Publications
Our teams aspire to make discoveries that impact everyone, and core to our approach is sharing our research and tools to fuel progress in the field.
Our teams aspire to make discoveries that impact everyone, and core to our approach is sharing our research and tools to fuel progress in the field.
Sort By
1 - 15 of 10795 publications
Preview abstract
AI coding assistants are rapidly becoming integral to modern software development. A key challenge in this space is the continual need to migrate and modernize codebases in response to evolving software ecosystems. Traditionally, such migrations have relied on rule-based systems and human intervention. With the advent of powerful large language models (LLMs), AI-driven agentic frameworks offer a promising alternative—but their effectiveness remains underexplored. In this paper, we introduce FreshBrew, a novel benchmark for evaluating AI-based agentic frameworks on project-level Java migrations. We benchmark several such frameworks, powered by state-of-the-art LLMs, and compare their performance against established rule-based tools. Our evaluation of AI agents on this benchmark of 228 repositories shows that the top-performing model, Gemini 2.5 Flash, can successfully migrate 56.5% of projects to JDK 17. Our empirical analysis reveals novel insights into the critical strengths and limitations of current agentic approaches, offering actionable insights into their real-world applicability. By releasing FreshBrew publicly upon acceptance, we aim to facilitate rigorous, reproducible evaluation and catalyze progress in AI-driven codebase modernization.
View details
Preview abstract
For many practical applications of quantum computing, the slowest and most costly steps involve coherently accessing classical data. We help address this challenge by applying mass production techniques, which can sometimes allow us to perform operations many times in parallel for a cost that is comparable to a single execution[1-3]. We combine existing mass-production results with modern approaches for loading classical data using ``quantum read-only memory.'' We show that quantum mass production techniques offer no benefit when we consider a cost model that focuses purely on the number of non-Clifford gates. However, analyzing the constant factors in a more nuanced cost model, we find that it may be possible to obtain a reduction in cost of an order or magnitude or more for a variety reasonably-sized fault-tolerant quantum algorithms. We present several applications of quantum mass-production techniques beyond naive parallelization, including a strategy for reducing the cost of serial calls to the same data loading step.
View details
Scaling Laws for Downstream Task Performance in Machine Translation
Natalia Ponomareva
Hussein Hazimeh
Sanmi Koyejo
International Conference on Learning Representations (ICLR) (2025) (to appear)
Preview abstract
Scaling laws provide important insights that can guide the design of large language models (LLMs). Existing work has primarily focused on studying scaling laws for pretraining (upstream) loss. However, in transfer learning settings, in which LLMs are pretrained on an unsupervised dataset and then finetuned on a downstream task, we often also care about the downstream performance. In this work, we study the scaling behavior in a transfer learning setting, where LLMs are finetuned for machine translation tasks. Specifically, we investigate how the choice of the \emph{pretraining} data and its size affect downstream performance (translation quality) as judged by: downstream cross-entropy and translation quality metrics such as BLEU and COMET scores. Our experiments indicate that the size of the finetuning dataset and the distribution alignment between the pretraining and downstream data significantly influence the scaling behavior. With sufficient alignment, both downstream cross-entropy and translation quality scores improve monotonically with more pretraining data. In such cases, we show that it is possible to predict the downstream translation quality metrics with good accuracy using a log-law. However, there are cases where moderate misalignment causes the downstream translation scores to fluctuate or get worse with more pretraining, whereas downstream cross-entropy monotonically improves. By analyzing these, we provide new practical insights for choosing appropriate pretraining data.
View details
Circadian rhythm of heart rate and activity: a cross-sectional study
Maryam Khalid
Logan Schneider
Aravind Natarajan
Conor Heneghan
Karla Gleichauf
Chronobiology International (2025)
Preview abstract
ABSTRACT
Background: Circadian rhythms are commonly observed in a number of physiological processes. Consumer wearable devices have made it possible to obtain continuous time series data from a large number of individuals. We study circadian rhythms from measurements of heart rate, movement, and sleep, from a cohort of nearly 20,000 participants over the course of 30 days.
Methods: Participation was restricted to Fitbit users of age 21 years or older residing in the United States or Canada. Participants were enrolled through a recruitment banner shown on the Fitbit App. The advertisement was shown to 531,359 Fitbit users, and 23,239 enrolled in the program. Of these, we obtained heart rate data from 19,350 participants. We obtain the underlying circadian rhythm from time series heart rate by modeling the circadian rhythm as a sum over the first two Fourier harmonics. The first Fourier harmonic accounts for the 24-hour rhythmicity, while the second harmonic accounts for non-sinusoidal perturbations.
Findings: We observe a circadian rhythm in both heart rate and acceleration. From the diurnal modulation, we obtain the following circadian parameters: (i) amplitude of modulation, (ii) bathyphase, (iii) acrophase, (iv) non-sinusoidal fraction, and (v) fraction of day when the heart rate is greater than the mean. The amplitude, bathyphase, and acrophase depend on sex, and decrease with age. The waketime on average, follows the bathyphase by 2.4 hours. In most individuals, the circadian rhythm of heart rate lags the circadian rhythm of activity.
Interpretation: Circadian metrics for heart rate and activity can be reliably obtained from commercially available wearable devices. Distributions of circadian metrics can be valuable tools for individual-level interpretation.
View details
Preview abstract
The global adoption of Large Language Models (LLMs) in healthcare shows promise for enhancing clinical workflows and improving patient outcomes. However, Automatic Speech
Recognition (ASR) errors in critical medical entities remain a significant challenge. These
errors can lead to severe consequences if undetected. This study investigates the prevalence and impact of ASR errors in medical transcription across Africa, Europe, and North America. By examining variations in accented English across three continents, we analyze the impact of regional speech patterns on ASR performance. Our research quantifies both the potential and limitations of LLMs in mitigating ASR inaccuracies within various medical settings, with particular attention to performance variations across regional accents and medical terminology. Our findings highlight significant disparities in ASR accuracy across regions and identify specific conditions under which LLM corrections prove most effective.
View details
Depth-Width tradeoffs in Algorithmic Reasoning of Graph Tasks with Transformers
Gilad Yehudai
Maya Bechler-Speicher
Orr Fischer
Ran Gilad-Bachrach
2025
Preview abstract
In particular, they can be used to solve complex algorithmic problems, including graph-based tasks. In such algorithmic tasks a key question is what is the minimal size of a transformer that can implement a task. Recent work has begun to explore this problem for graph-based tasks, showing that for sub-linear embedding dimension (i.e., model width) logarithmic depth suffices. However, an open question, which we address here, is what happens if width is allowed to grow linearly. Here we analyze this setting, and provide the surprising result that with linear width, constant depth suffices for solving a host of graph-based problems. This suggests that a moderate increase in width can allow much shallower models, which are advantageous in terms of inference time. For other problems, we show that quadratic width is required. Our results demonstrate the complex and intriguing landscape of transformer implementations of graph-based algorithms. We support our theoretical results with empirical evaluations.
View details
Preview abstract
This paper investigates the theoretical underpinnings of the widely successful
pretrain-then-adapt strategy for foundation models. We introduce a Bayesian
model selection criterion, termed the downstream free energy, which quantifies
the adaptability of a pretrained checkpoint by measuring, under the downstream
data distribution, the concentration of favorable solutions near the checkpoint.
However, minimizing this downstream free energy is infeasible without access to
downstream data. To address this, we show that under certain conditions, mini-
mizing the upstream free energy – which can be estimated using only upstream
data – can serve as a reliable proxy. We validate this theoretical insight through
preliminary experiments, showing that commonly used pretraining heuristics ef-
fectively lower upstream free energy, leading to better downstream performance.
View details
How Expressive are Knowledge Graph Foundation Models?
Xingyue Huang
Pablo Barcelo
Michael Bronstein
Ismail Ilkan Ceylan
Michael Galkin
Juan Reutter
Miguel Romero Orth
ICML 2025
Preview abstract
Knowledge Graph Foundation Models (KGFMs) are at the frontier for deep learning on knowledge graphs (KGs), as they can generalize to completely novel knowledge graphs with different relational vocabularies. Despite their empirical success, our theoretical understanding of KGFMs remains very limited. In this paper, we conduct a rigorous study of the expressive power of KGFMs. Specifically, we show that the expressive power of KGFMs directly depends on the motifs that are used to learn the relation representations. We then observe that the most typical motifs used in the existing literature are binary, as the representations are learned based on how pairs of relations interact, which limits the model's expressiveness. As part of our study, we design more expressive KGFMs using richer motifs, which necessitate learning relation representations based on, e.g., how triples of relations interact with each other. Finally, we empirically validate our theoretical findings, showing that the use of richer motifs results in better performance on a wide range of datasets drawn from different domains.
View details
The Role of Outgoing Connection Heterogeneity in Feedforward Layers of Large Language Models
Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Suzhou, China (to appear)
Preview abstract
We report on investigations into the characteristics of outgoing connections in feedforward layers of large language models. Our findings show that inner neurons with diverse outgoing connection strengths contribute more significantly to model performance than those with uniform connections. We propose a new fine-tuning loss that takes advantage of this observation by decreasing the outgoing connection entropy in feedforward layers. Using this loss yields gains over standard fine-tuning across two different model families (PaLM-2 and Gemma-2) for downstream tasks in math, coding, and language understanding. To further elucidate the role of outgoing connection heterogeneity, we develop a data-free structured pruning method, which uses entropy to identify and remove neurons. This method significantly surpasses the effectiveness of random and even magnitude-based pruning.
View details
Enhancing Performance of the Tesseract Decoder for Quantum Error Correction
DRAGANA GRBIC
Laleh Beni
Noah Shutty
2025
Preview abstract
In this paper I describe the performance enchantments I implemented in a quantum-error-correction decoder developed at Google. The decoder is an open-source project and I am documenting the speedups I achieved in this paper.
View details
Mufu: Multilingual Fused Learning for Low- Resource Translation with LLM
Zheng Lim
Honglin Yu
Trevor Cohn
International Conference on Learning Representations (ICLR) 2025
Preview abstract
Multilingual large language models (LLMs) are great translators, but this is largely limited to high-resource languages. For many LLMs, translating in and out of low-resource languages remains a challenging task. To maximize data efficiency in this low-resource setting, we introduce Mufu, which includes a selection of automatically generated multilingual candidates and an instruction to correct inaccurate translations in the prompt. Mufu prompts turn a translation task into a postediting one, and seek to harness the LLM's reasoning capability with auxiliary translation candidates, from which the model is required to assess the input quality, align the semantics cross-lingually, copy from relevant inputs and override instances that are incorrect. Our experiments on En-XX translations over the Flores-200 dataset show LLMs finetuned against Mufu-style prompts are robust to poor quality auxiliary translation candidates, achieving performance superior to NLLB 1.3B distilled model in 64% of low- and very-low-resource language pairs. We then distill these models to reduce inference cost, while maintaining on average 3.1 chrF improvement over finetune-only baseline in low-resource translations.
View details
Preview abstract
In Julia, JuMP is the go-to modelling package for mathematical optimisation. As of this writing, Google's award-winning solvers have not been accessible through JuMP; which offers Julia's ease of use. ORTools.jl is changing this. Julia users will now have access to Google's Glop, CP-SAT, and PDLP solvers through JuMP as provided by the ORTools.jl package.
This talk offers an introduction to the features of the package and an overview of the difficulties we encountered.
View details
Preview abstract
In modern datasets, where single records can have multiple owners, enforcing user-level differential privacy requires capping each user's total contribution. This "contribution bounding" becomes a significant combinatorial challenge. Existing sequential algorithms for this task are computationally intensive and do not scale to the massive datasets prevalent today. To address this scalability bottleneck, we propose a novel and efficient distributed algorithm. Our approach models the complex ownership structure as a hypergraph, where users are vertices and records are hyperedges. The algorithm proceeds in rounds, allowing users to propose records in parallel. A record is added to the final dataset only if all its owners unanimously agree, thereby ensuring that no user's predefined contribution limit is violated. This method aims to maximize the size of the resulting dataset for high utility while providing a practical, scalable solution for implementing user-level privacy in large, real-world systems.
View details
Preview abstract
Many AI applications of interest require specialized multi-modal models. Yet, relevant data for training these models is inherently scarce. Human annotation is prohibitively expensive, error-prone, and time-consuming. Meanwhile, existing synthetic data generation methods often rely on manual prompts, evolutionary algorithms, or extensive seed data from the target distribution - limiting scalability and control. In this paper, we introduce Simula, a novel, seedless framework that balances global and local reasoning to generate synthetic datasets. We utilize taxonomies to capture a global coverage space and use a series of agentic refinements to promote local diversity and complexity. Our approach allows users to define desired dataset characteristics through an explainable and controllable process, without relying on seed data. This unlocks new opportunities for developing and deploying AI in domains where data scarcity or privacy concerns are paramount.
View details
Performance of a Deep Learning Diabetic Retinopathy Algorithm in India
Arthur Brant
Xiang Yin
Lu Yang
Divleen Jeji
Sunny Virmani
Anchintha Meenu
Naresh Babu Kannan
Florence Thng
Lily Peng
Ramasamy Kim
JAMA Network Open (2025)
Preview abstract
Importance: While prospective studies have investigated the accuracy of artificial intelligence (AI) for detection of diabetic retinopathy (DR) and diabetic macular edema (DME), to date, little published data exist on the clinical performance of these algorithms.
Objective: To evaluate the clinical performance of an automated retinal disease assessment (ARDA) algorithm in the postdeployment setting at Aravind Eye Hospital in India.
Design, Setting, and Participants: This cross-sectional analysis involved an approximate 1% sample of fundus photographs from patients screened using ARDA. Images were graded via adjudication by US ophthalmologists for DR and DME, and ARDA’s output was compared against the adjudicated grades at 45 sites in Southern India. Patients were randomly selected between January 1, 2019, and July 31, 2023.
Main Outcomes and Measures: Primary analyses were the sensitivity and specificity of ARDA for severe nonproliferative DR (NPDR) or proliferative DR (PDR). Secondary analyses focused on sensitivity and specificity for sight-threatening DR (STDR) (DME or severe NPDR or PDR).
Results: Among the 4537 patients with 4537 images with adjudicated grades, mean (SD) age was 55.2 (11.9) years and 2272 (50.1%) were male. Among the 3941 patients with gradable photographs, 683 (17.3%) had any DR, 146 (3.7%) had severe NPDR or PDR, 109 (2.8%) had PDR, and 398 (10.1%) had STDR. ARDA’s sensitivity and specificity for severe NPDR or PDR were 97.0% (95% CI, 92.6%-99.2%) and 96.4% (95% CI, 95.7%-97.0%), respectively. Positive predictive value (PPV) was 50.7% and negative predictive value (NPV) was 99.9%. The clinically important miss rate for severe NPDR or PDR was 0% (eg, some patients with severe NPDR or PDR were interpreted as having moderate DR and referred to clinic). ARDA’s sensitivity for STDR was 95.9% (95% CI, 93.0%-97.4%) and specificity was 94.9% (95% CI, 94.1%-95.7%); PPV and NPV were 67.9% and 99.5%, respectively.
Conclusions and Relevance: In this cross-sectional study investigating the clinical performance of ARDA, sensitivity and specificity for severe NPDR and PDR exceeded 96% and caught 100% of patients with severe NPDR and PDR for ophthalmology referral. This preliminary large-scale postmarketing report of the performance of ARDA after screening 600 000 patients in India underscores the importance of monitoring and publication an algorithm's clinical performance, consistent with recommendations by regulatory bodies.
View details
