arXiv subjects

Subhadeep Roy

Publications and source records attributed to Subhadeep Roy.

At least 19 recordsLinked to original sources

arXiv2026

Who Annotates in NLP? A Large-scale Assessment of Human Annotation Reporting between 2018 and 2025

Human annotation is the empirical foundation of much NLP research, from dataset construction to model evaluation, but papers often leave unclear who produced the annotations and how the annotation process was controlled. We provide the first large-scale, task-level audit of human annotation reporting across major NLP venues, asking which annotation details are documented, which are missing, and how reporting varies across time, topic, venue, and intended use of human judgment. We introduce a unified taxonomy of annotation-reporting practices and validate an LLM-assisted extraction pipeline against Annotated-gold, a human-adjudicated gold standard of 41 papers and 72 annotation tasks, where the best model reaches human-comparable agreement with adjudicated labels, with Krippendorff's alpha of 0.606 versus 0.585 for human-human agreement. Using this pipeline, we construct Annotated-llm, a dataset covering ACL-venue papers from 2018-2025, with 2,667 extracted annotation tasks from 1,603 papers, and find that papers frequently report operational details such as recruitment strategies, annotator expertise, and annotation volume, but often omit details needed to assess annotation validity, including training, language proficiency, compensation, socio-demographics, adjudication, and agreement values, especially in model-evaluation studies. Our results show that annotation reporting in NLP has improved over time but remains uneven, and they establish a scalable framework and bare-minimum reporting recommendations for making human annotation more reliable, reproducible, and interpretable.

cs.CL↗

arXiv2026

Efficient fluid extraction through hydraulic fracture in capillary fiber bundle model

We have simulated a one dimensional capillary fiber bundle model with fracking events while acted between a pressure gradient across the system. The hydraulic fractures are incorporated through a decreasing nature of capillary thresholds for each tube that replicates an increment in pore spaces due to fracking. An increment in flow rate is evident through the evolved rheology we observe in our study. Analytical approaches for certain limits are adopted to understand the rheology which matches well with the numerical results. The overall hydraulic power increases with pressure gradient as well as with the percentage decrease in capillary threshold due to a single event, defines as the fracking amplitude. This combined with the early onset of linear Darcy flow increases the quality of the fluid extraction. We successfully point towards an optimum pressure gradient at which the fracking events are most effective - maximum change in fluid extracting with a maximum rate. We observed that it is possible to extract the information regarding the change from non-linear to Darcy flow due to fracking as well as the optimum pressure for fluid extraction through local flow profile, something which in much superior from the point of view of computational cost. The former is done by correlating the maximum fluctuation in local flow profile to the onset of Darcy flow. The later is done through the relative change in Shannon entropy with respect to the fracking amplitude that points towards the pressure associated with the maximum fluid extraction criterion.

physics.flu-dyn↗

arXiv2026

Relative transverse activity as a probe of collectivity-like long-range correlations in pp collisions at $\sqrt{s}=13$ TeV

Understanding the origin of collectivity-like signatures in small collision systems is a central open question in high-energy nuclear physics, and two-particle correlation functions offer unique sensitivity to the underlying-event (UE) dynamics that may drive such behavior in proton--proton (pp) collisions. In this work, the two-particle number ($R_{2}$) and transverse-momentum ($P_{2}$) correlation functions are studied in pp collisions at $\sqrt{s}=13$ TeV using PYTHIA 8, for final state charged hadrons within $|\eta|<0.8$ and $0.2<p_{\rm T}<2.0$ GeV/$c$, with events classified by the relative transverse activity $R_{\mathrm{T}}$ to probe how UE activity shapes correlation structures in the soft-QCD-dominated regime. A collectivity-like long-range near-side component is observed in the charge-independent correlator $R_{2}^{\mathrm{CI}}$ exclusively for the highest $R_{\mathrm{T}}$ class ($2.5 < R_{\mathrm{T}} \leq 5.0$), while no corresponding structure appears in the charge-dependent correlators. This indicates that enhanced UE activity, driven by multiple partonic interactions and color reconnection, can generate collectivity-like long-range correlations without hydrodynamic evolution. These findings establish $R_{\mathrm{T}}$ as a differential event classifier to provide a non-hydrodynamic baseline for interpreting such signatures in small-system measurements at the LHC.

hep-ph↗

arXiv2026

Prototypicality Bias Reveals Blindspots in Multimodal Evaluation Metrics

Automatic metrics are widely used to evaluate text-to-image models, often replacing human judgment in benchmarking, model selection, and large-scale data filtering. Yet they may reward images that look plausible or prototypical rather than images that faithfully satisfy the prompt. We identify prototypicality bias as a systematic blindspot in multimodal evaluation: metrics can prefer a semantically incorrect but visually or socially prototypical image over a correct but less prototypical one. We introduce PROTOBIAS, a controlled diagnostic benchmark across Animals, Objects, and Demography, where semantically correct images are contrasted with plausible prototypical adversaries containing a single controlled semantic violation. Grounded in prototype theory and social-category prototypicality, PROTOBIAS is constructed with multiple prompt generators, image generators, and independent VLM filters, and validated through prompt-quality, human-annotation, and image-quality controls. Using PROTOBIAS, we show that widely used embedding, reward, VQA-based, and VLM-as-judge metrics frequently fail these contrasts, while human judgments remain more faithful to semantic correctness. We further introduce PROTOSCORE, a lightweight contrastively trained evaluator, as an initial mitigation baseline. PROTOBIAS provides a focused benchmark for measuring prototypicality-driven metric failures and developing more semantically faithful T2I evaluators.

cs.CV↗

arXiv2025

Event-by-event fluctuations of mean transverse momentum in proton-proton collisions at $\sqrt{s}$ = 13 TeV with PYTHIA8 and HERWIG7 models

Estimations of event-by-event mean transverse momentum ($\langle p_{\rm T} \rangle$) fluctuations are reported in terms of the integral correlator, $\langle \Delta p_{\rm T} \Delta p_{\rm T}\rangle$, and the skewness of event-wise $\langle p_{\rm T} \rangle$ distribution in proton$-$proton (pp) collisions at $\sqrt{s}=13$ TeV with the Monte Carlo event generators PYTHIA8 and HERWIG7. The final-state charged particles with transverse momentum ($p_{\rm T}$) and pseudorapidity ($\eta$) ranges $0.15 \leq p_{\rm T}\leq 2.0$ GeV/$c$ and $|\eta| \leq 0.8$ were considered for the investigation. The correlator, $\langle \Delta p_{\rm T} \Delta p_{\rm T}\rangle$, is observed to follow distinct decreasing trends with average charged particle multiplicity ($\langle N_{\rm ch} \rangle$) for the models. Furthermore, both models yield positive finite skewness in low-multiplicity events. Fluctuations are additionally studied using the transverse spherocity estimator ($S_{\rm 0}$) to understand the relative contributions of hard scattering (jets) and other soft processes to the observed fluctuations. Comparing model predictions for $\langle p_{\rm T} \rangle$ fluctuations provides valuable insight into the sensitivity of these fluctuations to hadronization and parton shower models. This is essential for a reliable interpretation of the fluctuation dynamics in pp collisions. Moreover, such comparisons would help to establish a crucial baseline for identifying and studying non-trivial fluctuations in heavy-ion collisions.

hep-ph↗

arXiv2024

Inequality indices for heterogeneous systems: a tool for failure prediction

We have numerically studied a mean-field fiber bundle model of fracture at a non-zero temperature and acted by a constant external tensile stress. The individual fibers fail (local damage) due to creep-like dynamics that lead up to a catastrophic breakdown (global failure). We quantify the variations in sizes of the resulting avalanches by calculating the Lorenz function and two inequality indices -- Gini ($g$) and Kolkata ($k$) indices -- derived from the Lorenz function. We show that the two indices cross just prior to the failure point when the dynamics goes through intermittent avalanches. For a continuous failure dynamics (finite numbers of fibers breaking at each time step), the crossing does not happen. However, in that phase, the usual prediction method i.e., linear relation between the time of minimum strain-rate and failure time, holds. The boundary between continuous and intermittent dynamics is very close to the boundary between crossing and non-crossing of the two indices in the temperature-stress phase space.

cond-mat.stat-mech↗

arXiv2024

Critical crack-length during fracture

Through controlled numerical simulations in a one dimensional fiber bundle model with local stress concentration, we established an inverse correlation between the strength of the material and the cracks which grow inside it - both the maximum crack and the one that set in instability within the system, defined to be the critical crack. Through Pearson correlation function as well as probabilistic study of individual configurations, we found that the maximum and the critical crack often differ from each other unless the disorder strength is extremely low. A phase diagram on the plane of disorder vs system size demarcates between the regions where the largest crack is the most vulnerable one and where they differ from each other but still shows moderate correlation.

cond-mat.stat-mech↗

arXiv2023

Disorder-induced non-linear growth of viscously-unstable immiscible two-phase flow fingers in porous media

The immiscible displacement of a fluid by another one inside a porous medium produces different types of patterns depending on the capillary number Ca and viscosity ratio M. At high Ca, viscous fingers resulting from the viscous instability between fluid-fluid interfaces are believed to exhibit the same Laplacian growth behavior as viscously-unstable fingers observed in Hele-Shaw cells by Saffman and Taylor [1], or as diffusion limited aggregates (DLA) [2]. I.e., the interface velocity depends linearly on the local gradient of the physical field that drives the growth process (for two-phase flow, the pressure field). However, steady-state two-phase flow in porous media is known to exhibit a regime for which the flow rate depends as a non-linear power law on the global pressure drop, due to the disorder in the capillary barriers at pore throats. A similar nonlinear growth regime was also evidenced experimentally for viscously-unstable drainage in two-dimensional porous media 20 years ago [3]. Here we revisit this flow regime using dynamic pore-network modeling, and explore the non-linearity in the growth properties. We characterize the previously-unstudied dependencies of the statistical finger width and nonlinear growth law's exponent on Ca, and discuss quantitatively, based on theoretical arguments, how disorder in the capillary barriers controls the growth process' non-linearity, and why the flow regime crosses over to Laplacian growth at sufficiently high Ca. In addition, the statistical properties of the fingering patterns are compared to those of Saffman-Taylor fingers, DLA growth patterns, and the results from the aforementioned previous experimental study.

physics.flu-dyn↗

arXiv2023

Record statistics based prediction of fracture in the random spring network model

We study the role of record statistics of damage avalanches in predicting the fracture of a heterogeneous material under tensile loading. The material is modeled using a two-dimensional random spring network where disorder is introduced through randomness in the breakage threshold strains of the springs. It is shown that the waiting time between successive records of avalanches has a maximum for moderate disorder, thus showing an acceleration of records with impending fracture. Such a signature is absent for low disorder strength when the fracture is nucleation-dominated, and high disorder strength when the fracture is percolation type. We examine the correlation between the record with the maximum waiting time and the crossover record at which the avalanche statistics change from off-critical to critical. Compared to the avalanche based predictor for failure, we show that the record statistics have the advantage of both being real-time as well as able to predict final fracture at much smaller strains. We also show that in the avalanche-dominated regime, the failure strain is shown to have a linear relation with the strain at the maximum waiting time, making possible a quantitative prediction.

cond-mat.stat-mech↗

arXiv2023

Test Time Adaptation for Blind Image Quality Assessment

While the design of blind image quality assessment (IQA) algorithms has improved significantly, the distribution shift between the training and testing scenarios often leads to a poor performance of these methods at inference time. This motivates the study of test time adaptation (TTA) techniques to improve their performance at inference time. Existing auxiliary tasks and loss functions used for TTA may not be relevant for quality-aware adaptation of the pre-trained model. In this work, we introduce two novel quality-relevant auxiliary tasks at the batch and sample levels to enable TTA for blind IQA. In particular, we introduce a group contrastive loss at the batch level and a relative rank loss at the sample level to make the model quality aware and adapt to the target data. Our experiments reveal that even using a small batch of images from the test distribution helps achieve significant improvement in performance by updating the batch normalization statistics of the source model.

cs.CV↗

arXiv2023

Multiplicity and Transverse Spherocity dependence of $\langle p_{\rm T} \rangle$ fluctuations of charged particles in p$-$p collisions at $\sqrt{s}$ = 7 and 13 TeV

The multiplicity dependence of event-by-event fluctuations in mean transverse momentum, $\langle p_{\rm T} \rangle$, of charged particles has been studied in p$-$p collisions at $\sqrt{s}$ = 7 TeV and 13 TeV using the PYTHIA 8 event generator. The charged particles were selected in kinematic range of $0.15 < p_{\rm T}<2$ GeV$/c$ and $|\eta| < 0.8$. The dynamical fluctuations would indicate towards the correlated emission of particles. The measurements in A$-$A and p$-$p collisions has shown a decrease in the strength of $ \langle p_{\rm T} \rangle$ fluctuations with the average charged particle multiplicity. The effects of various microscopic processes like color reconnection and multi-partonic interactions has been studied. A minimal dependency on the collision energy is also observed. Furthermore, the fluctuation observables are investigated in the intervals of transverse spherocity in order to comprehend the relative contributions resulting from hard scattering and underlying events. The present study would act as a baseline for future measurements in A$-$A as well as p$-$p collisions at the LHC.

hep-ph↗

arXiv2023

Record statistics of emitted energies -- prediction of an upcoming failure

The article reports a numerical investigation of the breakdown of a disordered system considering the effect of local stress concentration under the action of an external tensile force. The statistics of the record-breaking magnitudes of emitted energies during the failure process, as well as the waiting time to achieve those record events, show rich behavior. The latter includes information about the acceleration and subsequent catastrophic failure through its non-monotonic behavior. The maximum waiting time is also correlated with the maximum change in elastic energy as the model evolves, a different way of predicting an upcoming failure, which is consistent with our hypothesis as well. At a moderate disorder, such a prediction can be done with higher accuracy while at a low disorder, due to the abrupt nature of the failure process our hypothesis does not hold well.

cond-mat.stat-mech↗

arXiv2022

Modeling crack propagation in heterogeneous materials: Griffith's law, intrinsic crack resistance and avalanches

Various kinds of heterogeneity in solids including atomistic discreteness affect the fracture strength as well as the failure dynamics remarkably. Here we study the effects of an initial crack in a discrete model for fracture in heterogeneous materials, known as the fiber bundle model. We find three distinct regimes for fracture dynamics depending on the initial crack size. If the initial crack is smaller than a certain value, it does not affect the rupture dynamics and the critical stress. While for a larger initial crack, the growth of the crack leads to a breakdown of the entire system, and the critical stress depends on the crack size in a power-law manner with a nontrivial exponent. The exponent, as well as the limiting crack size, depend on the strength of heterogeneity and the range of stress relaxation in the system.

cond-mat.stat-mech↗

arXiv2021

Opinion dynamics: Public and private

We study here the dynamics of opinion formation in a society where we take into account of the internally held beliefs and externally expressed opinions of the individuals, which are not necessarily the same at all times. While these two components can influence one another, their difference, both in dynamics and in the steady state, poses interesting scenarios in terms of the transition to consensus in the society and characterizations of such consensus. Here we study this public and private opinion dynamics and the critical behavior of the consensus forming transitions, using a kinetic exchange model.

physics.soc-ph↗

arXiv2021

The Co-Moving Velocity in Immiscible Two-Phase Flow in Porous Media

We present a continuum (i.e., an effective) description of immiscible two-phase flow in porous media characterized by two fields, the pressure and the saturation. Gradients in these two fields are the driving forces that move the immiscible fluids around. The fluids are characterized by two seepage velocity fields, one for each fluid. Following Hansen et al.\ (Transport in Porous Media, 125, 565 (2018)), we construct a two-way transformation between the velocity couple consisting of the seepage velocity of each fluid, to a velocity couple consisting of the average seepage velocity of both fluids and a new velocity parameter, the co-moving velocity. The co-moving velocity is related but not equal to velocity difference between the two immiscible fluids. The two-way mapping, the mass conservation equation and the constitutive equations for the average seepage velocity and the co-moving velocity form a closed set of equations that determine the flow. There is growing experimental, computational and theoretical evidence that constitutive equation for the average seepage velocity has the form of a power law in the pressure gradient over a wide range of capillary numbers. Through the transformation between the two velocity couples, this constitutive equation may be taken directly into account in the equations describing the flow of each fluid. This is e.g., not possible using relative permeability theory. By reverse engineering relative permeability data from the literature, we construct the constitutive equation for the co-moving velocity. We also calculate the co-moving constitutive equation using a dynamic pore network model over a wide range of parameters, from where the flow is viscosity dominated to where the capillary and viscous forces compete.

physics.flu-dyn↗

arXiv2021

Correlation between avalanches and emitted energies during fracture with variable stress release range

We observe the failure process of a fiber bundle model with a variable stress release range, $\gamma$, higher the value of $\gamma$ lower the stress release range. By tuning $\gamma$ from low to high, it is possible to go from the mean-field (MF) limit of the model to local load sharing (LLS) where local stress concentration plays a crucial role. In the MF limit, the avalanche size $s$ and energy $E$ emitted during the avalanche are highly correlated producing the same distribution for both $P(s)$ and $Q(E)$: a scale-free distribution with a universal exponent -5/2. With increasing $\gamma$, the model enters the LLS limit. In this limit, due to the presence of local stress concentration such correlation $C(\gamma)$ between $s$ and $E$ decreases where the nature of the decreases depends highly on the dimension of the bundle. In 1d, the $C(\gamma)$ stars from a high value for low $\gamma$ and decreases towards zero when $\gamma$ is increased. As a result, $Q(E)$ and $P(s)$ are similar at low $\gamma$, an exponential one, and then $Q(E)$ becomes power-law for high-stress release range though $P(s)$ remains exponential. On the other hand, in 2d, the $C(\gamma)$ decreases slightly with $\gamma$ but remains at a high value. Due to such a high correlation, the distribution of both $s$ and $E$ is exponential in the LLS limit independent of how large $\gamma$ is.

cond-mat.stat-mech↗

arXiv2021

From nucleation to percolation: the effect of system size and system disorder

A phase diagram for a one dimensional fiber bundle model is constructed with a continuous variation in two parameters guiding dynamics of the model: strength of disorder and system size. We monitor the successive events of fiber rupture in order to understand the spatial correlation associated with it. We observe three distinct regions with increasing disorder strength. (I) Nucleation - a crack propagates from a particular nucleus with very high spatial correlation and causes global failure; (II) Avalanche - the rupture events show precursors activities with a number of bursts. (III) Percolation - the rupture events are spatially uncorrelated like a percolation process. As the size of the bundle is increased, it favors the nucleating failure. In the thermodynamic limit, we only observe a nucleating failure unless the disorder strength is infinitely high.

cond-mat.stat-mech↗

arXiv2021

Rheology of immiscible two-phase flow in mixed wet porous media: Dynamic pore network model and capillary fiber bundle model results

Immiscible two-phase flow in porous media with mixed wet conditions was examined using a capillary fiber bundle model, which is analytically solvable, and a dynamic pore network model. The mixed wettability was implemented in the models by allowing each tube or link to have a different wetting angle chosen randomly from a given distribution. Both models showed that mixed wettability can have significant influence on the rheology in terms of the dependence of the global volumetric flow rate on the global pressure drop. In the capillary fiber bundle model, for small pressure drops when only a small fraction of the tubes were open, it was found that the volumetric flow rate depended on the excess pressure drop as a power law with an exponent equal to 3/2 or 2 depending on the minimum pressure drop necessary for flow. When all the tubes were open due to a high pressure drop, the volumetric flow rate depended linearly on the pressure drop, independent of the wettability. In the transition region in between where most of the tubes opened, the volumetric flow depended more sensitively on the wetting angle distribution function and was in general not a simple power law. The dynamic pore network model results also showed a linear dependence of the flow rate on the pressure drop when the pressure drop is large. However, out of this limit the dynamic pore network model demonstrated a more complicated behaviour that depended on the mixed wettability condition and the saturation. In particular, the exponent relating volumetric flow rate to the excess pressure drop could take on values anywhere between 1.0 and 1.8. The values of the exponent were highest for saturations approaching 0.5, also, the exponent generally increased when the difference in wettability of the two fluids were larger and when this difference was present for a larger fraction of the porous network.

physics.flu-dyn↗