arXiv subjects

Yash Aggarwal

Publications and source records attributed to Yash Aggarwal.

11 recordsLinked to original sources

arXiv2026

Fireworks at Cosmic Dawn: relieving BAO-CMB tensions with the Pop III.1 Flash

A Cosmic Microwave Background (CMB) optical depth of $\tau \sim 0.09$, several $\sigma$ in excess of the latest Planck low-$\ell$ EE polarization measurement, has been proposed as a way to reconcile the preference for a sub-minimal neutrino mass sum in a combined analysis with CMB and Dark Energy Spectroscopic Instrument (DESI) three-year data. Reionization, however, is not just probed by $\tau$. It is also constrained by Ly$\alpha$ forest observations that indicate a late end of reionization, and the patchy kinetic Sunyaev-Zel'dovich (pkSZ) effect which prefers a short duration. We explore whether an early phase of reionization can achieve a high $\tau$ while remaining consistent with both Ly$\alpha$ forest and pkSZ constraints. As a concrete example, we consider supermassive Pop III.1 stars, dark-matter-powered metal-free stars proposed as progenitors of supermassive black holes. Within this framework, self-regulating ionizing feedback imposes a minimum source separation of $\sim 1 \, \text{cMpc}$, consequently limiting large-scale ionization fluctuations and reducing the pkSZ power on observationally relevant scales. Our fiducial model realizes an optical depth of $\tau = 0.087$ with a Pop III.1-driven flash ionization phase centered at $z = 20$, while evading the most conservative $2\sigma$ upper limits on the pkSZ signal from the most recent South Pole Telescope data release. More broadly, our findings motivate further exploration of early reionization models with weakly clustered sources as a possible resolution of tensions between BAO and CMB measurements.

astro-ph.CO↗

arXiv2026

Moral Sensitivity in LLMs: A Tiered Evaluation of Contextual Bias via Behavioral Profiling and Mechanistic Interpretability

Large language models (LLMs) are increasingly deployed in settings that require nuanced ethical reasoning, yet existing bias evaluations treat model outputs as simply "biased" or "unbiased." This binary framing misses the gradual, context-sensitive way bias actually emerges. We address this gap in two stages: behavioral profiling and mechanistic validation. In the behavioral stage, we introduce the Moral Sensitivity Index (MSI), a metric that quantifies the probability of biased output across a graduated, seven-tier stress test ranging from abstract numerical problems to scenarios rooted in historical and socioeconomic injustice. Evaluating four leading models (Claude 3.5, Qwen 3.5, Llama 3, and Gemini 1.5), we identify distinct behavioral signatures shaped by alignment design: for instance, Gemini 1.5 reaches 72.7% MSI by Tier 5 under socioeconomic framing, while Claude exhibits sharp suppression consistent with identity-based safety training. We then verify these behavioral patterns mechanistically. We select criminal-bias scenarios, which produced the highest MSI scores across models, as probes and apply logit lens, attention analysis, activation patching, and semantic probing to a controlled set of six models spanning three capability tiers: small language models (SLMs), instruction-tuned base models, and reasoning-distilled variants. Circuit-level analysis reveals a U-curve of bias: SLMs exhibit strong criminal bias; scaling to instruction-tuned models eliminates it; reasoning distillation reintroduces bias to SLM-like levels despite identical parameter counts, suggesting distillation compresses reasoning traces in ways that reactivate shallow statistical associations. Critically, the socially loaded cues that drive high MSI scores activate the same bias-driving circuits identified mechanistically, providing cross-stage validation.

cs.LG↗

arXiv2026

Temporal Dependencies in In-Context Learning: The Role of Induction Heads

Large language models (LLMs) exhibit strong in-context learning capabilities, but how they track and retrieve information from context remains underexplored. Drawing on the free recall paradigm in cognitive science (where participants recall list items in any order), we show that several open-source LLMs consistently display a serial-recall-like pattern, assigning peak probability to tokens that immediately follow a repeated token in the input sequence. Through systematic ablation experiments, we show that induction heads, specialized attention heads that attend to the token following a previous occurrence of the current token, play an important role in this phenomenon. Removing heads with a high induction score substantially reduces the +1 lag bias, whereas ablating random heads does not reproduce the same reduction. We also show that removing heads with high induction scores impairs the performance of models prompted to do serial recall using few-shot learning to a larger extent than removing random heads. Our findings highlight a mechanistically specific connection between induction heads and temporal context processing in transformers, suggesting that these heads are especially important for ordered retrieval and serial-recall-like behavior during in-context learning.

cs.CL↗

arXiv2025

Beyond Semantics: How Temporal Biases Shape Retrieval in Transformer and State-Space Models

In-context learning is governed by both temporal and semantic relationships, shaping how Large Language Models (LLMs) retrieve contextual information. Analogous to human episodic memory, where the retrieval of specific events is enabled by separating events that happened at different times, this work probes the ability of various pretrained LLMs, including transformer and state-space models, to differentiate and retrieve temporally separated events. Specifically, we prompted models with sequences containing multiple presentations of the same token, which reappears at the sequence end. By fixing the positions of these repeated tokens and permuting all others, we removed semantic confounds and isolated temporal effects on next-token prediction. Across diverse sequences, models consistently placed the highest probabilities on tokens following a repeated token, but with a notable bias for those nearest the beginning or end of the input. An ablation experiment linked this phenomenon in transformers to induction heads. Extending the analysis to unique semantic contexts with partial overlap further demonstrated that memories embedded in the middle of a prompt are retrieved less reliably. Despite architectural differences, state-space and transformer models showed comparable temporal biases. Our findings deepen the understanding of temporal biases in in-context learning and offer an illustration of how these biases can enable temporal separation and episodic retrieval.

cs.CL↗

arXiv2025

On the origins, growth, and radiative efficiency of J0529-4351, reportedly the fastest-growing known black hole

SMSS J0521-4351 is reportedly the most luminous quasar known to date, and assuming a mean radiative efficiency of 0.1, it is inferred to be the fastest-growing black hole, accreting approximately one solar mass per day. Assessing the implications of this assumption on the seed mass and inception time of J0529-4351, we show that the inferred accretion rate is unreasonably high and that its radiative efficiency must be much greater than 0.1. Then, we derive its accretion rate and seed mass, and for comparison of three other similar-size (~1-2E+10 solar masses) black holes at various redshifts, using well-tested empirical scaling relations. The results indicate that J0529-4351 grew from a heavy seed (~2-3E+04 solar masses), and that its accretion rate (~10-13 solar masses/year) is the lowest of the four black holes. However, its radiative efficiency inferred from its bolometric luminosity and the derived accretion rate is the highest, which explains why it is the most luminous despite having the lowest accretion rate. This study challenges the prevailing notion that a higher luminosity or a higher Eddington ratio implies a higher accretion rate, highlights the dependence of a black hole's luminosity on radiative efficiency, reveals the pitfalls of inferring black hole properties assuming a standard value for radiative efficiency, and suggests that the Eddington ratios of high-luminosity BHs may be significantly overestimated.

astro-ph.GA↗

arXiv2025

Direct Collapse Black Hole Candidates from Decaying Dark Matter

Injecting 1-13.6 eV photons into the early universe can suppress the molecular hydrogen abundance and alter the star formation history dramatically enough to produce direct collapse black holes. These, in turn, could explain the recently observed population of puzzling high-redshift supermassive black holes that appear to require super-Eddington accretion. We show that axion dark matter decay in the intergalactic medium can account for this energy injection. We use a single zone model of the gas core and semi-analytically evolve its chemo-thermal properties to track the conditions for which the system becomes an atomic cooling halo-a necessary precursor for the production of heavy black hole seeds to explain the high-redshift black hole population. Windows of axions masses between 24.5-26.5 eV with photon couplings as low as $4\times 10^{-12}$/GeV may realize this atomic cooling halo condition. We highlight the significance of the band structure of molecular hydrogen on the effectiveness of this process and discuss estimates of the heavy seed population and prospects for testing this model.

hep-ph↗

arXiv2025

Emergence of Episodic Memory in Transformers: Characterizing Changes in Temporal Structure of Attention Scores During Training

We investigate in-context temporal biases in attention heads and transformer outputs. Using cognitive science methodologies, we analyze attention scores and outputs of the GPT-2 models of varying sizes. Across attention heads, we observe effects characteristic of human episodic memory, including temporal contiguity, primacy and recency. Transformer outputs demonstrate a tendency toward in-context serial recall. Importantly, this effect is eliminated after the ablation of the induction heads, which are the driving force behind the contiguity effect. Our findings offer insights into how transformers organize information temporally during in-context learning, shedding light on their similarities and differences with human memory and learning.

cs.LG↗

arXiv2024

Evidence that Eddington ratio depends upon a supermassive black hole's mass and redshift: Implications for radiative efficiency

Presently, it is unclear whether the Eddington ratio and radiative efficiency depend upon a supermassive black hole's (SMBH's) redshift z and mass MBH. We attempt to resolve this issue using published data for 132,000 SMBHs with MBH >1E+7 Msun (solar masses) at ~0.1<z<2.4 covering ~10 billion years of cosmic time, with MBH determined using MgII lines and bolometric luminosities (Lbol) based on a weighted mean of Lbol from two or more monochromatic luminosities and a single uniformly applied correction factor. The SMBHs are sorted into 7 MBH bins separated from each other by half an order of magnitude. The Eddington ratio and z data in each bin are subjected to spline regression analysis. The results unambiguously show that for similar-size SMBHs, the Eddington ratio decreases as z decreases and that for a given redshift larger SMBHs have a lower Eddington ratio. These findings require that either a SMBH's accretion rate and/or its radiative efficiency be a function of z and MBH and, in the context of the Bondi accretion model, imply that radiative efficiency is an inverse function of a SMBH's redshift z and mass MBH. These findings suggest that SMBHs become less efficient (higher radiative efficiency) in accreting gases as the ambient gas density decreases with z and that larger SMBHs are more efficient (lower radiative efficiency) than smaller ones. The results leave little doubt that the current widespread practice of assigning radiative efficiency a standard value is untenable and gives erroneous estimates of accretion rates and growth times of SMBHs.

astro-ph.GA↗

arXiv2023

Accretion rates and radiative efficiencies of Sagittarius A* and nearby supermassive black holes estimated using empirical relations: Implications for accretion models

The Bondi accretion rate of black holes in our and nearby galaxies Messier 87, NGC 3115, NGC 1600, and Cygnus A have been determined or constrained using Chandra or other observations. It, however, remains unknown how much mass from the Bondi radius reaches each black hole and how much is accreted. We determine the accretion rate and radiative efficiency for each black hole using two well-tested empirical relations: one relates a black hole's accretion rate to its mass and redshift, and the other relates the radiative efficiency to the Eddington ratio and redshift. We get an accretion rate of ~0.00002 solar mass per year and radiative efficiency of ~0.9 for Sagittarius A* and an accretion rate of ~0.09 solar masses per year and radiative efficiency of ~0.68 for NGC 1600; and values in between these extremes for the rest. The derived mass inflow rate onto each black hole (not the accretion rate) essentially matches the reported Bondi accretion rate. Thus, the results do not support the ADIOS and CDAF models, but whether the dissipated energy not reflected in a black hole's observed luminosity is advected as in the ADAF model remains uncertain. Furthermore, contrary to current model expectations, the derived radiative efficiencies are orders of magnitude higher and radiative efficiency increases as the accretion rate decreases and a BH ages. A physical basis is found relating the empirical formulation of accretion rate to Bondi accretion.

astro-ph.GA↗

arXiv2023

Comparative tests of accretion rates of quasars derived using a new empirical and an existing theoretical relation: Insights into black hole properties and growth

A scaling relation based on thin-disc accretion theory has been used by some workers to determine the mass-inflow rate onto 20 high-redshift (z) and 80 Palomar-Green quasars. Based on several assumptions, it inexplicably implies that the inflow rate is an inverse function of black-hole (BH) mass Mbh. Moreover, its results remain untested. This paper offers a simple empirical relation essentially free of assumptions, found using available data for 59 highest-z quasars and the so-called Salpeter relation. We find that the accretion rate is proportional to Mbh(1+z)3, consistent with conventional astrophysics that the accretion rate is a direct function of both Mbh and the ambient gas density. We apply it to the 20 high-z and a subset of Palomar-Green quasars. Comparative analyses show that all empirically derived accretion rates and radiative efficiencies pass the tests, but their theoretical counterparts fail in most cases. A secondary relation defines the Eddington ratio as a function of z and radiative efficiency. Consistent with the empirical relations, spline regression analysis of Kozlowski's data for 132,000 quasars at z<2.4 shows that both the Eddington ratio and radiative efficiency are functions of Mbh and z. The results show that bigger BHs accrete more efficiently the smaller ones. For BHs > a billion solar masses, we get radiative efficiency of ~0.23 at z>5.7 and ~0.84 at z<0.005. Notably, the empirical relations predict a mass-inflow rate of 0.11-0,21 solar mass/year on to the BH in M87 that matches its Bondi accretion rate determined using observed density and temperature profiles.

astro-ph.GA↗

arXiv2021

Deciphering the origins and growth of supermassive black holes

We present a well-tested, theoretically supported empirical relation that helps decipher the origins, growth, and properties of SMBHs (supermassive black holes). Based on theoretical considerations and analysis of mass (MBH) versus age (t) distribution of 93 high-redshift (z>5.6) SMBHs, we get MBH = Ms exp [14.6(t-100)/t (Myr)], which gives the SMBH's seed mass Ms and its derivative gives the instantaneous mass-accretion rate. It yields seeds of ~ 20-420 Msun (solar masses) for the recently discovered SMBHs GNz11, CEERS-19, and UHZ1 and (~ 3E+04 Msun) for the largest (1.24E+10 Msun) high-z SMBH. It is applied to 132446 SMBHs at z <2.4, cataloged by Kozlowski. The resultant seeds are classified based on size and likely formation mechanism: ~54 percent are classified as light (<350 Msun) deemed to be Pop III remnants; ~40 percent as intermediate (350-2 x 3E+03 Msun) and ~6 percent as heavier seeds (2 x 103-3E+04 Msun), both of which may have formed by mergers of Pop III remnants. The direct collapse black hole (DCBH) mechanism is not required but is not excluded. Furthermore, the results show the following. The mass-accretion rate increases exponentially from the seed's inception at z ~30, reaches a broad plateau at z ~ 8.5 to 6 coincident with the epoch of reionization, and decreases monotonically towards z=0. Sub-Eddington accretion is the norm, except during the first ~150 Myr, SMBHs experienced super-Eddington accretion, or the radiative efficiency was <0.1. The largest seed can potentially grow via luminous accretion up to (6.6+/-2.2)E+10 Msun, consistent with a theoretical limit of ~ 5+10 Msun proposed by King. The Eddington ratio decreases, and the radiative efficiency increases as z decreases, consistent with recent findings.

astro-ph.GA↗