing distributions from 27,628 popular songs across 59 countries, enabling large-scale cross-cultural comparison that bypasses traditional music annotations. Musical similarities between countries aligned with geographic and linguistic relationships, validating our approach. Substantial variation emerged in both melodic and rhythmic structures across countries, yet the diversity of the two components was not significantly correlated, challenging assumptions of coupled evolution. Only rhythmic diversity was significantly associated with ethnic and linguistic heterogeneity, while melodic diversity showed no such association. These findings suggest that melody and rhythm constitute partially independent systems shaped by distinct cultural and evolutionary pressures, rather than components of a single monolithic musical style.

cs.SD↗

arXiv2026

When Correct Isn't Usable: Improving Structured Output Reliability in Small Language Models

Deployed language models must produce outputs that are both correct and format-compliant. We study this structured-output reliability gap using two mathematical benchmarks -- GSM8K and MATH -- as a controlled testbed: ground truth is unambiguous and the output contract is strict (JSON with required fields). We evaluate three 7-9B models under five prompting strategies and report output accuracy -- the joint event of mathematical correctness and valid JSON structure -- as the primary metric. A systematic format failure emerges: NAIVE prompting (no system prompt) achieves up to 85% task accuracy on GSM8K but 0% output accuracy across all models and datasets. REFERENCE prompting (a minimal hand-written JSON format prompt) fares little better, yielding 0% output accuracy for two of four models tested. Constrained decoding enforces syntactic validity but incurs 3.6x-8.2x latency overhead and in several settings degrades task performance substantially. To overcome this limitation, we developed AloLab, an iterative system-prompt optimizer (meta-agent: Claude Sonnet 4.5) requiring only black-box API access to the target model; it reaches 84-87% output accuracy on GSM8K and 34-40% on MATH across five independent runs per model, with 29/30 paired McNemar comparisons against the best static prompt significant at p < 0.05, at near-NAIVE inference latency and without model fine-tuning. The same format failure extends to GPT-4o (OpenAI, 2024), a proprietary closed-source model: REFERENCE achieves 0% output accuracy due to systematic markdown-fence wrapping, while AloLab reaches 95.2% [94.8, 95.6]. An ablation replacing the Sonnet 4.5 meta-agent with Claude 3 Haiku reduces mean output accuracy to 61.0% and increases run-to-run standard deviation from <1 pp to 21.8 pp, confirming that meta-agent capability is a primary driver of optimization quality.

cs.CL↗

arXiv2025

MAVIS: A Benchmark for Multimodal Source Attribution in Long-form Visual Question Answering

Source attribution aims to enhance the reliability of AI-generated answers by including references for each statement, helping users validate the provided answers. However, existing work has primarily focused on text-only scenario and largely overlooked the role of multimodality. We introduce MAVIS, the first benchmark designed to evaluate multimodal source attribution systems that understand user intent behind visual questions, retrieve multimodal evidence, and generate long-form answers with citations. Our dataset comprises 157K visual QA instances, where each answer is annotated with fact-level citations referring to multimodal documents. We develop fine-grained automatic metrics along three dimensions of informativeness, groundedness, and fluency, and demonstrate their strong correlation with human judgments. Our key findings are threefold: (1) LVLMs with multimodal RAG generate more informative and fluent answers than unimodal RAG, but they exhibit weaker groundedness for image documents than for text documents, a gap amplified in multimodal settings. (2) Given the same multimodal documents, there is a trade-off between informativeness and groundedness across different prompting methods. (3) Our proposed method highlights mitigating contextual bias in interpreting image documents as a crucial direction for future research.

cs.CV↗

arXiv2025

LASTIST: LArge-Scale Target-Independent STance dataset

Stance detection has emerged as an area of research in the field of artificial intelligence. However, most research is currently centered on the target-dependent stance detection task, which is based on a person's stance in favor of or against a specific target. Furthermore, most benchmark datasets are based on English, making it difficult to develop models in low-resource languages such as Korean, especially for an emerging field such as stance detection. This study proposes the LArge-Scale Target-Independent STance (LASTIST) dataset to fill this research gap. Collected from the press releases of both parties on Korean political parties, the LASTIST dataset uses 563,299 labeled Korean sentences. We provide a detailed description of how we collected and constructed the dataset and trained state-of-the-art deep learning and stance detection models. Our LASTIST dataset is designed for various tasks in stance detection, including target-independent stance detection and diachronic evolution stance detection. We deploy our dataset on https://anonymous.4open.science/r/LASTIST-3721/.

cs.CL↗

arXiv2025

Data Compression with Noise Suppression for Inference under Noisy Covariance

In many fields including cosmology, statistical inference often relies on Gaussian likelihoods whose covariance matrices are estimated from a finite number of simulations. This finite-sample estimation introduces noise into the covariance, which propagates to parameter estimates, a phenomenon known as the Dodelson-Schneider (DS) effect, leading to inflated uncertainties. While the Massively Optimized Parameter Estimation and Data compression (MOPED) algorithm offers lossless Fisher information-preserving compression, it does not mitigate the DS effect when the compression matrix itself is derived from noisy covariances. In this paper, we propose a modified compression scheme, powered MOPED ($p$-MOPED), which suppresses noise propagation by balancing information retention and covariance estimate noise reduction through a tunable power-law transformation of the sample correlation matrix. We test $p$-MOPED against standard and diagonal MOPED on toy models and on cosmological data from the Subaru Hyper Suprime-Cam Year 3 weak lensing survey. Our results demonstrate that $p$-MOPED consistently outperforms other approaches, especially in regimes with limited simulations, offering a robust compression strategy for high-dimensional data analyses under practical constraints.

astro-ph.CO↗

arXiv2025

Solving Copyright Infringement on Short Video Platforms: Novel Datasets and an Audio Restoration Deep Learning Pipeline

Short video platforms like YouTube Shorts and TikTok face significant copyright compliance challenges, as infringers frequently embed arbitrary background music (BGM) to obscure original soundtracks (OST) and evade content originality detection. To tackle this issue, we propose a novel pipeline that integrates Music Source Separation (MSS) and cross-modal video-music retrieval (CMVMR). Our approach effectively separates arbitrary BGM from the original OST, enabling the restoration of authentic video audio tracks. To support this work, we introduce two domain-specific datasets: OASD-20K for audio separation and OSVAR-160 for pipeline evaluation. OASD-20K contains 20,000 audio clips featuring mixed BGM and OST pairs, while OSVAR-160 is a unique benchmark dataset comprising 1,121 video and mixed-audio pairs, specifically designed for short video restoration tasks. Experimental results demonstrate that our pipeline not only removes arbitrary BGM with high accuracy but also restores OSTs, ensuring content integrity. This approach provides an ethical and scalable solution to copyright challenges in user-generated content on short video platforms.

cs.MM↗

arXiv2025

Text-Driven Video Style Transfer with State-Space Models: Extending StyleMamba for Temporal Coherence

StyleMamba has recently demonstrated efficient text-driven image style transfer by leveraging state-space models (SSMs) and masked directional losses. In this paper, we extend the StyleMamba framework to handle video sequences. We propose new temporal modules, including a \emph{Video State-Space Fusion Module} to model inter-frame dependencies and a novel \emph{Temporal Masked Directional Loss} that ensures style consistency while addressing scene changes and partial occlusions. Additionally, we introduce a \emph{Temporal Second-Order Loss} to suppress abrupt style variations across consecutive frames. Our experiments on DAVIS and UCF101 show that the proposed approach outperforms competing methods in terms of style consistency, smoothness, and computational efficiency. We believe our new framework paves the way for real-time text-driven video stylization with state-of-the-art perceptual results.

cs.GR↗

arXiv2025

GlobalMood: A cross-cultural benchmark for music emotion recognition

Human annotations of mood in music are essential for music generation and recommender systems. However, existing datasets predominantly focus on Western songs with terms derived from English, which may limit generalizability across diverse linguistic and cultural backgrounds. We introduce 'GlobalMood', a novel cross-cultural benchmark dataset comprising 1,180 songs sampled from 59 countries, with large-scale annotations collected from 2,519 individuals across five culturally and linguistically distinct locations: U.S., France, Mexico, S. Korea, and Egypt. Rather than imposing predefined emotion and mood categories, we implement a bottom-up, participant-driven approach to organically elicit culturally specific music-related emotion terms. We then recruit another pool of human participants to collect 988,925 ratings for these culture-specific descriptors. Our analysis confirms the presence of a valence-arousal structure shared across cultures, yet also reveals significant divergences in how certain emotion terms (despite being dictionary equivalents) are perceived cross-culturally. State-of-the-art multimodal models benefit substantially from fine-tuning on our cross-culturally balanced dataset, particularly in non-English contexts. Broadly, our findings inform the ongoing debate on the universality versus cultural specificity of emotional descriptors, and our methodology can contribute to other multimodal and cross-lingual research.

cs.IR↗

arXiv2025

Visual and Auditory Aesthetic Preferences Across Cultures

Research on how humans perceive aesthetics in shapes, colours, and music has predominantly focused on Western populations, limiting our understanding of how cultural environments shape aesthetic preferences. We present a large-scale cross-cultural study examining aesthetic preferences across five distinct modalities extensively explored in the literature: shape, curvature, colour, musical harmony and melody. We gather 401,403 preference judgements from 4,835 participants across 10 countries, systematically sampling two-dimensional parameter spaces for each modality. The findings reveal both universal patterns and cultural variations. Preferences for shape and curvature cross-culturally demonstrate a consistent preference for symmetrical forms. While colour preferences are categorically consistent, ratio-like preferences vary across cultures. Musical harmony shows strong agreement in interval relationships despite differing regions of preference within the broad frequency spectrum, while melody shows the highest cross-cultural variation. These results suggest that aesthetic preferences emerge from an interplay between shared perceptual mechanisms and cultural learning.

cs.MM↗

arXiv2025

A Cholesky decomposition-based asset selection heuristic for sparse tangent portfolio optimization

In practice, including large number of assets in mean-variance portfolios can lead to higher transaction costs and management fees. To address this, one common approach is to select a smaller subset of assets from the larger pool, constructing more efficient portfolios. As a solution, we propose a new asset selection heuristic which generates a pre-defined list of asset candidates using a surrogate formulation and re-optimizes the cardinality-constrained tangent portfolio with these selected assets. This method enables faster optimization and effectively constructs portfolios with fewer assets, as demonstrated by numerical analyses on historical stock returns. Finally, we discuss a quantitative metric that can provide a initial assessment of the performance of the proposed heuristic based on asset covariance.

q-fin.MF↗

arXiv2025

Locally Convex Global Loss Network for Decision-Focused Learning

In decision-making problems under uncertainty, predicting unknown parameters is often considered independent of the optimization part. Decision-focused learning (DFL) is a task-oriented framework that integrates prediction and optimization by adapting the predictive model to give better decisions for the corresponding task. Here, an inevitable challenge arises when computing the gradients of the optimal decision with respect to the parameters. Existing research copes with this issue by smoothly reforming surrogate optimization or constructing surrogate loss functions that mimic task loss. However, they are applied to restricted optimization domains. In this paper, we propose Locally Convex Global Loss Network (LCGLN), a global surrogate loss model that can be implemented in a general DFL paradigm. LCGLN learns task loss via a partial input convex neural network which is guaranteed to be convex for chosen inputs while keeping the non-convex global structure for the other inputs. This enables LCGLN to admit general DFL through only a single surrogate loss without any sense for choosing appropriate parametric forms. We confirm the effectiveness and flexibility of LCGLN by evaluating our proposed model with three stochastic decision-making problems.

cs.LG↗

arXiv2025

Dimensionality Reduction Techniques for Statistical Inference in Cosmology

We explore linear and non-linear dimensionality reduction techniques for statistical inference of parameters in cosmology. Given the importance of compressing the increasingly complex data vectors used in cosmology, we address questions that impact the constraining power achieved, such as: Are currently used methods effectively lossless? Under what conditions do nonlinear methods, typically based on neural nets, outperform linear methods? Through theoretical analysis and experiments with simulated weak lensing data vectors we compare three standard linear methods and neural network based methods. We propose two linear methods that outperform all others while using less computational resources: a variation of the MOPED algorithm we call e-MOPED and an adaptation of Canonical Correlation Analysis (CCA), which is a method new to cosmology but well known in statistics. Both e-MOPED and CCA utilize simulations spanning the full parameter space, and rely on the sensitivity of the data vector to the parameters of interest. The gains we obtain are significant compared to compression methods used in the literature: up to 30% in the Figure of Merit for $Ω_m$ and $S_8$ in a realistic Simulation Based Inference analysis that includes statistical and systematic errors. We also recommend two modifications that improve the performance of all methods: First, include components in the compressed data vector that may not target the key parameters but still enhance the constraints on due to their correlations. The gain is significant, above 20% in the Figure of Merit. Second, compress Gaussian and non-Gaussian statistics separately -- we include two summary statistics of each type in our analysis.

astro-ph.CO↗

arXiv2024

Dynamic orders of a Quantum Spin Liquid at Non-zero Temperatures

A quantum spin liquid hosts massive quantum entanglement whose identification is one of the most significant problems in physics. Yet, its detection is known to be notoriously difficult because of featureless properties without a symmetry order parameter. Here, we demonstrate dynamic signatures of a quantum spin liquid state by investigating Kitaev's spin model on the hyper-honeycomb lattice, where a quantum spin liquid state is stabilized as a stable thermodynamic phase. The real-time dynamics of spin correlation function is obtained with the large-scale quantum Monte Carlo simulation. We find the onset of a characteristic oscillation in dynamic local spin correlation as entering the quantum spin liquid phase. Our results show that a quantum spin liquid may be characterized by a sharp growth of coherent spin dynamics of the system, which we name as a dynamic order. We further propose that a dynamic-order may naturally detect a featureless thermal phase transition, which has been reported in a class of strongly correlated materials.

cond-mat.str-el↗

arXiv2024

Inclusive content reduces racial and gender biases, yet non-inclusive content dominates popular culture

Images are often termed as representations of perceived reality. As such, racial and gender biases in popular culture and visual media could play a critical role in shaping people's perceptions of society. While previous research has made significant progress in exploring the frequency and discrepancies in racial and gender group appearances in visual media, it has largely overlooked important nuances in how these groups are portrayed, as it lacked the ability to systematically capture such complexities at scale over time. To address this gap, we examine two media forms of varying target audiences, namely fashion magazines and movie posters. Accordingly, we collect a large dataset comprising over 300,000 images spanning over five decades and utilize state-of-the-art machine learning models to classify not only race and gender but also the posture, expressed emotional state, and body composition of individuals featured in each image. We find that racial minorities appear far less frequently than their White counterparts, and when they do appear, they are portrayed less prominently. We also find that women are more likely to be portrayed with their full bodies, whereas men are more frequently presented with their faces. Finally, through a series of survey experiments, we find evidence that exposure to inclusive content can help reduce biases in perceptions of minorities, while racially and gender-homogenized content may reinforce and amplify such biases. Taken together, our findings highlight that racial and gender biases in visual media remain pervasive, potentially exacerbating existing stereotypes and inequalities.

cs.CY↗

arXiv2024

Improving Multi-lingual Alignment Through Soft Contrastive Learning

Making decent multi-lingual sentence representations is critical to achieve high performances in cross-lingual downstream tasks. In this work, we propose a novel method to align multi-lingual embeddings based on the similarity of sentences measured by a pre-trained mono-lingual embedding model. Given translation sentence pairs, we train a multi-lingual model in a way that the similarity between cross-lingual embeddings follows the similarity of sentences measured at the mono-lingual teacher model. Our method can be considered as contrastive learning with soft labels defined as the similarity between sentences. Our experimental results on five languages show that our contrastive loss with soft labels far outperforms conventional contrastive loss with hard labels in various benchmarks for bitext mining tasks and STS tasks. In addition, our method outperforms existing multi-lingual embeddings including LaBSE, for Tatoeba dataset. The code is available at https://github.com/YAI12xLinq-B/IMASCL

cs.CL↗

arXiv2023

Perception, performance, and detectability of conversational artificial intelligence across 32 university courses

The emergence of large language models has led to the development of powerful tools such as ChatGPT that can produce text indistinguishable from human-generated work. With the increasing accessibility of such technology, students across the globe may utilize it to help with their school work -- a possibility that has sparked discussions on the integrity of student evaluations in the age of artificial intelligence (AI). To date, it is unclear how such tools perform compared to students on university-level courses. Further, students' perspectives regarding the use of such tools, and educators' perspectives on treating their use as plagiarism, remain unknown. Here, we compare the performance of ChatGPT against students on 32 university-level courses. We also assess the degree to which its use can be detected by two classifiers designed specifically for this purpose. Additionally, we conduct a survey across five countries, as well as a more in-depth survey at the authors' institution, to discern students' and educators' perceptions of ChatGPT's use. We find that ChatGPT's performance is comparable, if not superior, to that of students in many courses. Moreover, current AI-text classifiers cannot reliably detect ChatGPT's use in school work, due to their propensity to classify human-written answers as AI-generated, as well as the ease with which AI-generated text can be edited to evade detection. Finally, we find an emerging consensus among students to use the tool, and among educators to treat this as plagiarism. Our findings offer insights that could guide policy discussions addressing the integration of AI into educational frameworks.

cs.CY↗