arXiv subjects

Anirban Das

Publications and source records attributed to Anirban Das.

At least 37 records · Page 2Linked to original sources

arXiv2025

WorldCuisines: A Massive-Scale Benchmark for Multilingual and Multicultural Visual Question Answering on Global Cuisines

Vision Language Models (VLMs) often struggle with culture-specific knowledge, particularly in languages other than English and in underrepresented cultural contexts. To evaluate their understanding of such knowledge, we introduce WorldCuisines, a massive-scale benchmark for multilingual and multicultural, visually grounded language understanding. This benchmark includes a visual question answering (VQA) dataset with text-image pairs across 30 languages and dialects, spanning 9 language families and featuring over 1 million data points, making it the largest multicultural VQA benchmark to date. It includes tasks for identifying dish names and their origins. We provide evaluation datasets in two sizes (12k and 60k instances) alongside a training dataset (1 million instances). Our findings show that while VLMs perform better with correct location context, they struggle with adversarial contexts and predicting specific regional cuisines and languages. To support future research, we release a knowledge base with annotated food entries and images along with the VQA data.

cs.CL↗

arXiv2025

Shedding Infrared Light on QCD Axion and ALP Dark Matter with JWST

James Webb Space Telescope (JWST) has opened up a new chapter in infrared astronomy. Besides the discovery and a deeper understanding of various astrophysical sources, JWST can also uncover the non-gravitational nature of dark matter (DM). If DM is QCD axion or an eV-scale Axion-like particle (ALP), it can decay into two photons in the infrared band. This will produce a distinct line signature in the spectroscopic observations made by JWST. Using the latest NIRSpec IFU spectroscopic observations from JWST, we put the strongest bound on the photon coupling for QCD axion/ ALP DM in the mass range between 0.47 and 2.55 eV. In particular, we are able to probe a new mass range for ALP DM between $\sim$ 0.47 eV to 0.78 eV beyond what can be probed by globular cluster observations. We constrain well-motivated and UV complete models of QCD axion and ALP DM, including predictions from some models derived from string theory and/ or various Grand Unification scenarios. Future JWST observations of DM-rich systems with a better understanding of the astrophysical and instrumental backgrounds can thus enable us to potentially discover QCD axion and ALP DM. The datasets used in this work are available at: https://dx.doi.org/10.17909/3e5f-nv69

hep-ph↗

arXiv2025

RainbowPO: A Unified Framework for Combining Improvements in Preference Optimization

Recently, numerous preference optimization algorithms have been introduced as extensions to the Direct Preference Optimization (DPO) family. While these methods have successfully aligned models with human preferences, there is a lack of understanding regarding the contributions of their additional components. Moreover, fair and consistent comparisons are scarce, making it difficult to discern which components genuinely enhance downstream performance. In this work, we propose RainbowPO, a unified framework that demystifies the effectiveness of existing DPO methods by categorizing their key components into seven broad directions. We integrate these components into a single cohesive objective, enhancing the performance of each individual element. Through extensive experiments, we demonstrate that RainbowPO outperforms existing DPO variants. Additionally, we provide insights to guide researchers in developing new DPO methods and assist practitioners in their implementations.

cs.AI↗

arXiv2025

Cosmological constraints on mass-varying dark matter

As one of the fundamental unknowns of our Universe, the mass of dark matter remains to be a topic of great interest. We consider the possibility of a time-variation of the dark matter mass. We study the cosmological constraints on a model where the dark matter mass transitions from zero to a finite value in the early Universe. In this model, the matter power spectrum exhibits power suppression below a certain scale that depends on the epoch of transition, and the angular power spectrum of the cosmic microwave background shows a distinctive phase shift and power suppression at small scales. We use the latest cosmic microwave background data and the $S_8$ priors from weak lensing data to place a lower limit on the transition redshift. We also find that the data from the ACT show a mild preference for the mass-varying dark matter model over $Λ$CDM.

astro-ph.CO↗

arXiv2024

Preference Tuning with Human Feedback on Language, Speech, and Vision Tasks: A Survey

Preference tuning is a crucial process for aligning deep generative models with human preferences. This survey offers a thorough overview of recent advancements in preference tuning and the integration of human feedback. The paper is organized into three main sections: 1) introduction and preliminaries: an introduction to reinforcement learning frameworks, preference tuning tasks, models, and datasets across various modalities: language, speech, and vision, as well as different policy approaches, 2) in-depth exploration of each preference tuning approach: a detailed analysis of the methods used in preference tuning, and 3) applications, discussion, and future directions: an exploration of the applications of preference tuning in downstream tasks, including evaluation methods for different modalities, and an outlook on future research directions. Our objective is to present the latest methodologies in preference tuning and model alignment, enhancing the understanding of this field for researchers and practitioners. We hope to encourage further engagement and innovation in this area.

cs.CL↗

arXiv2024

Dynamical freezing in the thermodynamic limit: the strongly driven ensemble

The ergodicity postulate, a foundational pillar of Gibbsian statistical mechanics predicts that a periodically driven (Floquet) system in the absence of any conservation law heats to a featureless `infinite temperature' state. Here, we find--for a clean and interacting generic spin chain subject to a {\it strong} driving field--that this can be prevented by the emergence of {\it approximate but stable} conservation-laws not present in the undriven system. We identify their origin: they do not necessarily owe their stability to familiar protections by symmetry, topology, disorder, or even high energy costs. We show numerically, {\it in the thermodynamic limit,} that when required by these emergent conservation-laws, the entanglement-entropy density of an infinite subsystem remains zero over our entire simulation time of several decades in natural units. We further provide a recipe for designing such conservation laws with high accuracy. Finally, we present an ensemble description, which we call the strongly driven ensemble incorporating these constraints. This provides a way to control many-body chaos through stable Floquet-engineering. Strong signatures of these conservation-laws should be experimentally accessible since they manifest in all length and time scales. Variants of the spin model we have used, have already been realized using Rydberg-dressed atoms.

cond-mat.stat-mech↗

arXiv2024

LLM Surgery: Efficient Knowledge Unlearning and Editing in Large Language Models

Large language models (LLMs) have revolutionized various domains, yet their utility comes with significant challenges related to outdated or problematic knowledge embedded during pretraining. This paper addresses the challenge of modifying LLMs to unlearn problematic and outdated information while efficiently integrating new knowledge without retraining from scratch. Here, we propose LLM Surgery, a framework to efficiently modify LLM behaviour by optimizing a three component objective function that: (1) Performs reverse gradient on unlearning dataset (problematic and outdated information), (2) Performs gradient descent on the update dataset (new and updated information), and (3) Minimizes the KL divergence on the retain dataset (small subset of unchanged text), ensuring alignment between pretrained and modified model outputs. Due to the lack of publicly available datasets specifically tailored for our novel task, we compiled a new dataset and an evaluation benchmark. Using Llama2-7B, we demonstrate that LLM Surgery can achieve significant forgetting on the unlearn set, a 20\% increase in accuracy on the update set, and maintain performance on the retain set.

cs.CL↗

arXiv2024

Sub-MeV Dark Matter Detection with Bilayer Graphene

The light dark matter mass regime has emerged as the next frontier in the direct detection experiment due to the lack of any detection signal in the higher mass range. In this paper, we propose a new detector material, a bilayer stack of graphene to detect sub-MeV dark matter. Its voltage-tunable low energy sub-eV electronic band gap makes it an excellent choice for the detector material of a light dark matter search experiment. We compute its dielectric function using the random phase approximation and estimate the projected sensitivity for sub-MeV dark matter-electron scattering and sub-eV dark matter absorption. We show that a bilayer graphene dark matter detector can have competitive sensitivity as other candidate target materials, like a superconductor, but with a tunable threshold energy in this mass regime. The dark matter scattering rate in bilayer graphene is also characterized by a daily modulation from the rotation of the Earth which may help us mitigate the backgrounds in a future experiment. We also outline a detector design concept and provide noise estimates that can be followed to setup an experiment in future.

hep-ph↗

arXiv2024

Transmon Qubit Constraints on Dark Matter-Nucleon Scattering

We recently pointed out that power measurements of single quasiparticle devices can be used to detect dark matter. These devices have the lowest known energy thresholds, far surpassing standard direct detection experiments, requiring energy deposition above only about an meV. We calculate dark matter induced quasiparticle densities in transmon qubits, and use the latest transmon qubit measurements that provide one of the strongest existing lab-based bounds on dark matter-nucleon scattering below about 100 MeV. We strongly constrain sub-component dark matter, using both a dark matter population thermalized in the Earth as well as the dark matter wind from the Galactic halo. We demonstrate future potential sensitivities using devices with low quasiparticle densities.

hep-ph↗

arXiv2024

A Dataset for Multi-intensity Continuous Human Activity Recognition through Passive Sensing

Human activity recognition (HAR) is essential in healthcare, elder care, security, and human-computer interaction. The use of precise sensor data to identify activities passively and continuously makes HAR accessible and ubiquitous. Specifically, millimeter wave (mmWave) radar is promising for passive and continuous HAR due to its ability to penetrate non-metallic materials and provide high-resolution wireless sensing. Although mmWave sensors are effective at capturing macro-scale activities, like exercising, they fail to capture micro-scale activities, such as typing. In this paper, we introduce mmDoppler, a novel dataset that utilizes off-the-shelf (COTS) mmWave radar in order to capture both macro and micro-scale human movements using a machine-learning driven signal processing pipeline. The dataset includes seven subjects performing 19 distinct activities and employs adaptive doppler resolution to enhance activity recognition. By adjusting the radar's doppler resolution based on the activity type, our system captures subtle movements more precisely. mmDoppler includes range-doppler heatmaps, offering detailed motion dynamics, with data collected in a controlled environment with single as well as multiple subjects performing activities simultaneously. The dataset aims to bridge the gap in HAR systems by providing a more comprehensive and detailed resource for improving the robustness and accuracy of mmWave radar activity recognition.

cs.HC↗

arXiv2024

Cross-Silo Federated Learning for Multi-Tier Networks with Vertical and Horizontal Data Partitioning

We consider federated learning in tiered communication networks. Our network model consists of a set of silos, each holding a vertical partition of the data. Each silo contains a hub and a set of clients, with the silo's vertical data shard partitioned horizontally across its clients. We propose Tiered Decentralized Coordinate Descent (TDCD), a communication-efficient decentralized training algorithm for such two-tiered networks. The clients in each silo perform multiple local gradient steps before sharing updates with their hub to reduce communication overhead. Each hub adjusts its coordinates by averaging its workers' updates, and then hubs exchange intermediate updates with one another. We present a theoretical analysis of our algorithm and show the dependence of the convergence rate on the number of vertical partitions and the number of local updates. We further validate our approach empirically via simulation-based experiments using a variety of datasets and objectives.

cs.LG↗

arXiv2024

Dark Matter Induced Power in Quantum Devices

We point out that power measurements of single quasiparticle devices open a new avenue to detect dark matter (DM). The threshold of these devices is set by the Cooper pair binding energy, and is therefore so low that they can detect DM as light as about an MeV incoming from the Galactic halo, as well as the low-velocity thermalized DM component potentially present in the Earth. Using existing power measurements with these new devices, as well as power measurements with SuperCDMS-CPD, we set new constraints on the spin-independent DM scattering cross section for DM masses from about 10 MeV to 10 GeV. We outline future directions to improve sensitivity to both halo DM and a thermalized DM population in the Earth using power deposition in quantum devices.

hep-ph↗

arXiv2024

Energy-dependent Boosted Dark Matter from Diffuse Supernova Neutrino Background

Diffuse neutrinos from past supernovae in the Universe present us with a unique opportunity to test dark matter (DM) interactions. These neutrinos can scatter and boost the DM particles in the Milky Way halo to relativistic energies allowing us to detect them in terrestrial laboratories. Focusing on generic models of DM-neutrino and electron interactions, mediated by a vector or a scalar boson, we implement energy-dependent scattering cross-sections and perform detailed numerical analysis of DM attenuation due to electron scattering in-medium while propagating towards terrestrial experiments. We set new limits on DM-neutrino and electron interactions for DM with masses in the range $\sim (0.1, 10^4)~$MeV, using recent data from XENONnT, LUX-ZEPLIN, and PandaX-4T direct detection experiments. We demonstrate that consideration of energy-dependent cross-sections for DM interactions can significantly affect constraints previously derived under the assumption of constant cross-sections, modifying them by multiple orders of magnitude.

hep-ph↗

arXiv2024

Mass varying dark matter and its cosmological signature

Nontrivial dark sector physics continues to be an interesting avenue in our quest to the nature of dark matter. In this paper, we study the cosmological signatures of mass-varying dark matter where its mass changes from zero to a nonzero value in the early Universe. We compute the changes in various observables, such as, the linear matter power spectrum and the cosmic microwave background anisotropy power spectrum. We explain the origin of the effects and point out a qualitative similarity between this model and a warm dark matter cosmology with no sudden mass transition. Finally, we do a simple analytical study to estimate the constraint on the parameters of this model from the Lyman-$α$ forest data.

astro-ph.CO↗

arXiv2023

Detecting Dark Matter Induced Power in Quantum Devices

In the past few years, many mesoscale systems have been proposed as possible detectors of sub-GeV dark matter particles. In this work, we point out the feasibility of probing dark matter-nucleon scattering cross section using superconductor-based quantum devices with meV-scale energy threshold. We compute new limits on spin-independent dark matter scattering cross section using the existing power measurement data from three different experiments for MeV to 10 GeV mass. We derive the limits for both halo and thermalized dark matter populations.

hep-ph↗

arXiv2023

A gate-tunable quantum phase transition in a topological excitonic insulator

Coulomb interactions among electrons and holes in two-dimensional (2D) semimetals with overlapping valence and conduction bands can give rise to a correlated insulating ground state via exciton formation and condensation. One candidate material in which such excitonic state uniquely combines with non-trivial band topology are atomic monolayers of tungsten ditelluride (WTe2), in which a 2D topological excitonic insulator (2D TEI) forms. However, the detailed mechanism of the 2D bulk gap formation in WTe2, in particular with regard to the role of Coulomb interactions, has remained a subject of ongoing debate. Here, we show that WTe2 is susceptible to a gate-tunable quantum phase transition, evident from an abrupt collapse of its 2D bulk energy gap upon ambipolar field-effect doping. Such gate tunability of a 2D TEI, into either n- and p-type semimetals, promises novel handles of control over non-trivial 2D superconductivity with excitonic pairing.

cond-mat.str-el↗

arXiv2023

The magnificent ACT of flavor-specific neutrino self-interaction

We revisit the cosmology of neutrino self-interaction and use the latest cosmic microwave background data from the Atacama Cosmology Telescope (ACT) and the Planck experiment to constrain the interaction strength. In both flavor-universal and nonuniversal coupling scenarios, we find that the ACT data prefers strong neutrino self-interaction that delays neutrino free streaming until just before the matter-radiation equality. When combined with the Planck 2018 data, the preference for strong interaction decreases due to the Planck polarization data. For the combined dataset, the flavor-specific interaction still provides a better fit to the CMB data than $Λ$CDM. This trend persists even when neutrino mass is taken into account and extra radiation is added. We also study the prospect of constraining such strong interaction by future terrestrial and space telescopes, and find that the upcoming CMB-S4 experiment will improve the upper limit on neutrino self-interaction by about a factor of three.

astro-ph.CO↗

arXiv2023

Continuous Multi-user Activity Tracking via Room-Scale mmWave Sensing

Continuous detection of human activities and presence is essential for developing a pervasive interactive smart space. Existing literature lacks robust wireless sensing mechanisms capable of continuously monitoring multiple users' activities without prior knowledge of the environment. Developing such a mechanism requires simultaneous localization and tracking of multiple subjects. In addition, it requires identifying their activities at various scales, some being macro-scale activities like walking, squats, etc., while others are micro-scale activities like typing or sitting, etc. In this paper, we develop a holistic system called MARS using a single Commercial off the-shelf (COTS) Millimeter Wave (mmWave) radar, which employs an intelligent model to sense both macro and micro activities. In addition, it uses a dynamic spatial time sharing approach to sense different subjects simultaneously. A thorough evaluation of MARS shows that it can infer activities continuously with a weighted F1-Score of > 94% and an average response time of approx 2 sec, with 5 subjects and 19 different activities.

cs.HC↗