arXiv subjects

Chirag Chawla

Publications and source records attributed to Chirag Chawla.

11 recordsLinked to original sources

arXiv2026

Gaia's promise to detect compact-object binaries: where we stand with the third data release

With its third data release (DR3), Gaia begins unveiling dormant candidate compact object (CO) binaries with luminous companions (LC) as predicted by several past theoretical studies. To date, 3 black hole (BH), 21 neutron star (NS), and 3200 white dwarf (WD) candidates have been identified with LCs in detached orbits using astrometry. We adopt an observationally motivated sampling scheme for the star formation history of the Milky Way, and initial zero-age main-sequence binary properties, incorporate all relevant binary interaction processes during evolution to obtain a realistic present-day intrinsic population of CO--LC binaries. We apply Gaia's selection criteria to identify the CO-LC binaries detectable using the observational cuts applicable for DR3 as well as its end-of-mission (EOM). We find that under the DR3 selection cuts, our detectable population includes no BH--LCs, approximately 10-40 NS--LCs, and around ~4300 WD--LCs. Our predicted NS--LC population is in good agreement with the current DR3 census, both in its predicted yield and in the orbital and stellar properties, and we recover a close analogue of the Gaia NS1 candidate together with its detailed formation pathway. For WD--LCs, we find that a moderate natal kick of 5-15 km/s imparted at WD formation is required to match the observed orbital properties of WD-LC candidates in DR3. We further show that Gaia BH3-like binaries can form through standard isolated binary evolution without invoking any additional modelling assumptions, whereas reproducing Gaia BH1 and BH2 remains challenging within this framework. Looking ahead to the EOM, we predict detection of ~30-300 BH--LCs, ~1500-5000 NS-LCs, and ~10^5-10^6 WD-LC binaries, primarily due to the significantly longer observational baseline.

astro-ph.SR↗

arXiv2026

ALIGNBEAM : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing

Domain fine-tuning degrades the safety of large language models: fine-tuned specialists readily comply with harmful prompts framed in domain language. Existing inference-time defenses that mix logits from a safe anchor model require both models to share a vocabulary, which rules them out for the cross-family specialists where safety is most degraded. We present ALIGNBEAM, a training-free method that lifts this restriction by translating anchor logits into the target model's vocabulary token-by-token at each decoding step; a small LLM judge then selects the safest among K candidate continuations. No weights are changed, and the safety-utility trade-off can be tuned at deployment without retraining. Across both cross-vocabulary and same-vocabulary evaluation pairs, ALIGNBEAM substantially raises refusal on adversarial benchmarks while keeping task accuracy and inference overhead within practical bounds. The results show that safety alignment can be transferred between model families at inference time, without touching either model's weights.

cs.CL↗

arXiv2026

Selective-Advantage Entropy-Adaptive Horizon GRPO: Asymmetric Token-Level Discounting for Efficient Reinforcement Learning of Language Models

Group Relative Policy Optimisation (GRPO) has emerged as an effective reinforcement-learning algorithm for aligning language models on reasoning tasks, but it treats every token position and every sampled rollout symmetrically. We introduce two complementary extensions: (i) Adaptive-Horizon GRPO (AH-GRPO), which weights each token's policy gradient using a cumulative entropy-based discount that reduces the effective horizon when the model is uncertain, and (ii) Selective-Advantage AH-GRPO (SA-AH-GRPO), which applies this discounting only to negative-advantage rollouts, leaving positive-advantage, successful trajectories unattenuated. We evaluate standard GRPO with alpha = 0, AH-GRPO with alpha = 0.5, and SA-AH-GRPO with alpha = 0.5 on the GSM8K mathematical reasoning benchmark using both Qwen 2.5-1.5B-Instruct and Qwen 2.5-3B-Instruct fine-tuned with LoRA. On the 3B model, SA-AH-GRPO achieves Pass@1 = 0.858 at its peak at step 30 and maintains 0.846 at 180 steps, with training variance reduced to 0.0246, a 3.6 times reduction relative to GRPO while matching its peak accuracy. On the 1.5B model, SA-AH-GRPO achieves a peak Pass@1 of 0.686, improving over the zero-shot baseline of 0.637. Our analysis shows that asymmetric discounting preserves the full gradient signal on correct solutions, prevents entropy collapse, and substantially stabilises training, suggesting a principled inductive bias for reinforcement learning with verifiable rewards on structured generation tasks.

cs.LG↗

arXiv2026

Detecting the Machine: A Comprehensive Benchmark of AI-Generated Text Detectors Across Architectures, Domains, and Adversarial Conditions

The rapid proliferation of large language models (LLMs) has created an urgent need for robust and generalizable detectors of machine-generated text. Existing benchmarks typically evaluate a single detector on a single dataset under ideal conditions, leaving open questions about cross-domain transfer, cross-LLM generalization, and adversarial robustness. We present a comprehensive benchmark evaluating diverse detection approaches across two corpora: HC3 (23,363 human-ChatGPT pairs) and ELI5 (15,000 human-Mistral-7B pairs). Methods include classical classifiers, fine-tuned transformer encoders (BERT, RoBERTa, ELECTRA, DistilBERT, DeBERTa-v3), a CNN, an XGBoost stylometric model, perplexity-based detectors, and LLM-as-detector prompting. Results show that transformer models achieve near-perfect in-distribution performance but degrade under domain shift. The XGBoost stylometric model matches performance while remaining interpretable. LLM-based detectors underperform and are affected by generator-detector identity bias. Perplexity-based methods exhibit polarity inversion, with modern LLM outputs showing lower perplexity than human text, but remain effective when corrected. No method generalizes robustly across domains and LLM sources.

cs.CL↗

arXiv2026

AlignTune: Modular Toolkit for Post-Training Alignment of Large Language Models

Post-training alignment is central to deploying large language models (LLMs), yet practical workflows remain split across backend-specific tools and ad-hoc glue code, making experiments hard to reproduce. We identify backend interference, reward fragmentation, and irreproducible pipelines as key obstacles in alignment research. We introduce AlignTune, a modular toolkit exposing a unified interface for supervised fine-tuning (SFT) and RLHF-style optimization with interchangeable TRL and Unsloth backends. AlignTune standardizes configuration, provides an extensible reward layer (rule-based and learned), and integrates evaluation over standard benchmarks and custom tasks. By isolating backend-specific logic behind a single factory boundary, AlignTune enables controlled comparisons and reproducible alignment experiments.

cs.CL↗

arXiv2025

PERCS: Persona-Guided Controllable Biomedical Summarization Dataset

Automatic medical text simplification plays a key role in improving health literacy by making complex biomedical research accessible to diverse readers. However, most existing resources assume a single generic audience, overlooking the wide variation in medical literacy and information needs across user groups. To address this limitation, we introduce PERCS (Persona-guided Controllable Summarization), a dataset of biomedical abstracts paired with summaries tailored to four personas: Laypersons, Premedical Students, Non-medical Researchers, and Medical Experts. These personas represent different levels of medical literacy and information needs, emphasizing the need for targeted, audience-specific summarization. Each summary in PERCS was reviewed by physicians for factual accuracy and persona alignment using a detailed error taxonomy. Technical validation shows clear differences in readability, vocabulary, and content depth across personas. Along with describing the dataset, we benchmark four large language models on PERCS using automatic evaluation metrics that assess comprehensiveness, readability, and faithfulness, establishing baseline results for future research. The dataset, annotation guidelines, and evaluation materials are publicly available to support research on persona-specific communication and controllable biomedical summarization.

cs.CL↗

arXiv2025

Realistic predictions for Gaia black hole discoveries: comparison of isolated binary and dynamical formation models

Astrometry from Gaia has enabled discovery of three dormant black holes (BHs) in au-scale binaries. Numerous models have been proposed to explain their formation, including several that have forecasted Gaia detections. However, previous works have used simplified detectability metrics that do not capture key elements of the Gaia astrometric orbit selection function. We apply a realistic forward-model of Gaia astrometric orbit catalogs to BH binary populations generated through (a) isolated binary evolution (IBE) and (b) dynamical formation in star clusters. For both formation channels, we analyze binary populations in a simulated Milky Way-like galaxy with a realistic metallicity-dependent star formation history and 3D dust map. We generate epoch astrometry for each binary from the Gaia scanning law and fit it with the cascade of astrometric models used in Gaia DR3. The IBE model of Chawla et al. (2022) predicts that no BH binaries should have been detected in DR3 and thus significantly underpredicts the formation rate of Gaia BHs. In contrast, the dynamical model of Di Carlo et al. (2024) overpredicts the number of BHs receiving DR3 orbital solutions by a factor of $\sim$8. The two models predict very different orbital period distributions, with the IBE model predicting only binaries that avoided common envelope evolution and have $P_{\text{orb}} \gtrsim 2,000$ d to be detectable, and the dynamical formation model predicting a period distribution that is roughly log-uniform. Adopting the dynamical channel as a fiducial model and rescaling by a factor of 1/8 to match DR3, we predict that $\sim$30 BH binaries will be detected in Gaia DR4, representing $\sim0.1\%$ of Milky Way BHs with luminous companions in au-scale orbits.

astro-ph.GA↗

arXiv2024

Detecting Detached Black Hole binaries through Photometric Variability

Understanding the connection between the properties of black holes (BHs) and their progenitors is interesting in many branches of astrophysics. Discovering BHs in detached orbits with luminous companions (LCs) promises to help create this map since the LC and BH progenitor are expected to have the same metallicity and formation time. We explore the possibility of detecting BH-LC binaries in detached orbits using photometric variations of the LC flux, induced by tidal ellipsoidal variation, relativistic beaming, and self-lensing. We create realistic present-day populations of detached BH-LC binaries in the Milky Way (MW) using binary population synthesis where we adopt observationally motivated initial stellar and binary properties, star formation history and present-day distribution of these sources in the MW based on detailed cosmological simulations. We test detectability of these sources via photometric variability by Gaia and TESS missions by incorporating their respective detailed detection biases as well as interstellar extinction. We find that Gaia is expected to resolve 300--1,000 (700--1,500) detached BH--LC binaries with SNR>10 (1) depending on the photometric precision and details of supernova physics. Similarly, the number of resolved BH--LC binaries with TESS are ~50--200 (140--350). We find that 136^{+15}_{-15} BH--LC binaries would be common between Gaia and TESS. Moreover, between ~60--70 (50--200) BH--LC binaries identifiable using photometry with SNR >10 may also be resolved using Gaia's radial velocity (astrometry).

astro-ph.GA↗

arXiv2023

Dynamically Forming Extremely Low-Mass White Dwarf Binaries in Wide Orbits

The detection of a 0.2 Msun extremely low-mass white dwarf (EW) in a wide orbit (Porb about 450 days) with a 1.1 Msun main-sequence (MS) companion, KIC 8145411, challenges our current understanding of how EWs form. The traditional channel for EW formation, via mass transfer from the WD progenitor, is expected to form EW binaries in tight orbits. Indeed, the majority of known EWs are found in tight binaries with a median Porb about 5.4 hrs. Using numerical scattering experiments, we find that binary-binary strong encounters in star clusters can sufficiently widen the orbit of a typical EW binary, to explain the observed wide orbit of the KIC 8145411 system. The Porb distribution for EW binaries produced through binary-binary encounters is bimodal: one mode corresponds to the initial orbital period of the EW binary, while the other is near Porb about a few 100 days, similar to the orbital period of the KIC 8145411 system. We find that the production of wide EW binaries that are also ejected from the cluster peaks at a star cluster mass of about 10^5 Msun with a rate of about 10^{-3} Gyr^{-1}. Assuming that 50% of all stars form in star clusters and an initial cluster mass function proportional to m^{-2}, we estimate a galactic formation rate of about 4.16 x 10^3 Gyr^{-1} for wide EW binaries.

astro-ph.SR↗

arXiv2022

Gaia May Detect Hundreds of Well-characterised Stellar Black Holes

Detection of black holes (BHs) with detached luminous companions (LCs) can be instrumental in connecting the BH properties with their progenitors' since the latter can be inferred from the observable properties of the LC. Past studies showed the promise of Gaia astrometry in detecting BH-LC binaries. We build upon these studies by: 1) initialising the zero-age binary properties based on realistic, metallicity-dependent star-formation history in the Milky Way (MW), 2) evolving these binaries to current epoch to generate realistic MW populations of BH-LC binaries, 3) distributing these binaries in the MW preserving the complex age-metallicity-Galactic position correlations, 4) accounting for extinction and reddening using three-dimensional dust maps, 5) examining the extended Gaia mission's ability to resolve BH-LC binaries. We restrict ourselves to detached BH-LC binaries with orbital period <10 yr such that Gaia can observe at least one full orbit. We find: 1) the extended Gaia mission can astrometrically resolve 30-300 detached BH-LC binaries depending on our assumptions of supernova physics and astrometric detection threshold; 2) Gaia's astrometry alone can indicate BH candidates for 10-100 BH-LC binaries by constraining the dark primary mass >3 Msun; 3) distributions of observables including orbital periods, eccentricities, and component masses are sensitive to the adopted binary evolution model, hence can directly inform binary evolution models. Finally, we comment on the potential to further characterise these BH binaries through radial velocity measurements and observation of X-ray counterparts.

astro-ph.GA↗

arXiv2021

Weighing the Darkness II: Astrometric Measurement of Partial Orbits with Gaia

Over the course of several years, stars trace helical trajectories as they traverse across the sky due to the combined effects of proper motion and parallax. It is well known that the gravitational pull of an unseen companion can cause deviations to these tracks. Several studies have pointed out that the astrometric mission Gaia will be able to identify a slew of new exoplanets, stellar binaries, and compact object companions with orbital periods as short as tens of days to as long as Gaia's lifetime. Here, we use mock astrometric observations to demonstrate that Gaia can identify and characterize black hole companions to luminous stars with orbital periods longer than Gaia's lifetime. Such astrometric binaries have orbital periods too long to exhibit complete orbits, and instead are identified through curvature in their characteristic helical paths. By simultaneously measuring the radius of this curvature and the orbital velocity, constraints can be placed on the underlying orbit. We quantify the precision with which Gaia can measure orbital accelerations and apply that to model predictions for the population of black holes orbiting stars in the stellar neighborhood. Although orbital degeneracies imply that many of the accelerations induced by hidden black holes could also be explained by faint low-mass stars, we discuss how the nature of certain putative black hole companions can be confirmed with high confidence using Gaia data alone.

astro-ph.SR↗