arXiv subjects

Christine Ye

Publications and source records attributed to Christine Ye.

13 recordsLinked to original sources

arXiv2026

Efficiently Aligning Language Models with Online Natural Language Feedback

Reinforcement learning with verifiable rewards has been used to elicit impressive performance from language models in many domains. But, broadly beneficial deployments of AI may require us to train models with strong capabilities in "fuzzy", hard-to-supervise domains. In this paper, we develop methods to align language models in fuzzy domains where human experts are still able to provide high-quality supervision signal, but only for a small number of model outputs, using online natural language feedback. Specifically, we train models by iteratively optimizing against proxy reward signals, stopping at the point of over-optimization, collecting fresh expert supervision, and updating the proxy reward. We construct proxy reward models from language models using in-context learning (ICL) and fine-tuning. We test our methods by eliciting creative writing and alignment research capabilities in Qwen3-8B and Haiku 4.5 respectively. For Qwen3-8B, ICL methods recover up to 35% of performance with 50x fewer expert samples, while fine-tuning methods recover 80% with up to 20x fewer samples and 100% with 3x fewer samples. For Haiku 4.5, ICL methods recover up to 35% of performance with 30x fewer samples, and fine-tuning methods recover 100% with 10x fewer samples. Our results suggest that online natural language feedback can substantially improve the data efficiency of expert supervision.

cs.LG↗

arXiv2025

The NANOGrav 12.5-year Data Set: Chromatic Noise Characterization & Mitigation with Time-Domain Kernels

Pulsar timing arrays (PTAs) have recently entered the detection era, quickly moving beyond the goal of simply improving sensitivity at the lowest frequencies for the sake of observing the stochastic gravitational wave background (GWB), and focusing on its accurate spectral characterization. While all PTA collaborations around the world use Fourier-domain Gaussian processes to model the GWB and intrinsic long time-correlated (red) noise, techniques to model the time-correlated radio frequency-dependent (chromatic) processes have varied from collaboration to collaboration. Here we test a new class of models for PTA data, Gaussian processes based on time-domain kernels that model the statistics of the chromatic processes starting from the covariance matrix. As we will show, these models can be effectively equivalent to Fourier-domain models in mitigating chromatic noise. This work presents a method for Bayesian model selection across the various choices of kernel as well as deterministic chromatic models for non-stationary chromatic events and the solar wind. As PTAs turn towards high frequency (>1/yr) sensitivity, the size of the basis used to model these processes will need to increase, and these time-domain models present some computational efficiencies compared to Fourier-domain models.

astro-ph.IM↗

arXiv2025

ReplicationBench: Can AI Agents Replicate Astrophysics Research Papers?

Frontier AI agents show increasing promise as scientific research assistants, and may eventually be useful for extended, open-ended research workflows. However, in order to use agents for novel research, we must first assess the underlying faithfulness and correctness of their work. To evaluate agents as research assistants, we introduce ReplicationBench, an evaluation framework that tests whether agents can replicate entire research papers drawn from the astrophysics literature. Astrophysics, where research relies heavily on archival data and computational study while requiring little real-world experimentation, is a particularly useful testbed for AI agents in scientific research. We split each paper into tasks which require agents to replicate the paper's core contributions, including the experimental setup, derivations, data analysis, and codebase. Each task is co-developed with the original paper authors and targets a key scientific result, enabling objective evaluation of both faithfulness (adherence to original methods) and correctness (technical accuracy of results). ReplicationBench is extremely challenging for current frontier language models: even the best-performing language models score under 20%. We analyze ReplicationBench trajectories in collaboration with domain experts and find a rich, diverse set of failure modes for agents in scientific research. ReplicationBench establishes the first benchmark of paper-scale, expert-validated astrophysics research tasks, reveals insights about agent performance generalizable to other domains of data-driven science, and provides a scalable framework for measuring AI agents' reliability in scientific research.

cs.CL↗

arXiv2025

Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignment

Large language models are sometimes trained with imperfect oversight signals, leading to undesired behaviors such as reward hacking and sycophancy. Improving oversight quality can be expensive or infeasible, motivating methods that improve learned behavior despite an imperfect training signal. We introduce Inoculation Prompting (IP), a simple but counterintuitive technique that prevents learning of an undesired behavior by modifying training prompts to explicitly request it. For example, to inoculate against reward hacking, we modify the prompts used in supervised fine-tuning to request code that only works on provided test cases but fails on other inputs. Across four settings we find that IP reduces the learning of undesired behavior without substantially reducing the learning of desired capabilities. We also show that prompts which more strongly elicit the undesired behavior prior to fine-tuning more effectively inoculate against the behavior when used during training; this serves as a heuristic to identify promising inoculation prompts. Overall, IP is a simple yet effective way to control how models generalize from fine-tuning, preventing learning of undesired behaviors without substantially disrupting desired capabilities.

cs.LG↗

arXiv2024

Generative AI for Overall Mission Effectiveness at the Habitable Worlds Observatory

Here we present several use cases for using Generative AI (Gen AI) to improve systems engineering and cognitive knowledge management related to the future of astronomy from a culmination of working meetings and presentations as part of the Gen AI Task Group for the NASA Habitable Worlds Observatory (HWO) Science and Technology Architecture Review Team (START) AI/ML Working Group. Collectively, our group mission statement is "Where is the Human-in-the-loop as Gen AI systems become more powerful and autonomous?" with an emphasis on the ethical applications of Gen AI, guided by using these systems to remove drudgery from human work while simultaneously increasing opportunities for humans to experience more collective creativity and innovation. The HWO mission stands to benefit dramatically from generative models for different data types including text, time series/spectra, and image data. These cover a wide range of applications in science and engineering for HWO, including: mission development acceleration, data analysis and interpretation, enhancing imaging capabilities, anomaly detection, predictive modeling and simulation, data augmentation for machine learning, instrument calibration and optimization, public engagement and education, and assisting in mission planning. As an example, through sensitivity analysis of simulated exoplanet population science data sets of various generative model complexity, we can reverse engineer the measurement uncertainty requirements for HWO instruments to produce data that can constrain population models and thus inform HWO design requirements. This approach to HWO design is one example of a strategy that can ensure that HWO remains AI-ready. Through presenting herein a combination of visionary ideas balanced with grounded validated use case examples, we aim to support the development of a long-term strategy to keep HWO AI-ready as it moves forward.

astro-ph.IM↗

arXiv2024

Disentangling Dense Embeddings with Sparse Autoencoders

Sparse autoencoders (SAEs) have shown promise in extracting interpretable features from complex neural networks. We present one of the first applications of SAEs to dense text embeddings from large language models, demonstrating their effectiveness in disentangling semantic concepts. By training SAEs on embeddings of over 420,000 scientific paper abstracts from computer science and astronomy, we show that the resulting sparse representations maintain semantic fidelity while offering interpretability. We analyse these learned features, exploring their behaviour across different model capacities and introducing a novel method for identifying ``feature families'' that represent related concepts at varying levels of abstraction. To demonstrate the practical utility of our approach, we show how these interpretable features can be used to precisely steer semantic search, allowing for fine-grained control over query semantics. This work bridges the gap between the semantic richness of dense embeddings and the interpretability of sparse representations. We open source our embeddings, trained sparse autoencoders, and interpreted features, as well as a web app for exploring them.

cs.LG↗

arXiv2024

pathfinder: A Semantic Framework for Literature Review and Knowledge Discovery in Astronomy

The exponential growth of astronomical literature poses significant challenges for researchers navigating and synthesizing general insights or even domain-specific knowledge. We present Pathfinder, a machine learning framework designed to enable literature review and knowledge discovery in astronomy, focusing on semantic searching with natural language instead of syntactic searches with keywords. Utilizing state-of-the-art large language models (LLMs) and a corpus of 350,000 peer-reviewed papers from the Astrophysics Data System (ADS), Pathfinder offers an innovative approach to scientific inquiry and literature exploration. Our framework couples advanced retrieval techniques with LLM-based synthesis to search astronomical literature by semantic context as a complement to currently existing methods that use keywords or citation graphs. It addresses complexities of jargon, named entities, and temporal aspects through time-based and citation-based weighting schemes. We demonstrate the tool's versatility through case studies, showcasing its application in various research scenarios. The system's performance is evaluated using custom benchmarks, including single-paper and multi-paper tasks. Beyond literature review, Pathfinder offers unique capabilities for reformatting answers in ways that are accessible to various audiences (e.g. in a different language or as simplified text), visualizing research landscapes, and tracking the impact of observatories and methodologies. This tool represents a significant advancement in applying AI to astronomical research, aiding researchers at all career stages in navigating modern astronomy literature.

astro-ph.IM↗

arXiv2024

Designing an Evaluation Framework for Large Language Models in Astronomy Research

Large Language Models (LLMs) are shifting how scientific research is done. It is imperative to understand how researchers interact with these models and how scientific sub-communities like astronomy might benefit from them. However, there is currently no standard for evaluating the use of LLMs in astronomy. Therefore, we present the experimental design for an evaluation study on how astronomy researchers interact with LLMs. We deploy a Slack chatbot that can answer queries from users via Retrieval-Augmented Generation (RAG); these responses are grounded in astronomy papers from arXiv. We record and anonymize user questions and chatbot answers, user upvotes and downvotes to LLM responses, user feedback to the LLM, and retrieved documents and similarity scores with the query. Our data collection method will enable future dynamic evaluations of LLM tools for astronomy.

astro-ph.IM↗

arXiv2024

Searching for Bumps in the Cosmological Road: Do Type Ia Supernovae with Early Excesses Have Biased Hubble Residuals?

Flux excesses in the early time light curves of Type Ia supernovae (SNe\,Ia) are predicted by multiple theoretical models and have been observed in a number of nearby SNe\,Ia over the last decade. However, the astrophysical processes that cause these excesses may affect their use as standardizable candles for cosmological parameter measurements. In this paper, we perform a systematic search for early-time excesses in SNe\,Ia observed by the Zwicky Transient Facility (ZTF) to study whether SNe\,Ia with these excesses yield systematically different Hubble residuals. We analyze two compilations of ZTF SN\,Ia light curves from its first year of operations: 127 high-cadence light curves from \citet{Yao19} and 305 light curves from the ZTF cosmology data release of \citet{Dhawan22}. We detect significant early-time excesses for 17 SNe\,Ia in these samples and find that the excesses have an average $g-r$ color of $0.06\pm0.09$~mag; we do not find a clear preference for blue excesses as predicted by several models. Using the SALT3 model, we measure Hubble residuals for these two samples and find that excess-having SNe\,Ia may have lower Hubble residuals (HR) after correcting for shape, color, and host-galaxy mass, at $\sim$2-3$σ$ significance; our baseline result is $ΔHR = -0.056 \pm 0.026$~mag ($2.2 σ$). We compare the host-galaxy masses of excess-having and no-excess SNe\,Ia and find they are consistent, though at marginal significance excess-having SNe\,Ia may prefer lower-mass hosts. Additional discoveries of early excess SNe\,Ia will be a powerful way to understand potential biases in SN\,Ia cosmology and probe the physics of SN\,Ia progenitors.

astro-ph.HE↗

arXiv2023

SN2023ixf in Messier 101: the twilight years of the progenitor as seen by Pan-STARRS

The nearby type II supernova, SN2023ixf in M101 exhibits signatures of early-time interaction with circumstellar material in the first week post-explosion. This material may be the consequence of prior mass loss suffered by the progenitor which possibly manifested in the form of a detectable pre-supernova outburst. We present an analysis of the long-baseline pre-explosion photometric data in $g$, $w$, $r$, $i$, $z$ and $y$ filters from Pan-STARRS as part of the Young Supernova Experiment, spanning $\sim$5,000 days. We find no significant detections in the Pan-STARRS pre-explosion light curve. We train a multilayer perceptron neural network to classify pre-supernova outbursts. We find no evidence of eruptive pre-supernova activity to a limiting absolute magnitude of $-7$. The limiting magnitudes from the full set of $gwrizy$ (average absolute magnitude $\approx$-8) data are consistent with previous pre-explosion studies. We use deep photometry from the literature to constrain the progenitor of SN2023ixf, finding that these data are consistent with a dusty red supergiant (RSG) progenitor with luminosity $\log\left(L/L_\odot\right)$$\approx$5.12 and temperature $\approx$3950K, corresponding to a mass of 14-20 M$_\odot$

astro-ph.SR↗

arXiv2023

The role of natal kicks in forming asymmetric compact binary mergers

In their most recent observing run, the LIGO-Virgo-KAGRA (LVK) Collaboration observed gravitational waves (GWs) from compact binary mergers with highly asymmetric mass ratios, including both binary black holes (BBHs) and neutron star--black holes (NSBHs). It appears that NSBHs with mass ratios $q\simeq0.2$ are more common than equally asymmetric BBHs, but the reason for this remains unclear. We use the binary population synthesis code COSMIC to investigate the evolutionary pathways leading to the formation and merger of asymmetric compact binaries. We find that within the context of isolated binary stellar evolution, most asymmetric mergers start off as asymmetric stellar binaries. Because of the initial asymmetry, these systems tend to first undergo a dynamically unstable mass transfer phase. However, after the first star collapses to a compact object, the mass ratio is close to unity and the second phase of mass transfer is usually stable. According to our simulations, this stable mass transfer fails to shrink the orbit enough on its own for the system to merge. Instead, the natal kick received by the second-born compact object during its collapse is key in determining how many of these systems can merge. For the most asymmetric systems with mass ratios $q\leq0.1$, the merging systems in our models receive an average kick magnitude of 255 km s$^{-1}$ during the second collapse, while the average kick for non-merging systems is 59 km s$^{-1}$. Because lower mass compact objects, like NSs, are expected to receive larger natal kicks than higher mass BHs, this may explain why asymmetric NSBH systems merge more frequently than asymmetric BBH systems.

astro-ph.HE↗

arXiv2022

Inferring the neutron star maximum mass and lower mass gap in neutron star-black hole systems with spin

Gravitational-wave (GW) detections of merging neutron star-black hole (NSBH) systems probe astrophysical neutron star (NS) and black hole (BH) mass distributions, especially at the transition between NS and BH masses. Of particular interest are the maximum NS mass, minimum BH mass, and potential mass gap between them. While previous GW population analyses assumed all NSs obey the same maximum mass, if rapidly spinning NSs exist, they can extend to larger maximum masses than nonspinning NSs. In fact, several authors have proposed that the $\sim2.6\,M_\odot$ object in the event GW190814 -- either the most massive NS or least massive BH observed to date -- is a rapidly spinning NS. We therefore infer the NSBH mass distribution jointly with the NS spin distribution, modeling the NS maximum mass as a function of spin. Using 4 LIGO-Virgo NSBH events including GW190814, if we assume that the NS spin distribution is uniformly distributed up to the maximum (breakup) spin, we infer the maximum non-spinning NS mass is $2.7^{+0.5}_{-0.4}\,M_\odot$ (90\% credibility), while assuming only nonspinning NSs, the NS maximum mass must be $>2.53 M_\odot$ (90\% credibility). The data support the mass gap's existence, with a minimum BH mass at $5.4^{+0.7}_{-1.0} M_\odot$. With future observations, under simplified assumptions, 150 NSBH events may constrain the maximum nonspinning NS mass to $\pm0.02\,M_\odot$, and we may even measure the relation between the NS spin and maximum mass entirely from GW data. If rapidly rotating NSs exist, their spins and masses must be modeled simultaneously to avoid biasing the NS maximum mass.

astro-ph.HE↗

arXiv2021

Cosmology with Standard Sirens at Cosmic Noon

Gravitational waves (GWs) directly measure the luminosity distance to the merger, which, when combined with an independent measurement of the source's redshift, provides a novel probe of cosmology. The proposed next generation of ground-based GW detectors, Einstein Telescope and Cosmic Explorer, will detect tens of thousands of binary neutron stars (BNSs) out to cosmological distances ($z>2$), beyond the peak of the star formation rate (SFR), or "cosmic noon." At these distances, it will be challenging to measure the sources' redshifts by observing electromagnetic (EM) counterparts or statistically marginalizing over a galaxy catalog. In the absence of an EM counterpart or galaxy catalog, Ding et al. showed that theoretical priors on the merger redshift distribution can be used to infer parameters in a $w$CDM cosmology. We argue that in the BNS case, the redshift distribution will be measured by independent observations of short gamma ray bursts (GRBs), kilonovae, and known BNS host galaxies. We show that, in addition to measuring the background cosmology, this method can constrain the effects of dark energy on modified GW propagation. We consider the simple case in which the BNS rate is \textit{a priori} known to follow the SFR. If the SFR is perfectly known, $\mathcal{O}(10,000)$ events (to be expected within a year of observation with Cosmic Explorer) would yield a sub-tenth percent measurement of the combination $H_0^{2.8}Ω_M$. Fixing $H_0$ and $Ω_M$, this method may enable a 5\% measurement of the dark energy equation of state parameter. Fixing the background cosmology and probing modified GW propagation, the running of the Planck mass parameter $c_M$ may be measured to $\pm0.02$. Although realistically, the redshift evolution of the merger rate will be uncertain, prior knowledge of the peak redshift will provide valuable information for standard siren analyses.

astro-ph.CO↗