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 10431 publications
Preview abstract
We present a scalable and agile approach for ads image content moderation at Google, addressing the challenges of moderating massive volumes of ads with diverse content and evolving policies. The proposed method utilizes human-curated textual descriptions and cross-modal text-image co-embeddings to enable zero-shot classification of policy violating ads images, bypassing the need for extensive supervised training data and human labeling. By leveraging large language models (LLMs) and user expertise, the system generates and refines a comprehensive set of textual descriptions representing policy guidelines. During inference, co-embedding similarity between incoming images and the textual descriptions serves as a reliable signal for policy violation detection, enabling efficient and adaptable ads content moderation. Evaluation results demonstrate the efficacy of this framework in significantly boosting the detection of policy violating content.
View details
Permission Rationales in the Web Ecosystem: An Exploration of Rationale Text and Design Patterns
Yusra Elbitar
Soheil Khodayari
Marian Harbach
Gianluca De Stefano
Balazs Engedy
Giancarlo Pellegrino
Sven Bugiel
CHI 2025, ACM
Preview abstract
Modern web applications rely on features like camera and geolocation for personalized experiences, requiring user permission via browser prompts. To explain these requests, applications provide rationales—contextual information on why permissions are needed. Despite their importance, little is known about how rationales appear on the web or their influence on user decisions.
This paper presents the first large-scale study of how the web ecosystem handles permission rationales, covering three areas: (i) identifying webpages that use permissions, (ii) detecting and classifying permission rationales, and (iii) analyzing their attributes to understand their impact on user decisions. We examined over 770K webpages from Chrome telemetry, finding 3.6K unique rationale texts and 749 rationale UIs across 85K pages. We extracted key rationale attributes and assessed their effect on user behavior by cross-referencing them with Chrome telemetry data. Our findings reveal nine key insights, providing the first evidence of how different rationales affect user decisions.
View details
Beyond Digital Literacy: Building Youth Digital Resilience Through Existing “Information Sensibility” Practices
Mia Hassoun
Ian Beacock
Todd Carmody
Patrick Gage Kelley
Beth Goldberg
Devika Kumar
Laura Murray
Rebekah Park
Behzad Sarmadi
Social Sciences Journal, 14(4) (2025)
Preview abstract
Youth media consumption and disordered eating practices have historically been subjects of moral panics, often resulting in protective, deficit-based interventions like content removal. We argue for interventions which instead equip youth to evaluate and manage risks in their online environments, building upon their existing “information sensibility” practices. Drawing upon ethnographic research and intervention testing with 77 participants in the US and India, we analyze how youth (aged 13–26), including those with diverse political perspectives and those recovering from disordered eating (DE), engage with online news and health information. Participants generally algorithmically encountered (rather than searched for) information online, and their engagement was shaped more by social motivations—like belonging—than truth seeking. Participants interpreted online information collaboratively, relying on social cues and peer validation within their online communities. They demonstrated preference for personal testimonies and relatable sources, particularly those with similar social identities. We propose resilience-building interventions that build upon these youth online information practices by: (1) leveraging peer networks, promoting critical information engagement through collaborative learning and peer-to-peer support within online communities; (2) developing social media sensibility, equipping youth to critically evaluate information sources in situ; (3) providing pathways offline, connecting youth to desired in-person communities; and (4) encouraging probabilistic thinking.
View details
Preview abstract
As large language models (LLMs) improve in their capacity to serve as personal AI assistants, their ability to output uniquely tailored, personalized responses that align with the soft preferences of their users is imperative for maximizing user satisfaction and retention. However, lay users are notoriously bad at prompt specification and often struggle with conveying their latent preferences to AI assistants. To resolve this, we demonstrate that activation steering, an inference-time method, can effectively control the response of the LLMs towards expressing different preferences. In contrast to memory-based personalization methods that require long user history, steering is extremely lightweight and easily-controllable via an interpretable linear strength factor. We further conduct a within-subjects user study (n=14) to investigate how end users personalize their conversations through three different steerable chatbot interfaces. The results demonstrate the effectiveness of preference-based steering for aligning real-world conversations with user preferences, and we discuss qualitative findings on how diverse values around control, transparency, and usability of personalization lead users to prefer different interfaces.
View details
Matryoshka Model Learning for Improved Elastic Student Models
Cho-Jui Hsieh
Chetan Verma
Aditya Srinivas Timmaraju
Inderjit Dhillon
Xin Liu
Wen Chen
Ngot Bui
Yang Zhang
2025
Preview abstract
Production machine learning models in the industry are often devel-oped with a primary focus on maximizing model quality. However,these models must ultimately operate within the resource con-straints of their serving infrastructure, including limitations on com-pute, memory and bandwidth. The rapid evolution of serving hard-ware, particularly with advancements in accelerator technology,necessitates periodic retraining to leverage newer, more efficientinfrastructure. This cyclical retraining process is resource-intensive,demanding significant model development time and incurring sub-stantial training costs. This challenge is further amplified by thetrend towards increasingly complex models, which inherently re-quire greater computational resources for training and deployment.While prior work has explored techniques like supernet sub-modelextraction to address training efficiency, a critical gap remains: theefficient generation of a spectrum of high-quality models froman existing production model, a common requirement in diverseindustrial applications. To bridge this gap, we introduce a novel ap-proach leveraging a "Teaching Assistant" (TA) model, derived froma given production model (referred to as the Student model). Wedemonstrate that through co-training the Student and TA modelswith Matryoshka structure while using online distillation, we notonly enhance the Student model’s performance but also enable theflexible creation of a model family offering a compelling trade-offbetween model quality and model size.
View details
Preview abstract
Despite the advent of legislation such as the General Data Protection Regulation (GDPR) with its associated "Right to be Forgotten" (RTBF), few, if any, studies have measured user reactions to realistic edge cases with public-interest content. Surveying both users covered by and excluded from RTBF, this vignette-based survey experiment sought to better understand how users think of delisting content from search engine results and what factors influence user perceptions. While leaving information accessible in search engine results generally leads to warmer feelings towards those search engines than delisting it, we find that users do prefer different outcomes
depending on contextual elements specific to given cases. We also find that whether a country has active RTBF legislation does seem to be associated with both knowledge and attitudes about RTBF, but is unlikely to explain all of it. These results indicate a complex context around removing public-interest content from search engines’ results; it is essential that experts sensitive to local context perform the review in order to ensure that removal requests are handled in a way that meets users’ expectations.
View details
Validation of Quantum Elliptic Curve Point Addition Circuits
(2025) (to appear)
Preview abstract
Specific quantum algorithms exist to—in theory—
break elliptic curve cryptographic protocols. Implementing these
algorithms requires designing quantum circuits that perform elliptic curve arithmetic. To accurately judge a cryptographic protocol’s resistance against future quantum computers, researchers
figure out minimal resource-count circuits for performing these
operations while still being correct. To assure the correctness of
a circuit, it is integral to restore all ancilla qubits used to their
original states. Failure to do so could result in decoherence of the
computation’s final result. Through rigorous classical simulation
and unit testing, I surfaced four inconsistencies in the state-ofthe-art quantum circuit for elliptic curve point addition where
the circuit diagram states the qubits are returned in the original
(|0⟩) state, but the intermediate values are not uncomputed. I
provide fixes to the circuit without increasing the leading-order
gate cost.
View details
Preview abstract
Augmenting LLMs with context leads to improved performance across many applications. Despite much research on Retrieval Augmented Generation (RAG) systems, an open question is whether errors arise because LLMs fail to utilize the context from retrieval or the context itself is insufficient to answer the query. To shed light on this, we develop a new notion of sufficient context, along with a way to classify instances that have enough information to answer the query. We then use sufficient context to analyze several models and datasets. By stratifying errors based on context sufficiency, we find that proprietary LLMs (Gemini, GPT, Claude) excel at answering queries when the context is sufficient, but often output incorrect answers instead of abstaining when the context is not. On the other hand, open-source LLMs (Llama, Mistral, Gemma) hallucinate or abstain often, even with sufficient context. We further categorize cases when the context is useful, and improves accuracy, even though it does not fully answer the query and the model errs without the context. Building on our findings, we explore ways to reduce hallucinations in RAG systems, including a new selective generation method that leverages sufficient context information for guided abstention. Our method improves the fraction of correct answers among times where the model responds by 2--10% for Gemini, GPT, and Gemma.
View details
Unprecedented Insights into Maternal Sleep: A Large-scale Longitudinal Analysis of Real-world Wearable Device Data Before, During, and After Pregnancy
Nichole Young-Lin
Conor Heneghan
Logan Schneider
Logan Niehaus
Ariel Haney
Karla Gleichauf
Jacqueline Shreibati
Belen Lafon
Lancet eBioMedicine (2025)
Preview abstract
Introduction: Current understanding of pregnancy and postpartum sleep is driven by limited lab or self-reported data. Consumer wearable devices may help reveal longitudinal, real-world sleep patterns.
Methods: We analyzed de-identified wearable device data from 2,540 users in the United States and Canada who met strict wear-time requirements (≥80% daily usage for ≥80% of the time periods of interest [12 weeks prepregnancy, throughout pregnancy, and 20 weeks immediately postpartum]). We tracked sleep time and staging using Fitbit devices.
Results: Compared to prepregnancy, total sleep time (TST) increased from an average of 425.3±43.5 min to a peak of 447.6±47.6 min at gestational week 10 with ongoing declines throughout pregnancy. Time in bed (TIB) followed a similar pattern. Increased light sleep drove the initial TST rise. Deep and REM sleep decreased significantly throughout pregnancy, with maximum reductions of 19.2±13.8 min (p<0.01) and 9.0±19.2 min (p<0.01) respectively by pregnancy end. Sleep efficiency also declined slightly during pregnancy (median drop from 88.3% to 86.8%). After delivery, TIB remained below the prepregnancy baseline by 14.7±45.7 min at one year postpartum and 15.2±47.7 min at 1.5 years postpartum.
Conclusion: This unprecedented look at large-scale, real-world sleep and pregnancy patterns revealed a previously unquantified initial increase in sleep followed by decreases in both quantity and quality as pregnancy progresses. Sleep deficits persist for at least 1.5 years postpartum. These quantified trends can assist clinicians and patients in understanding what to expect.
View details
Preview abstract
Given copies of a quantum state $\rho$, a shadow tomography protocol aims to learn all expectation values from a fixed set of observables, to within a given precision $\epsilon$. We say that a shadow tomography protocol is \textit{triply efficient} if it is sample- and time-efficient, and only employs measurements that entangle a constant number of copies of $\rho$ at a time. The classical shadows protocol based on random single-copy measurements is triply efficient for the set of local Pauli observables. This and other protocols based on random single-copy Clifford measurements can be understood as arising from fractional colorings of a graph $G$ that encodes the commutation structure of the set of observables. Here we describe a framework for two-copy shadow tomography that uses an initial round of Bell measurements to reduce to a fractional coloring problem in an induced subgraph of $G$ with bounded clique number. This coloring problem can be addressed using techniques from graph theory known as \textit{chi-boundedness}. Using this framework we give the first triply efficient shadow tomography scheme for the set of local fermionic observables, which arise in a broad class of interacting fermionic systems in physics and chemistry. We also give a triply efficient scheme for the set of all $n$-qubit Pauli observables. Our protocols for these tasks use two-copy measurements, which is necessary: sample-efficient schemes are provably impossible using only single-copy measurements. Finally, we give a shadow tomography protocol that compresses an $n$-qubit quantum state into a $\poly(n)$-sized classical representation, from which one can extract the expected value of any of the $4^n$ Pauli observables in $\poly(n)$ time, up to a small constant error.
View details
Data Quality Issues in Multilingual Speech Datasets: The Need for Sociolinguistic Awareness and Proactive Language Planning
Preview
Mingfei Lau
Allen Chen
Yeming Fang
Tingting Xu
Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics (ACL), Vienna, Austria (2025), 7466–7492
Validation of a Deep Learning Model for Diabetic Retinopathy on Patients with Young-Onset Diabetes
Tony Tan-Torres
Pradeep Praveen
Divleen Jeji
Arthur Brant
Xiang Yin
Lu Yang
Tayyeba Ali
Ilana Traynis
Dushyantsinh Jadeja
Rajroshan Sawhney
Sunny Virmani
Pradeep Venkatesh
Nikhil Tandon
Ophthalmology and Therapy (2025)
Preview abstract
Introduction
While many deep learning systems (DLSs) for diabetic retinopathy (DR) have been developed and validated on cohorts with an average age of 50s or older, fewer studies have examined younger individuals. This study aimed to understand DLS performance for younger individuals, who tend to display anatomic differences, such as prominent retinal sheen. This sheen can be mistaken for exudates or cotton wool spots, and potentially confound DLSs.
Methods
This was a prospective cross-sectional cohort study in a “Diabetes of young” clinic in India, enrolling 321 individuals between ages 18 and 45 (98.8% with type 1 diabetes). Participants had fundus photographs taken and the photos were adjudicated by experienced graders to obtain reference DR grades. We defined a younger cohort (age 18–25) and an older cohort (age 26–45) and examined differences in DLS performance between the two cohorts. The main outcome measures were sensitivity and specificity for DR.
Results
Eye-level sensitivity for moderate-or-worse DR was 97.6% [95% confidence interval (CI) 91.2, 98.2] for the younger cohort and 94.0% [88.8, 98.1] for the older cohort (p = 0.418 for difference). The specificity for moderate-or-worse DR significantly differed between the younger and older cohorts, 97.9% [95.9, 99.3] and 92.1% [87.6, 96.0], respectively (p = 0.008). Similar trends were observed for diabetic macular edema (DME); sensitivity was 79.0% [57.9, 93.6] for the younger cohort and 77.5% [60.8, 90.6] for the older cohort (p = 0.893), whereas specificity was 97.0% [94.5, 99.0] and 92.0% [88.2, 95.5] (p = 0.018). Retinal sheen presence (94% of images) was associated with DME presence (p < 0.0001). Image review suggested that sheen presence confounded reference DME status, increasing noise in the labels and depressing measured sensitivity. The gradability rate for both DR and DME was near-perfect (99% for both).
Conclusion
DLS-based DR screening performed well in younger individuals aged 18–25, with comparable sensitivity and higher specificity compared to individuals aged 26–45. Sheen presence in this cohort made identification of DME difficult for graders and depressed measured DLS sensitivity; additional studies incorporating optical coherence tomography may improve accuracy of measuring DLS DME sensitivity.
View details
A Strategic Framework for AI Product Development and Evaluation in Enterprise Software
International Journal of Computer Engineering and Technology (IJCET), Volume 16, Issue 1 (2025)
Preview abstract
This article presents a comprehensive framework for developing and evaluating AI products in enterprise software systems, addressing the critical challenges organizations face during AI transformation initiatives. The article introduces a structured approach to decision-making for AI integration, encompassing ROI evaluation, user value assessment, and business impact analysis. It establishes distinct methodologies for both assistive and autonomous AI systems, providing detailed metrics for measuring success and performance across different implementation scenarios. Across various industries, the framework has shown potential in reducing implementation time, increasing user adoption rates, and enhancing overall project success rates, highlighting its practical applicability. The article methodology combines theoretical analysis with practical case studies, resulting in a flexible yet robust framework that can adapt to various organizational contexts. The framework's primary contribution lies in its practical approach to bridging the gap between theoretical AI capabilities and real-world implementation challenges, offering product leaders a systematic methodology for AI product development and evaluation. By addressing both current implementation challenges and future scalability requirements, this framework provides organizations with a foundational tool for navigating their AI transformation journey while maintaining a focus on measurable business outcomes and user value creation.
View details
Astute RAG: Overcoming Imperfect Retrieval Augmentation and Knowledge Conflicts for Large Language Models
Fei Wang
The Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025) (2025) (to appear)
Preview abstract
Retrieval-Augmented Generation (RAG), while effective in integrating external knowledge to address the limitations of large language models (LLMs), can be undermined by imperfect retrieval, which may introduce irrelevant, misleading, or even malicious information. Despite its importance, previous studies have rarely explored the behavior of RAG through joint analysis on how errors from imperfect retrieval attribute and propagate, and how potential conflicts arise between the LLMs' internal knowledge and external sources. We find that imperfect retrieval augmentation might be inevitable and quite harmful, through controlled analysis under realistic conditions. We identify the knowledge conflicts between LLM-internal and external knowledge from retrieval as a bottleneck to overcome in the post-retrieval stage of RAG. To render LLMs resilient to imperfect retrieval, we propose Astute RAG, a novel RAG approach that adaptively elicits essential information from LLMs' internal knowledge, iteratively consolidates internal and external knowledge with source-awareness, and finalizes the answer according to information reliability. Our experiments using Gemini and Claude demonstrate that Astute RAG significantly outperforms previous robustness-enhanced RAG methods. Notably, Astute RAG is the only approach that matches or exceeds the performance of LLMs without RAG under worst-case scenarios. Further analysis reveals that Astute RAG effectively resolves knowledge conflicts, improving the reliability and trustworthiness of RAG systems.
View details
Reasoning-SQL: Reinforcement Learning with Partial Rewards for Reasoning-Enhanced Text-to-SQL
Mohammadreza Pourreza
Shayan Talaei
Hailong Li
Azalia Mirhoseini
Amin Saberi
Conference on Language Modeling (COLM) (2025) (to appear)
Preview abstract
Text-to-SQL is a challenging task involving multiple reasoning-intensive subtasks, including natural language understanding, database schema comprehension, and precise SQL query formulation. Existing approaches often rely on handcrafted reasoning paths with inductive biases that can limit their overall effectiveness. Motivated by the recent success of reasoning-enhanced models such as DeepSeek R1 and OpenAI o1, which effectively leverage reward-driven self-exploration to enhance reasoning capabilities and generalization, we propose a novel set of partial rewards tailored specifically for the Text-to-SQL task. Our reward set includes schema-linking, AI feedback, n-gram similarity, and syntax check, explicitly designed to address the reward sparsity issue prevalent in reinforcement learning (RL). Leveraging group relative policy optimization (GRPO), our approach explicitly encourages large language models (LLMs) to develop intrinsic reasoning skills necessary for accurate SQL query generation. With models of different sizes, we demonstrate that RL-only training with our proposed rewards consistently achieves higher accuracy and superior generalization compared to supervised fine-tuning (SFT). Remarkably, our RL-trained 14B-parameter model significantly outperforms larger proprietary models, e.g. o3-mini by 4% and Gemini-1.5-Pro-002 by 3% on the BIRD benchmark. These highlight the efficacy of our proposed RL-training framework with partial rewards for enhancing both accuracy and reasoning capabilities in Text-to-SQL tasks.
View details
