Abstract

Central bank communications influence market expectations, but their
linguistic attributes are inadequately examined in emerging markets. We examine
337 Reserve Bank of India Monetary Policy Committee documents (2010 2025),
pairing the DAST framework for semantic complexity with a fine-tuned FinBERT
that we validate directly on human-annotated central-bank text ($81.3%$
accuracy, macro-$F_1 = 0.784$, Cohen's $kappa = 0.714$ against a human
majority vote), and relate the resulting textual features to minute-level and
daily volatility in NIFTY indices. Within a pre-specified nine-test confirmatory family, negative-sentiment
probability is associated with event-window equity volatility and CLS-based
sentiment with event-window turbulence in India VIX; both survive
Benjamini Hochberg correction, rank correlation, bootstrap intervals and
regime-conditional controls, and both strengthen under a time-matched
volatility baseline. A third candidate, attention concentration, is
Bonferroni-significant under Pearson correlation yet reverses sign under
Spearman and has a bootstrap interval spanning zero; we report it as a
cautionary methodological result rather than a finding. A pre-specified
February~20 Chow test and an endogenous break search near March~11, 2020
indicate pronounced COVID-period regime dependence. At the daily horizon a
sentiment$times$complexity interaction yields only small exploratory
in-sample gains that are not significant under heteroskedasticity-robust
inference and do not improve out-of-sample prediction, even after controlling
for document length. Textual features add in-sample explanatory power over a
GARCH plus policy-rate baseline but no incremental out-of-sample predictive
value on any index.

Abstract

This paper examines the cascaded deployment of
beyond diagonal (BD) reconfigurable intelligent surfaces (RISs)
and explores its potential to enhance the performance of MIMO
systems. We first derive the jointly optimal closed-form solutions
for the RISs in cascade with SVD-water-filling (SVD-WF) and
uniform power allocation (UPA) precoding strategies. The optimally
configured cascaded-RIS with UPA is shown to achieve
performance comparable to that with the SVD-WF approach,
suggesting that cascaded-RISs can also aid in reducing transmitter
complexity. Furthermore, we derive the approximate ergodic
capacity for Rician fading and obtain its high-SNR closed form
approximation for Rayleigh. These analytical results reveal key
tradeoffs among transmit power, RIS size, and achievable capacity,
offering practical deployment guidelines for high-SNR cascaded-
RIS MIMO systems.

Abstract

Abstract--Camera-based water-level monitoring provides a low-cost alternative to vulnerable physical sensors but often requires extensive site-specific calibration. To address this, an end-to-end image classification framework is proposed to estimate water levels directly from staff-gauge images. To validate generalizability, cameras were deployed across five diverse water bodies to collect an extensive multi-site dataset. The architecture employs a ResNet-18 backbone with four feature pooling variants: Global Average Pooling (GAP), Spatial Attention Pooling (SAP), Feature Pyramid Network (FPN), and a combination of SAP and FPN (SAP+FPN). Performance was evaluated in four phases: First, it surpassed the recent Adaptive-SAM model on the public Mishui Lake benchmark. Second, testing on the custom multi-site dataset demonstrated consistent accuracy across highly variable environmental conditions. Third, the classification formulation significantly reduced estimation error compared to traditional regression models. Finally, the ResNet-18 backbone paired with FPN outperformed alternative models, including ConvNeXt, EfficientNet-B0, MobileNetV3, and ViT. These results demonstrate that robust, scalable water-level monitoring can be achieved purely from images without specialized sensor infrastructure.

Abstract

As Large Language Models(LLMs)areincreas ingly deployed in multilingual contexts, evalu ating their fairness across languages is essen tial. We investigate occupational gender bias in 16 LLMs across eight Indic languages us ing a controlled coreference resolution bench mark that employs transliterated English oc cupational loanwords to minimize grammati cal gender cues. In addition to this generative evaluation, we compute intrinsic bias metrics (WEATand DirectBias) from the models' em bedding spaces. Across models, we observe systematic disparities between pro-stereotype Competence vs. Bias Across Languages 0.6 0.4 0.2 Accuracy Gap (%) -- Stereotype Bias 0.0 0.5 0.6 0.7 Control Accuracy (%) -- Syntactic Competence 0.9 0.8 1.0 Language Bengali Hindi Kannada Malayalam Marathi Odiya Tamil Telugu Figure 1: Relationship between syntactic competence (control accuracy) and stereotype bias (accuracy gap) across model-language pairs. Each point represents one and anti-stereotype conditions. These gaps are largest in high-accuracy models, whereas mod els lacking baseline competence exhibit smaller gaps driven by general prediction uncertainty. Error analysis shows a consistent asymmetry, with higher error rates when the correct refer ent is female. However, intrinsic embedding bias metrics show no statistically significant correlation with these generative disparities (p > 0.05). These results suggest that strong linguistic competence does not necessarily cor respond to stereotype-neutral generation and that intrinsic embedding metrics alone may not reliably predict bias in multilingual language model behavior

Abstract

Pretrained Vision Transformers (ViTs) such as DINOv2 and MAE provide generic image features that can be applied to a variety of downstream tasks such as retrieval, classification, and segmentation. However, such representations tend to focus on the most salient visual cues in the image, with no way to direct them toward less prominent concepts of interest. In contrast, Multimodal LLMs can be guided with textual prompts, but the resulting representations tend to be language-centric and lose their effectiveness for generic visual tasks. To address this, we introduce Steerable Visual Representations, a new class of visual representations, whose global and local features can be steered with natural language. While most vision-language models (e.g., CLIP) fuse text with visual features after encoding (late fusion), we inject text directly into the layers of the visual encoder (early fusion) via lightweight cross-attention. We introduce benchmarks for measuring representational steerability, and demonstrate that our steerable visual features can focus on any desired objects in an image while preserving the underlying representation quality. Our method also matches or outperforms dedicated approaches on anomaly detection and personalized object discrimination, exhibiting zero-shot generalization to out-of-distribution tasks.

Abstract

Image-based air quality estimation offers a scalable alternative to sparse sensor-based monitoring. However, existing methods rely on narrow field-of-view imagery and overlook both the spatial structure of panoramic scenes and the ordinal nature of Air Quality Index (AQI) prediction. To address these limitations, we introduce PAN-AQI, a large-scale dataset comprising 33,982 panoramic 360◦ street-view images collected over 18 days and 1000 km across the twin cities of Hyderabad and Secunderabad, India, with co-located PM2.5, PM10, Temperature, and Humidity measurements. We further propose PANQIFormer, a multimodal transformer framework that jointly models panoramic visual context and environmental metadata for AQI estimation. The proposed architecture combines spatialzone reasoning, attention-based multimodal fusion, and ordinalaware learning to perform both AQI category classification and continuous AQI regression. Experiments demonstrate consistent improvements over prior image-based AQI methods, reducing AQI MAE from 21.59 to 10.45 while improving classification performance. Together, PAN-AQI and PANQIFormer establish panoramic sensing as a promising direction for scalable visionbased air-quality monitoring. We will release code, dataset, and experimental results at Link.

Abstract

Audio Descriptions (ADs) narrate visual content for Blind and Low Vision (BLV) audiences during gaps in audiovisual media. There is growing momentum around ADs in movies and TV shows, and with mandates from India's Central Board of Film Certification (CBFC), there is a need to expand ADs beyond English. Yet, there is no work that generates ADs for any Indian language. To address this gap, we present the first systematic study of ADs in Hindi, contributing to aspects such as data, generation, and evaluation. We introduce Andha-Dhun, the first dataset of human-authored Hindi ADs collected from 8 full-length movies. We explore two approaches for generating ADs in Hindi: (i) directly from English dense video descriptions, and (ii) translating English ADs into Hindi. We evaluate these approaches using perplexity and LLM-as-a-judge metrics to assess fluency and quality respectively. We also analyze movies that have both English and Hindi human-authored ADs and find that naïve translation introduces artifacts and narrows diversity compared to original Hindi ADs. Direct machine translation fails to adapt cultural references, while human-translated ADs do better but still fall short. Our findings emphasize that the purpose of Hindi ADs is accessibility for Indian BLV audiences, and that this requires adapting content for the audience more than strict fidelity to the source.

Abstract

Low-dimensional reductions provide a useful framework
for studying high-dimensional dynamics on
complex networks, but most existing approaches are
restricted to pairwise interactions. Here, we develop a
one-dimensional reduction for dynamical systems on
networks with purely higher-order interactions. The
reduction is formulated through an effective higherorder
interaction strength (), associated with the triangular
interactions of the underlying network and
the dynamical system's effective state. We present a
theoretical framework for the dimension-reduction approach
and validate it across three dynamical models
with exclusively higher-order interactions.We find
that the reduction accuracy is mainly determined by
the homogeneity of node states, that is the deviations
in state values become very small. Numerical results
on synthetic and real networks show that the reduced
model captures the effective steady states and transitions of the full system with good accuracy.

Abstract

Vaccination campaigns play a pivotal role in controlling infectious diseases. Their success, however, depends not only on vaccine efficacy and availability but also significantly on public opinion and the willingness of individuals to vaccinate. This paper investigates a coupled opinion-epidemic model on heterogeneous networks, where individual opinions influence vaccination probability, and opinions themselves evolve through a combination of peer interaction and local risk perception derived from observed infection rates. Embedding the coupled dynamics in scale-free networks, particularly Barabási-Albert structures, allows us to examine the role of network heterogeneity beyond homogeneous-mixing assumptions. Using Monte Carlo simulations and a semi-analytical microscopic Markov-chain approach, we derive and numerically validate analytical expressions for the critical infection threshold and stable vaccinated population where risk perception dominated peer influence. Our results show that stronger local risk perception enhances pro-vaccination opinions and suppresses infection, while dominant peer influence can increase long-term infection levels. These findings underscore the importance of accounting for social behavior and network structure when designing effective vaccination and epidemic control strategies.

Abstract

Multilingual scene text recognition in the cropped word-level image setting is difficult to scale, especially for Indic scripts, where larger combined character inventories, diacritics, and fine-grained visual ambiguities increase cross-script confusion under limited real-world data availability. We propose mSTR, a single-head multilingual recognizer that operates in a shared character embedding space. In Stage 1, mCLIP learns this space by aligning synthetic character images with prompted character labels. In Stage 2a, mSequenceSpotter predicts word-level embeddings and decodes them via cosine similarity against a frozen multilingual bank of character embeddings, replacing a separately learned vocabulary-sized character classifier with fixed character prototypes. In Stage 2b, mRefiner improves near-miss predictions using visual and contextual memory. This design keeps inference single-head, avoids explicit script routing, and supports scalable multilingual decoding. We evaluate mSTR in a pooled public-data cropped word-level image setting across six languages spanning six distinct scripts and multiple multilingual and Indic benchmarks. mSTR shows strong multilingual recognition performance, compares favorably to multilingual alternatives, including routed and spotting-derived formulations under recognition-only evaluation, and remains largely non-destructive as the language set is expanded from one to six languages in the evaluated setting. Overall, the results support shared character embeddings as a practical basis for jointly trained, single-head multilingual scene text recognition and provide an initial demonstration of stable scaling across the six evaluated scripts.