arXiv subjects

Mohammadtaher Safarzadeh

Publications and source records attributed to Mohammadtaher Safarzadeh.

At least 19 recordsLinked to original sources

arXiv2026

When Vision-Language Models Judge Without Seeing: Exposing Informativeness Bias

The reliability of VLM-as-a-Judge is critical for the automatic evaluation of vision-language models (VLMs). Despite recent progress, our analysis reveals that VLM-as-a-Judge often pays limited attention to the image when making decisions. Instead, they often blindly favor the more informative answer, even when they can recognize it conflicts with the image content. We call this problem informativeness bias, which significantly undermines judge reliability. To address it, we propose BIRCH (Balanced Informativeness and CoRrectness with a Truthful AnCHor), a judging paradigm that first corrects inconsistencies with the image content in candidate answers, and then compares the answers against this corrected version. This shifts the judge's focus from informativeness to image-grounded correctness. Experiments on multiple models and benchmarks show that BIRCH reduces informativeness bias by up to 17%, resulting in performance gains of up to 9.8%. Our work reveals an overlooked but fundamental flaw in current VLM-as-a-Judge systems and highlights the need for more principled designs.

cs.AI↗

arXiv2026

SPENCE: A Syntactic Probe for Detecting Contamination in NL2SQL Benchmarks

Large language models (LLMs) have achieved strong performance on natural language to SQL (NL2SQL) benchmarks, yet their reported accuracy may be inflated by contamination from benchmark queries or structurally similar patterns seen during training. We introduce SPENCE (Syntactic Probing and Evaluation of NL2SQL Contamination Effects), a controlled syntactic probing framework for detecting and quantifying such contamination. SPENCE systematically generates syntactic variants of test queries for four widely used NL2SQL datasets-Spider, SParC, CoSQL, and the newer BIRD benchmark. We use SPENCE to evaluate multiple high-capacity LLMs under execution-based scoring. For each model, we measure changes in execution accuracy across increasing levels of syntactic divergence and quantify rank sensitivity using Kendall's tau with bootstrap confidence intervals. By aligning these robustness trends with benchmark release dates, we observe a clear temporal gradient: older benchmarks such as Spider exhibit the strongest negative values and thus the highest likelihood of training leakage, whereas the more recent BIRD dataset shows minimal sensitivity and appears largely uncontaminated. Together, these findings highlight the importance of temporally contextualized, syntactic-probing evaluation for trustworthy NL2SQL benchmarking.

cs.CL↗

arXiv2025

Evaluating NL2SQL via SQL2NL

Robust evaluation in the presence of linguistic variation is key to understanding the generalization capabilities of Natural Language to SQL (NL2SQL) models, yet existing benchmarks rarely address this factor in a systematic or controlled manner. We propose a novel schema-aligned paraphrasing framework that leverages SQL-to-NL (SQL2NL) to automatically generate semantically equivalent, lexically diverse queries while maintaining alignment with the original schema and intent. This enables the first targeted evaluation of NL2SQL robustness to linguistic variation in isolation-distinct from prior work that primarily investigates ambiguity or schema perturbations. Our analysis reveals that state-of-the-art models are far more brittle than standard benchmarks suggest. For example, LLaMa3.3-70B exhibits a 10.23% drop in execution accuracy (from 77.11% to 66.9%) on paraphrased Spider queries, while LLaMa3.1-8B suffers an even larger drop of nearly 20% (from 62.9% to 42.5%). Smaller models (e.g., GPT-4o mini) are disproportionately affected. We also find that robustness degradation varies significantly with query complexity, dataset, and domain -- highlighting the need for evaluation frameworks that explicitly measure linguistic generalization to ensure reliable performance in real-world settings.

cs.CL↗

arXiv2022

A new approach to constrain the Hubble expansion rate at high redshifts by gravitational waves

Detection of massive binary black hole (BBH) mergers at high redshifts is a target for LISA space mission. While the individual masses of a BBH merger are redshifted, the mass ratio of BBH mergers is independent of their redshift. Therefore, if there is an independent correlation between the mass ratio and redshift, such a relationship can be used to i) infer the redshift of the merging binaries, and together with the luminosity distance measurement ($D_L$), constrain the expansion rate of the universe at high redshifts $H(z)$, and ii) constrain models of supermassive black hole seed formation in the universe assuming a fixed cosmology. We discuss why there is an expected relation between the mass ratio of the massive BBHs with their redshift and show the clues for this relation by analyzing cosmological hydrodynamical simulations of BBH mergers. This approach opens up the possibility of directly measuring the expansion rate at redshift $z \approx 2$ with LISA for the first time. Moreover, we discover a trend between seed mass and mass ratio of massive BBHs which by itself is a major result that could be exploited to constrain the formation scenarios of supermassive BH seeds.

astro-ph.CO↗

arXiv2022

A Statistical Detection of Wide Binary Systems in the Ultra-Faint Dwarf Galaxy Reticulum II

Binary stars can inflate the observed velocity dispersion of stars in dark matter dominated systems such as ultra-faint dwarf galaxies (UFDs). However, the population of binaries in UFDs is poorly constrained by observations, with preferred binary fractions for individual galaxies ranging from a few percent to nearly unity. Searching for wide binaries through nearest neighbor (NN) statistics (or the two-point correlation function) has been suggested in the literature, and we apply this method for the first time to detect wide binaries in a UFD. By analyzing the positions of stars in Reticulum II (Ret II) from Hubble Space Telescope images, we search for angularly resolved wide binaries in Ret II. We find that the distribution of their NN distances shows an enhancement at projected separations of $<1$ arcsec relative to a model containing no binaries. We show that such an enhancement can be explained by a wide~binary fraction of $f_b\approx0.007^{+0.008}_{-0.003}$ at separations of more than 3000 AU. Under the assumption that the binary separation distribution is similar to that in the Milky Way, the total binary fraction in Ret II may be on the order of 50%. We also use the observed magnitude distribution of stars in Ret II to constrain the initial mass function over the mass range $0.34-0.78~M_{\odot}$, finding that a shallow power-law slope of $1.01 \le α\le 1.15$ matches the data.

astro-ph.GA↗

arXiv2022

Interpreting a Machine Learning Model for Detecting Gravitational Waves

We describe a case study of translational research, applying interpretability techniques developed for computer vision to machine learning models used to search for and find gravitational waves. The models we study are trained to detect black hole merger events in non-Gaussian and non-stationary advanced Laser Interferometer Gravitational-wave Observatory (LIGO) data. We produced visualizations of the response of machine learning models when they process advanced LIGO data that contains real gravitational wave signals, noise anomalies, and pure advanced LIGO noise. Our findings shed light on the responses of individual neurons in these machine learning models. Further analysis suggests that different parts of the network appear to specialize in local versus global features, and that this difference appears to be rooted in the branched architecture of the network as well as noise characteristics of the LIGO detectors. We believe efforts to whiten these "black box" models can suggest future avenues for research and help inform the design of interpretable machine learning models for gravitational wave astrophysics.

gr-qc↗

arXiv2021

A New Challenge for Dark Matter Models

Cold dark matter (CDM) has faced a number of challenges mainly at small scales, such as the too-big-to-fail problem, and core-cusp density profile of dwarf galaxies. Such problems were argued to have a solution either in the baryonic physics sector or in modifying the nature of dark matter to be self-interacting, or self-annihilating, or ultra-light. Here we present a new challenge for CDM by showing that two of Milky Way's satellites (Horologium I, and Tucana II) are too dense, requiring the formation masses and redshifts of halos in CDM not compatible with being a satellite. These too-dense-to-be-satellite systems are dominated by dark matter and exhibit a surface density above mean dark energy cosmic surface density $\simΩ_Λ ρ_c c/H_0\approx 600~\rm M_{\odot}/pc^2$. This value corresponds to dark matter pressure of $\approx 10^{-9}{\rm erg/cm^3}$. Along with the recently reported excess in small-scale substructures found in cluster-lenses, this problem, unlike other issues facing CDM, has no solution in the baryonic sector and none of the current alternatives of dark matter can account for it. The too-dense-to-be-satellite problem presented in this work provides a new clue for the nature of dark matter, never accounted for before. We note that Horologium I and Tucana II have only been discovered in the past $\sim6$ years with DES, and future surveys (such as those done by Vera C. Rubin Observatory) may uncover a broader population of such galaxies. Moreover, we find that a number of MW's satellite require formation halo masses below the atomic cooling limit which by itself is another challenging observation to account for in CDM.

astro-ph.GA↗

arXiv2021

The Challenge to MOND from ultra faint dwarf galaxies

Modified Newtonian Dynamics (MOND) at low acceleration has been astonishingly powerful at explaining the flat rotation curve of galaxies and the relation between the baryonic content of the galaxies and their observed circular velocity, known as the Baryonic Tully-Fisher Relationship (BTFR). It is known that MOND fails at explaining the observed velocity dispersion of the ultra-faint dwarf galaxies (UFDs) with the justification that UFDs are more prone to tidal disruption in MOND compared to cold dark matter model. We show that: (i) the ratio of tidal to internal acceleration in UFDs is extremely low, (ii) there is no correlation between the deviation of UFDs from MOND's prediction as a function of tidal susceptibility, and (iii) recent constraints from Gaia proper motion analysis on the orbital parameters of the UFDs exacerbates the challenge to MOND. In particular, Gaia data indicates that Ursa Major I is experiencing a recent infall into the Milky Way's halo, and its inconsistency with MOND at 7-$σ$ level can not be attributed to being an early infall satellite. Moreover, the new data from Gaia DR2 shows Willman I to have the least eccentric orbit of all UFDs, and its deviation from MOND at 4-$σ$ level can not be attributed to a highly eccentric orbit as previously suggested. Finally, given that Tucana III is the only UFD observed to show tidal features, Reticulum II and Segue I are two other UFDs that potentially challenge MOND as they have comparable galactocentric distances to Tucana III while showing no tidal features. Whether wide binaries have inflated the velocity dispersion of the UFDs remains an open question to be addressed with future multi-epoch observations.

astro-ph.GA↗

arXiv2021

Explaining the LIGO black hole mass function with field binaries: Revisiting Stellar Evolution at low Metallicity or Invoking Growth via gas accretion?

Our understanding of the formation and evolution of binary black holes (BBHs) is significantly impacted by the recent discoveries made by the LIGO/Virgo collaboration. Of utmost importance is the detection of the most massive BBH system, GW190521. Here we investigate what it takes for field massive stellar binaries to account for the formation of such massive BBHs. Whether the high mass end of the BH mass function is populated by remnants of massive stars that either formed at extremely low metallicities and avoid the pair-instability mass gap or increase their birth mass beyond the pair-instability mass gap through the accretion of gas from the surrounding medium. We show that assuming that massive stars at very low metallicities can form massive BHs by avoiding pair-instability supernova, coupled with a correspondingly high formation efficiency for BBHs, can explain the observed BH mass function. To this end, one requires a relation between the initial and final mass of the progenitor stars at low metallicities that is shallower than what is expected from wind mass loss alone. On the other hand, assuming pair-instability operates at all metallicities, one can account for the observed BH mass function if at least about 10% of the BHs born at very low metallicities double their mass before they merge because of accretion of ambient gas. Such BBHs will have to spend about a Gyr within a parsec length-scale of their parent atomic cooling halos or a shorter timescale if they reside in the inner sub-parsecs of their host dark matter halos. Future stellar evolution calculations of massive stars at very low metallicity and hydrodynamical simulations of gas accretion onto BBHs born in atomic cooling halos can shed light on this debate.

astro-ph.HE↗

arXiv2021

How can LISA probe a population of GW190425-like binary neutron stars in the Milky Way?

The nature of GW190425, a presumed binary neutron star (BNS) merger detected by the LIGO/Virgo Scientific Collaboration (LVC) with a total mass of $3.4^{+0.3}_{-0.1}$ M$_{\odot}$, remains a mystery. With such a large total mass, GW190425 stands at five standard deviations away from the total mass distribution of Galactic BNSs of $2.66\pm 0.12$ M$_{\odot}$. LVC suggested that this system could be a BNS formed from a fast-merging channel rendering its non-detection at radio wavelengths due to selection effects. BNSs with orbital periods less than a few hours - progenitors of LIGO/Virgo mergers - are prime target candidates for the future Laser Interferometer Space Antenna (LISA). If GW190425-like binaries exist in the Milky Way, LISA will detect them within the volume of our Galaxy and will measure their chirp masses to better than 10% for those binaries with gravitational wave frequencies larger than 2 mHz. This work explores how we can probe a population of Galactic GW190425-like BNSs with LISA and investigate their origin. We assume that the Milky Way's BNS population consists of two distinct sub-populations: a fraction $w_1$ that follows the observed Galactic BNS chirp mass distribution and $w_2$ that resembles chirp mass of GW190425. We show that LISA's accuracy on recovering the fraction of GW190425-like binaries depends on the BNS merger rate. For the merger rates reported in the literature, $21 - 212\,$Myr$^{-1}$, the error on the recovered fractions varies between $\sim 30 - 5$%.

astro-ph.HE↗

arXiv2021

A Common Origin for Low Mass Ratio Events Observed by LIGO and Virgo in the First Half of the Third Observing Run

In its third observing run, the LIGO/Virgo collaboration has announced a potential neutron star-black hole (NSBH) merger candidate, GW190426_152155. Together with GW190814, these two events belong to a class of binaries with a secondary mass less than $3 \, M_{\odot}$. While the secondary system in GW190426_152155 is consistent with being a neutron star with a mass of $1.5^{+0.8}_{-0.5} \, M_{\odot}$, that of GW190814 is a $2.59^{+0.08}_{-0.09} \, M_{\odot}$ object and counts as the first confirmed detection of a mass-gap object. Here we argue that these two events could have a common origin as follows: both are formed as NSBH systems; however, the larger escape velocity of a system with more massive primary BH increases the bound fraction of the ejecta material from the supernova explosion leading to the formation of a NS. This bound material forms a disk, which is preferentially accreted onto the NS. This scenario predicts the secondary component mass should correlate with the primary component mass, which is consistent with GW190426_152155 and GW190814. If this hypothesis is corroborated by upcoming observations, GW190814-like events can be excluded from the binary black hole population when inferring their global characteristics.

astro-ph.HE↗

arXiv2020

Confronting the Magnetar Interpretation of Fast Radio Bursts Through Their Host Galaxy Demographics

We explore the prompt magnetar progenitor scenario in the context of fast radio burst (FRB) host galaxies demographics and offset distributions. Magnetars are neutron stars with strong magnetic fields on the order of $10^{15}$ G with a short decay lifetime of less than $10^4$ years. Due to their extremely short lifetimes, magnetars should follow the demographics of galaxies according to their current star-formation rate (SFR). Moreover, we hypothesize that magnetars should follow the SFR profile within galaxies, which we assume to follow an exponential profile. We construct a simple model for the host galaxies of magnetars assuming these events track SFR in all galaxies and compare it to observed properties from a sample of \nsecure\ secure FRB hosts. We find the distribution of observed SFRs is inconsistent with the model at $>95\%$ c.l. The offset distribution is consistent with this scenario; however, this could be due to the limited sample size and the seeing limited estimates for the effective radii of the FRB host galaxies. Despite the recent association of an FRB with a magnetar in the Milky Way, magnetars may not be the only source of FRBs in the universe, yet any other successful model must account for the demographics of the FRB host in SFR and their observed galactocentric offsets.

astro-ph.HE↗

arXiv2020

Radiogenic Heating and its Influence on Rocky Planet Dynamos and Habitability

The thermal evolution of rocky planets on geological timescales (Gyr) depends on the heat input from the long-lived radiogenic elements potassium, thorium, and uranium. Concentrations of the latter two in rocky planet mantles are likely to vary by up to an order of magnitude between different planetary systems because Th and U, like other heavy r-process elements, are produced by rare stellar processes. Here we discuss the effects of these variations on the thermal evolution of an Earth-size planet, using a 1D parameterized convection model. Assuming Th and U abundances consistent with geochemical models of the Bulk Silicate Earth based on chondritic meteorites, we find that Earth had just enough radiogenic heating to maintain a persistent dynamo. According to this model, Earth-like planets of stars with higher abundances of heavy r-process elements, indicated by the relative abundance of europium in their spectra, are likely to have lacked a dynamo for a significant fraction of their lifetimes, with potentially negative consequences for hosting a biosphere. Because the qualitative outcomes of our 1D model are strongly dependent on the treatment of viscosity, further investigations using fully 3D convection models are desirable.

astro-ph.EP↗

arXiv2020

Formation of GW190521 via gas accretion onto Population III stellar black hole remnants born in high-redshift minihalos

The recent gravitational wave merger event, GW190521, has challenged our understanding of the stellar-mass black hole (BH) formation. The primary and secondary BH are both inferred to fall inside the pair-instability (PI) mass gap. Here we propose that the formation of such binaries is possible through gas accretion onto the BH remnants of Population III (Pop~III) stars born in high-redshift ($z>10$) minihalos. Once the parent halo has grown to the atomic-cooling limit, even brief episodes of gas accretion in the dense central regions of the halo can increase the masses of Pop~III remnant BHs above the PI limit. Starting with a BBH with an initial mass of O(100) ${\rm M_{\odot}}$ we find that it would only need to spend about 100~Myr in the inner few pc of an atomic-cooling halo to accrete about 50~${\rm M_{\odot}}$ of material and resemble a system similar to GW190521. The dynamical friction timescale for the binary to sink to the dense inner region of its parent halo is comparable or shorter than the accretion timescale required to increase their mass above the PI limit. Once in the core of the halo, the binary can enter a phase of hyper-Eddington accretion, where it would only take a few thousand years to exceed the PI limit through accretion. Even more massive BBHs could form through this channel, and be detectable by detectors with improved low-frequency sensitivity. Single Pop~III BH remnants would also grow through accretion and could later form binaries dynamically. As little as a few percent of Pop~III BH remnants may be sufficient to match the rate of massive BBH mergers inferred from GW190521 of $0.13^{+0.3}_{-0.11}\rm Gpc^{-3} yr^{-1}$.

astro-ph.HE↗

arXiv2020

Constraining the delay time distribution of compact binary objects from the stochastic gravitational wave background searches

The initial separation of massive star binaries sets the timescale over which their compact remnants merge through the emission of gravitational waves. We show that the delay time distribution (DTD) of binary neutron stars or black holes can be inferred from the stochastic gravitational wave background (SGWB). If the DTD of a population is long, most of the mergers take place at low redshifts and the background would be rather quiet compared to a scenario in which the DTD is short leading to few individual detections at low redshift but a rather loud background. We show that different DTDs predict a factor of 5 difference in the magnitude of the gravitational wave background energy density ($Ω_{\rm GW}$) and have the dominant effect on $Ω_{\rm GW}$ over other factors such as the mass function of the primary BH mass, $p(m_1)$, the maximum considered BH mass ($M_{\rm max}$), and the effective spin of the black hole ($χ_{\rm eff}$). A non-detection of such a background can rule out the short DTD scenario. We show that SGWB searches can rule out the short DTD scenario for the BBHs within about four years of observing time at advanced LIGO design sensitivty for a local merger rate of 30 $\rm Gpc^{-3} yr^{-1}$ assuming $p(m_1)\propto m_1^{-1}$, and $M_{\rm max}=50 M_{\odot}$.

astro-ph.HE↗

arXiv2020

LB-1 is inconsistent with the X-ray source population and pulsar-black hole binary searches in the Milky Way

If confirmed, a wide binary system of 70 $M_{\odot}$ black hole (BH) and an 8 $M_{\odot}$ main sequence star (LB-1) is observed to reside in the Milky Way (MW). We show that long term evolution of an 8 $M_{\odot}$ star around a BH with mass between 5-70 $M_{\odot}$ makes them visible as ultra-luminous X-ray (ULX) sources in the sky. Given the expected ULX phase lifetime ($\approx0.1$ Myr) and their lack of detection in the MW, we conclude that the frequency of an 8-20 $M_{\odot}$ star to be in binary around a stellar mass BH should be less than ($f<2\times10^{-3}$). This is in tension with Liu et al. (2019) claimed detection frequency of LB-1 like system around 8-20$M_{\odot}$ stars ($f\approx3\times10^{-2}$). Moreover, the 8 $M_{\odot}$ star is likely to end as a neutron star (NS) born with a very small kick from an electron capture supernova (ECSN), leaving behind a wide NS-BH binary. So far less than 1\% of all the detectable pulsars in the MW are mapped and there has been no detection of any pulsars in binary systems around BHs which sets an upper bound of about 100 possible pulsar-BH systems in the MW. We show if the NS is born from ECSN, a frequency upper limit of ($f=5\times10^{-4}$) for stars with masses $\approx 8-20~M_{\odot}$ in the MW to have a BH companion. The rate discrepancy will further increase as more pulsars are mapped in the MW, yet these searches would not be able to rule out the Liu et al. detection frequency if NSs are instead born in core collapse SNe with the commonly inferred high kick velocities.

astro-ph.HE↗

arXiv2020

Numerical Relativity Simulations of the Neutron Star Merger GW170817: Long-Term Remnant Evolutions, Winds, Remnant Disks, and Nucleosynthesis

We present a systematic numerical-relativity study of the dynamical ejecta, winds and nucleosynthesis in neutron star merger remnants. Binaries with the chirp mass compatible with GW170817, different mass ratios, and five microphysical equations of state (EOS) are simulated with an approximate neutrino transport and a subgrid model for magnetohydrodynamics turbulence up to 100 milliseconds postmerger. Spiral density waves propagating from the neutron star remnant to the disk trigger a wind with mass flux ${\sim}0.1{-}0.5\,{\rm M_\odot/s}$ persisting for the entire simulation as long as the remnant does not collapse to black hole. This wind has average electron fraction $\gtrsim 0.3$ and average velocity ${\sim}0.1-0.17\,$c and thus is a site for the production of weak $r$-process elements (mass number $A<195$). Disks around long-lived remnants have masses ${\sim}0.1{-}0.2\,{\rm M_\odot}$, temperatures peaking at $\lesssim10\,$MeV near the inner edge, and a characteristic double-peak distribution in entropy resulting from shocks propagating through the disk. The dynamical and spiral-wave ejecta computed in our targeted simulations are not compatible with those inferred from AT2017gfo using two-components kilonova models. Rather, they indicate that multi-component kilonova models including disk winds are necessary to interpret AT2017gfo. The nucleosynthesis in the combined dynamical ejecta and spiral-wave wind in the comparable-mass long-lived mergers robustly accounts for all the $r$-process peaks, from mass number ${\sim}75$ to actinides in terms of solar abundances. Total abundandes are weakly dependent on the EOS, while the mass ratio affect the production of first peak elements.

astro-ph.HE↗

arXiv2020

Formation of mass gap objects in highly asymmetric mergers

The LIGO/Virgo Collaboration (LVC) recently reported the detection of GW190814, a merger of a $23^{+1.0}_{-0.9}~M_{\odot}$ primary black hole (BH), and a $2.6^{+0.08}_{-0.08}~M_{\odot}$ secondary. The secondary's mass falls into the mass-gap regime, which refers to the scarcity of compact objects in the mass range of 2-5 $M_{\odot}$. The first clue to the formation of the GW190814 lies in the fact that the primary is a very massive BH. We suggest that the secondary was born as a neutron star (NS) where a significant amount of the supernova ejecta mass from its formation remained bound to the binary due to the presence of the massive BH companion. The bound mass forms a circumbinary accretion disk, and its accretion onto the NS created a mass-gap object. In this scenario, LIGO/Virgo will only detect mass-gap objects in binary mergers with an extreme mass ratio. We also predict a correlation between the mass of the secondary and the mass of the primary in such asymmetric mergers. Our model can be tested with future data from the LVC's third-observing run.

astro-ph.HE↗