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 11402 publications
Preview abstract
While Large Language Models (LLMs) excel at many tasks, they frequently struggle with complex reasoning that requires long-horizon planning and iterative error correction. Furthermore, standard single-stream prompting proves brittle when models encounter novel abstractions or rigorous domain constraints. We introduce PoTRE (Poly-Topological Reasoning Ensembles), a heterogeneous framework that decouples inference into four agents: (1) Adversarial Refinement Agent, (2) Hierarchical strategic Planning Agent, (3) Spectrum Search Agent, and (4) Direct Chain Agent. A final Task-Adaptive Aggregation Layer dynamically reconciles these perspectives -- via final candidate selection, semantic synthesis, or neuro-symbolic verification -- to produce a robust global solution. We evaluate PoTRE on three frontier benchmarks: ARC-AGI-2, Humanity's Last Exam (HLE), and PRBench Finance. PoTRE achieves state-of-the-art accuracy of 49.92% on HLE, surpassing the previous best official score. We demonstrate that this architectural heterogeneity achieves improved reasoning performance using similar or fewer inference tokens compared to heavily scaled homogeneous baselines.
View details
Towards Expert-level Medical AI for Real-time Video Consultations
Mahvish Nagda
Jihyeon Lee
Matthew Thompson
CJ Park
Tim Strother
Roma Ruparel
Teya Bergamaschi
Suhana Bedi
Meet Shah
Pavel Dubov
Toshiyuki Fukuzawa
Sam Schmidgall
Craig Schiff
Joseph Xu
Aliya Rysbek
Yana Lunts
Jan Freyberg
Rebecca Hemenway
Sunny Virmani
David Racz
Carey Radebaugh
Joelle Barral
Kavi Goel
Kat Chou
James Manyika
Gregory Wayne
Yun Liu
Ethan Goh
Christina Chen
Ryutaro Tanno
arXiv, Google (2026)
Preview abstract
Audio-visual interaction is the standard for patient-physician consultations, enabling natural communication and effective assessment of illness through non-verbal cues. While text-based AI has shown promise, it discards essential perceptual dimensions and limits patients who cannot articulate symptoms in writing. Early efforts to extend medical AI to audio-visual interaction have demonstrated feasibility, but not reached clinician-level performance. Here, we provide the first demonstration of expert-level AI in real-time clinical video consultations using AMIE (Articulate Medical Intelligence Explorer) in a video configuration. AMIE (Video) is a Gemini-based multi-agent system integrating low-latency dialogue, clinical reasoning, and real-time audio-visual perception. To guide development, we established a taxonomy and automated evaluations for clinical audio-visual cues in telehealth settings. In a randomized Objective Structured Clinical Examination (OSCE) study with 30 primary care physicians (PCPs), 15 patient actors and 100 clinical scenarios, we compared AMIE (Video), its text-only counterpart AMIE (Text), and PCPs consulting via video. Clinical evaluators rated AMIE (Video) on par or better than PCPs in history-taking, diagnosis, management, and physical observation and examination. Patient actors preferred AMIE's approach to assessing and explaining conditions, while PCPs were preferred for rapport and partnership building. In modality ablation, patient actors preferred AMIE (Video)'s interface over text chat for communicative effectiveness, convenience, and feeling understood. Limitations remain in fine anatomical precision, subtle affective nuances, and high-frequency movements. While further research is needed before real-world translation, these results mark an important milestone toward AI systems capable of augmenting care across the sensory complexity of clinical practice.
View details
Preview abstract
Generative AI’s humanlike qualities are driving its rapid adoption in professional domains. However, this anthropomorphic appeal raises concerns from HCI and responsible AI scholars about potential hazards and harms, such as overtrust in system outputs. To investigate how technology workers navigate these humanlike qualities and anticipate emergent harms, we conducted focus groups with 30 professionals across six job functions (ML engineering, product policy, UX research and design, product management, technology writing, and communications). Our findings reveal an unsettled knowledge environment surrounding humanlike generative AI, where workers’ varying perspectives illuminate a range of potential risks for individuals, knowledge work fields, and society. We argue that workers require comprehensive support, including clearer conceptions of “humanlikeness” to effectively mitigate these risks. To aid in mitigation strategies, we provide a conceptual map articulating the identified hazards and their connection to conflated notions of “humanlikeness.”
View details
Linear-Core Surrogates: Smooth Loss Functions with Linear Rates for Classification and Structured Prediction
Proceedings of the 43rd International Conference on Machine Learning (ICML 2026)
Preview abstract
A fundamental dichotomy in the theory of classification sets smoothness against statistical efficiency: smooth surrogate losses such as the logistic loss enable fast $O(1/T)$ optimization but yield slow square-root $H$-consistency bounds, while piecewise-linear losses like the Hinge loss achieve optimal linear $H$-consistency rates but are non-differentiable. We introduce Linear-Core (LC) Surrogates, the first family of explicit convex loss functions that provably resolve this tension. By stitching a linear core to a smooth tail, we construct surrogates that are differentiable everywhere ($C^1$, and even $C^2$ under mild conditions) while retaining strict linear $H$-consistency bounds, the strongest known form of consistency guarantee. We establish these linear bounds across three increasingly complex settings: binary classification, multi-class classification, and structured prediction. To our knowledge, this is the first explicit construction to simultaneously achieve smoothness and linear $H$-consistency in any of these settings. Beyond their theoretical appeal, Linear-Core Surrogates offer practical advantages. In multi-class classification, their constant gradient profile near the decision boundary provides natural robustness to instance-dependent label noise, outperforming Cross-Entropy by 2.6% on corrupted CIFAR-10. In structured prediction, their smoothness enables an unbiased stochastic gradient estimator that bypasses the $O(|Y|^2)$ per-step complexity of exact inference, yielding a 23$\times$ speedup over Structured SVMs on large-vocabulary sequence tagging tasks.
View details
Towards A Human-in-the-Loop Framework for Reliable Patch Evaluation using an LLM-as-a-Judge
Renyao Wei
Michele Tufano
José Cambronero
AI-SQE '26: Proceedings of the 1st International Workshop on AI for Software Quality Evaluation - Judgment, Metrics, Benchmarks, and Beyond, ACM (Association for Computing Machinery), New York, NY, USA (2026), pp. 19 - 28
Preview abstract
Reliable evaluation is crucial for advancing Automated Program Repair (APR), but prevailing benchmarks that rely on execution-based evaluation methods (pass@k) often fail to capture the patch quality required for real-world adoption. This creates a significant gap between automated metrics and true patch validity (valid@k), a discrepancy observed across several state-of-the-art techniques.
To develop a scalable solution for measuring valid@k, we first study the human evaluation process itself. While manual assessment can determine validity, we find it suffers from poor inter-rater reliability (Fleiss' Kappa k=0.307). Our foundational insight is that this inconsistency is largely resolved when evaluators use a shared, high-quality rubric, which significantly improves agreement.
Building on this finding, we propose an LLM-as-a-Judge framework that operationalizes rubric-guided evaluation at scale. Our method employs a human-in-the-loop workflow where an LLM first generates a candidate rubric for a given bug, which a human expert then reviews and refines into a "golden" evaluation standard. This golden rubric is then used by an LLM judge to assess the validity of candidate patches. In an evaluation on 48 bugs and 115 patches, our LLM judge demonstrates substantial agreement with the consensus of human developers. This work contributes a scalable and reliable methodology for approximating valid@k, providing a much-needed high-fidelity signal for measuring true progress in the field of automated program repair.
View details
Preview abstract
Using generative artificial intelligence with sensitive data may present challenges, as transmitting personally identifiable information or protected health information to third-party providers can introduce security risks, and some data masking techniques can reduce reasoning capabilities. A described system uses a proxy, masking layer that can intercept data within an enterprise's secure perimeter. This layer can substitute sensitive strings with persistent, structured semantic tokens that may be enriched with non-sensitive metadata hints to help preserve context. An external artificial intelligence can perform reasoning on this abstracted data, and its tokenized response can be re-hydrated into readable text on a client device (e.g., a smartphone, computer, or wearable device). This approach may allow third-party models to reason on proprietary information without direct access to the underlying plaintext data, which can assist organizations in managing data sovereignty while maintaining functional utility.
View details
Preview abstract
PURPOSE: To introduce Cardio Load (CL), a metric quantifying cardiovascular work from all activities across the day, and to investigate its distribution by age, gender, and workout profiles. CL adapts the Training Impulse (TRIMP) model by leveraging continuous heart rate and movement data from wearables, enabling minute-level intensity estimation. We also discuss the derivation of weekly target loads, intended to guide fitness maintenance.
METHODS: A retrospective analysis was conducted on 31.2 million hours of wrist-worn wearable data collected over a six-week period. The dataset comprised a 40,000-subject subset (37.9% female) of consenting Google Pixel Watch® users in the United States, aged 18 to 80 years (18-39: 41.8%, 40-59: 43.5%, 60+: 14.6%). Measured data included minute-interval heart rate averages, resting and maximum heart rates, minute-interval averaged accelerometer log energy, and manually-logged or auto-detected activity types. Cardio Load scores and target loads were calculated daily for each subject and compared across age and gender. We also compared the proportions of CL gained during workouts and incidental daily activities for these groups.
RESULTS: Overall, the study population's mean ± SD weekly CL scores were 221 ± 156 (female) and 259 ± 169 (male). Median weekly Cardio Load (CL) values exhibited consistency for individuals between 30 and 75 years of age. When analyzed in five-year age groups, the coefficient of variation (CV%) of median weekly CL values within this age range was less than 4.5%, with younger and older subjects demonstrating higher and lower median CL, respectively. The median proportion of CL accumulated during structured workouts versus incidental daily activity was 41.0% (female) and 49.0% (male) for all subjects, though this varied considerably with average weekly workout duration. CV% of weekly target load and daily target load over 6 weeks was 23.6% and 35.2% respectively.
CONCLUSION: Cardio Load provides a continuous quantification of activity load from wearables, acknowledging both structured workouts and everydayincidental activity. CL is equitably rewarded for age ranges spanning 30-75 years. Weekly target loads were found to have little measurement variability and be more consistent and, consequently, more practical for planning training and physical activity than daily targets.
View details
Preview abstract
**Agentic Engineering** is the rigorous discipline of treating Large Language Models as semi-autonomous systems that execute complex, multi-step workflows (trajectories) based on verifiable specifications, rather than using them as simple autocomplete engines.
Here is a brief summary of its core principles:
* **Main Goals:** It aims to maximize the agent's autonomous run-time, multiply a single engineer's impact by running parallel tasks, and offload tedious boilerplate coding.
* **The "Harness":** A raw model is virtually useless without heavy investment in a harness—comprising tools, system prompts, and strict guardrails—to reliably guide the model and enforce coding policies.
* **Loss of Micro-Control:** Engineers must surrender idiosyncratic stylistic preferences; if the agent's code passes automated linters and tests, it is accepted.
* **Meta-Debugging:** When failures occur, engineers no longer fix code syntax. Instead, they debug the workflow itself—adjusting the agent's tools, search queries, or prompt constraints to ensure repeatable success.
View details
ARM MTE Performance in Practice
Taehyun Noh
Yingchen Wang
Tal Garfinkel
Mahesh Madhav
Mattan Erez
Shravan Narayan
Usenix Security (2026)
Twenty years of Bigtable
Fabio Baltieri
Bora Beran
Igor Bernstein
Aimee Borda
Adrian Chan
Mark D'Andrea
Artak Dashyan
Ramesh Dharan
Gabor Dinnyes
Mike Dominguez
dorland .
Jose Duenas
Gary Elliott
Bruno Furtado
Madison Garcia
Marçal Garolera Huguet
Brendan Gleason
Alexis Hawkins
Anoshak Irani
Rohit Jog
Sudarshan Kadambi
Vikram Khemka
Sailesh Krishnamurthy
Maxim Krivokon
Bruce Lee
Tom Magrino
Matt Maly
Mark Mangrich
Douglas McErlean
Pablo Montes
Li Moore
Eduardo Morales
Greg Morris
Steve Niemitz
Gaurav Prabhu Gaonkar
Jim Rutherford
Stephen Ryan
Sho Saha
Kanoj Sarcar
Cristina Schmidt
Andrii Shyshkalov
Pratibha Suryadevara
Nick Suttle
Anvit Tawar
John Tobin
Justin Uang
Phaneendhar Vemuru
Harendra Verma
Shitanshu Verma
Jinghang (Frank) Wang
Michal Wegorek
Simon Yau
Andrius Ziukas
SIGMOD Companion '26: Companion of the International Conference on Management of Data, ACM (2026), pp. 188-200
Preview abstract
Bigtable is a pioneering and influential non-relational database system. The original Bigtable paper has been widely cited and it inspired and influenced many other systems such as HBase and Cassandra. Since then, Bigtable has continued to grow and has become one of the largest database systems inside Google. In this paper, we tell the journey of Bigtable inside Google for the last twenty years. We present new features added and improvements made to Bigtable, and we share our experience of running this storage system at scale, continually improving all aspects to accommodate the ever-growing demands of users.
View details
Preview abstract
Automating AI research differs from general software engineering due to computationally expensive evaluation (e.g., model training) and opaque performance attribution. Current LLM-based agents struggle here, often generating monolithic scripts that ignore execution costs and causal factors. We introduce MARS (Modular Agent with Reflective Search), a framework optimized for autonomous AI research. MARS relies on three pillars: (1) Budget-Aware Planning via cost-constrained Monte Carlo Tree Search (MCTS) to explicitly balance performance with execution expense; (2) Modular Construction, employing a "Design-Decompose-Implement" pipeline to manage complex research repositories; and (3) Comparative Reflective Memory, which addresses credit assignment by analyzing solution differences to distill high-signal insights. MARS achieves state-of-the-art performance among open-source frameworks on MLE-Bench under comparable settings, maintaining competitiveness with the global leaderboard's top methods. Furthermore, the system exhibits qualitative "Aha!" moments, where 63% of all utilized lessons originate from cross-branch transfer, demonstrating that the agent effectively generalizes insights across search paths.
View details
Beyond PII: How Users Perceive and Attempt to Mitigate Implicit LLM Inference
Synthia Wang
Nick Feamster
Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI), Association for Computing Machinery
Preview abstract
Large Language Models (LLMs) such as ChatGPT can infer personal attributes from seemingly innocuous text, raising privacy risks beyond memorized data leakage. While prior work has demonstrated these risks, little is known about how users estimate and respond. We conducted a survey with 240 U.S. participants who judged text snippets for inference risks, reported concern levels, and attempted rewrites to block inference. We compared their rewrites with those generated by ChatGPT and Rescriber, a state-of-the-art sanitization tool. Results show that participants struggled to anticipate inference, performing a little better than chance. User rewrites were effective in just 28% of cases - better than Rescriber but worse than ChatGPT. We examined our participants’ rewriting strategies, and observed that while paraphrasing was the most common strategy it is also the least effective; instead abstraction and adding ambiguity were more successful. Our work highlights the importance of inference-aware design in LLM interactions.
View details
Mining Attribute Subspaces for Efficient Fine-tuning of 3D Foundation Models
Yu Jiang
Hanwen Jiang
Vincent Chu
Brandon Y. Feng
Zhangyang Wang
Qixing Huang
IEEE/CVF Conference on Computer Vision and Pattern Recognition (2026)
Preview abstract
With the emergence of 3D foundation models, there is growing interest in fine-tuning them for downstream tasks, where LoRA is the dominant fine-tuning paradigm. As 3D datasets exhibit distinct variations in texture, geometry, camera motion, and lighting, there are interesting fundamental questions: 1) Are there LoRA subspaces associated with each type of variation? 2) Are these subspaces disentangled (i.e., orthogonal to each other)? 3) How do we compute them effectively? This paper provides answers to all these questions. We introduce a robust approach that generates synthetic datasets with controlled variations, fine-tunes a LoRA adapter on each dataset, and extracts a LoRA sub-space associated with each type of variation. We show that these subspaces are approximately disentangled. Integrating them leads to a reduced LoRA subspace that enables efficient LoRA fine-tuning with improved prediction accuracy for downstream tasks. In particular, we show that such a reduced LoRA subspace, despite being derived entirely from synthetic data, generalizes to real datasets. An ablation study validates the effectiveness of the choices in our approach.
View details
Preview abstract
A common problem in private data analysis is the partition selection problem, where each user holds a set of partitions (e.g. keys in a GROUP BY operation) from a possibly unbounded set. The challenge here is in maximizing the set of released partitions while respecting a differential privacy constraint. Previous work [DVGM21] presented an optimal (ε, δ)-DP algorithm when each user submits only a single partition. We generalize this approach to find the optimal algorithm under δ-approximate (α, ε)-R´enyi differential privacy (RDP), which allows much tighter analysis under composition. Motivated by the non-existence of a general optimality result in the case where users submit multiple partitions each, we present a simple extension of our optimal algorithm tuned for L2 bounded weighted partition selection which can be used as a drop-in improvement over the Gaussian mechanism any time the partition frequency is not also needed. We show that our primitive can be easily plugged into state of the art partition selection algorithms (PolicyGaussian from [GGK+20] and MAD2R from [CCAEZ25]), improving performance both for parallel and sequential algorithms. Finally, we show that there is an inherent cost to algorithm which do support releasing the frequency as well as the partitions. Specifically, we formulate a basic notion of optimal approximate RDP algorithm for partition selection using additive noise, and show that there is a numerical separation between additive and non-additive noise mechanisms for this problem.
View details
Preview abstract
Generative AI is reshaping software development, yet its psychological impact remains under-researched. During May and August 2025 we conducted reflexive thematic analysis of interviews with 12 senior engineers (≥5 years experience) recruited from Western technology hubs to explore shifts in professional identity. We identify a central transition from "coder to conductor," where AI acts as a cognitive partner. Key findings include: (1) a re-architecting of focus from implementation to strategy; (2) a shift in productivity metrics from output to impact; and (3) a dual-impact on agency, where AI empowers autonomy but threatens competence through de-skilling anxieties. These findings suggest that as implementation becomes commoditised, organisational training and career progression must prioritise architectural mastery and metacognitive oversight to ensure sustained developer motivation and system integrity.
View details
