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 11421 publications
Preview abstract
In this work, we prove computational lower bounds against differentially private (DP) coreset. Specifically, assuming the existence of one-way functions, we show that no polynomial-time $(\epsilon, 1/n^{\omega(1)})$-DP algorithm can compute $(\alpha, \beta)$-coreset for $k$-means in the $\ell_\infty$ metric for some constant $\alpha > 1$. For the Euclidean metric, we show a similar result but only for $\alpha = 1 + \Theta\left(\frac{1}{d^2}\right)$ where $d$ is the dimension.
View details
ConvApparel: A Benchmark Dataset and Validation Framework for User Simulators in Conversational Recommenders
Ofer Meshi
Guy Tennenholtz
Jihwan Jeong
Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL-26), Rabat, Morocco (2026), pp. 5270-5304
Preview abstract
LLM-based user simulators are a scalable solution for improving conversational AI, but a critical realism gap undermines their effectiveness. To close this gap, we introduce a framework for building and validating high-fidelity simulators. We present a novel dataset of human-AI shopping conversations designed to capture a wide spectrum of user experiences. To measure fidelity, we propose a hybrid evaluation protocol that combines statistical alignment with a learned, discriminator-based Human-Likeness Score. Our most sophisticated simulator, trained via reinforcement learning with iterative critique, achieves a significant leap in realism. Critically, we demonstrate through counterfactual validation that our simulator—trained exclusively on optimal interactions—realistically adapts its behavior to suboptimal system responses, mirroring real user reactions and marking a key advance in creating reliable simulators for robust AI development.
View details
Taming the Variants Multi-Architecture Continuous Testing at Google
Chandrakanth Chittappa
Ali Esmaeeli
Laura Macaddino
Sam Manfreda
David Margolin
Dharma Naidu
Sabuj Pattanayek
Sachin Sable
Ruslan Sakevych
Dushyant Acharya
Adrian Berding
Kevin Crossan
Wolff Dobson
Abhay Singh
19th IEEE International Conference on Software Testing, Verification and Validation (ICST) 2026, Daejeon, Republic of Korea, IEEE
Preview abstract
Enterprises are increasingly adopting multiple general-purpose computer architectures in the data center. This leads to new testing challenges as it creates demand to qualify the software for the additional architectures. Naively double-testing all software for both architectures is costly and unnecessary. Further, reconfiguring CI/CD to take advantage of the new architecture can be non-trivial at scale. This paper introduces CI/CD variants and an optimized testing cycle to solve these twin challenges. We empirically evaluate our solution's impact on human and machine expenses using 44k projects at Google on real production data. First, we estimate saving ~25% of machine expenses at the negligible cost of a few delayed breakage detections per day. Second, we estimate a 90+% reduction in human cost for migrating the configuration. All features described in this paper are now Generally Available at Google and we report this as an empirical case study in scaling CI/CD to new architectures.
View details
Preview abstract
Enterprise service centers, particularly in domains like People Operations, are critical hubs of organizational knowledge work. They face a persistent difficulty in disseminating the tacit, case-specific expertise of senior agents, which can lead to inconsistent service and slower onboarding for new hires. While existing Knowledge Management (KM) and Case-Based Reasoning (CBR) systems have improved the retrieval of historically similar cases, they inadvertently shift the cognitive burden of synthesizing this information to the time-constrained agent. This paper introduces the Dynamic Case Precedent (DCP) architecture, a novel socio-technical framework designed to address this gap. The DCP architecture moves beyond simple precedent recommendation to automated precedent synthesis. It achieves this by integrating a semantic retrieval model with the large-context reasoning capabilities of a generative Large Language Model (LLM). We propose a three-pillar framework—(1) Contextual Similarity Indexing, (2) Generative Insight Synthesis, and (3) Human-in-the-Loop Refinement. By analyzing multiple relevant historical cases to generate a concise summary of resolution patterns, the DCP architecture aims to reduce agent cognitive load, accelerate proficiency, and improve service consistency. This conceptual framework offers a new model for human-AI collaboration, framing the AI not as a mere information tool, but as an active partner in sensemaking.
View details
Preview abstract
We introduce AASE (Activation-based AI Safety Enforcement), a framework for post-perception safety monitoring in large language models. Unlike pre-perception approaches that analyze input or output text, AASE monitors the model's internal activation patterns—what the model "understands" rather than what text it processes or generates—enabling detection of safety-relevant states before harmful outputs are produced. The framework comprises three techniques: Activation Fingerprinting (AF) for harmful content detection, Agent Action Gating (AAG) for prompt injection defense, and Activation Policy Compliance (APC) for enterprise policy enforcement. We introduce paired contrastive training to isolate safety-relevant signals from confounding factors such as topic and style, addressing signal entanglement in polysemantic activations. Validation across 7 models from 3 architecture families shows strong class separation: Gemma-2-9B achieves AUC 1.00 with 7.2σ separation across all probes; AAG achieves AUC ≥0.88 across all models on the InjecAgent benchmark; APC achieves 0.97-1.00 AUC across three enterprise policies. Model size correlates with probe quality—Gemma-2-9B (7.2σ separation) outperforms Gemma-2-2B (4.3σ). All techniques survive INT4 quantization with minimal separation degradation. AASE is 9× faster than Llama Guard 3 (33ms vs 306ms) with higher TPR (88% vs 50%) at a tunable threshold that trades FPR for detection sensitivity, adding only 0.002ms probe overhead to existing inference.
View details
Mull-Tokens: Modality-Agnostic Latent Thinking
Arijit Ray
Chengzhi Mao
Bryan A. Plummer
Kate Saenko
Ranjay Krishna
Leonidas Guibas
Vincent Chu
IEEE/CVF Conference on Computer Vision and Pattern Recognition (Findings) (2026) (to appear)
Preview abstract
Reasoning goes beyond language; the real world requires reasoning about space, time, affordances, and much more that words alone cannot convey. Existing multimodal models exploring the potential of reasoning with images are brittle and do not scale. They rely on calling specialist tools, costly generation of images, or handcrafted reasoning data to switch between text and image thoughts. Instead, we offer a simpler alternative -- Mull-Tokens -- modality-agnostic latent tokens pre-trained to hold intermediate information in either image or text modalities to let the model think free-form towards the correct answer. We investigate best practices to train Mull-Tokens inspired by latent reasoning frameworks. We first train Mull-Tokens using supervision from interleaved text-image traces, and then fine-tune without any supervision by only using the final answers. Across four challenging spatial reasoning benchmarks involving tasks such as solving puzzles and taking different perspectives, we demonstrate that Mull-Tokens improve upon several baselines utilizing text-only reasoning or interleaved image-text reasoning, achieving a +3% average improvement and up to +16% on a puzzle solving reasoning-heavy split compared to our strongest baseline. Adding to conversations around challenges in grounding textual and visual reasoning, Mull-Tokens offers a simple solution to abstractly think in multiple modalities.
View details
Preview abstract
Every team building Large Language Models (LLMs) faces a core challenge: offline benchmarks show performance gains and user engagement rises post deployment, but isolating cause from effect remains difficult. Simultaneous marketing, media coverage, and seasonal demand obscure whether model updates truly drive engagement gains. This paper presents a novel causal estimation approach that leverages non uniform quality improvements across capabilities within a single model version. Because capabilities improve unevenly (e.g., strong gains in coding versus modest gains in writing), users experience varied quality depending on their task distribution. This variation in experienced quality provides causal signal for estimation. We validate this approach on synthetic data with known ground truth. The raw estimator recovers 81% to 88% of the true effect, with the remainder lost to measurement error in usage estimates. Applying an errors in variables disattenuation adjustment (using a test retest reliability ratio mean correlation of 0.861) corrects the estimate to 1.017 (bootstrapped 95% CI: 0.915 to 1.109). By contrast, naive methods fail significantly, recovering only 63% without version controls and 62% using real time rather than frozen usage patterns. Permutation tests confirm the framework distinguishes true causal effects from noise. Sensitivity analysis indicates recovery improves monotonically from 78% to 87% as pre period length increases from 3 to 7 weeks, highlighting a clear tradeoff between sample duration and estimator precision. Seed sensitivity tests further confirm stability across random draws.
View details
Preview abstract
Quantization methods have significantly improved the compute and memory efficiency of Large Language Model (LLM) training. However, existing approaches still rely on accumulating their updates into high precision: concretely, gradient updates must be applied to a high-precision weight buffer, known as \textit{master weights}. This buffer introduces substantial memory overhead, particularly for Sparse Mixture of Experts (SMoE) models, where model parameters and optimizer states dominate memory usage. In this work, we introduce the Error-Compensating Optimizer (ECO), which \textit{for the first time} enables the complete elimination of master weights by directly accumulating updates into quantized parameters, by leveraging existing optimizer states. ECO quantizes the weights after every gradient step and injects the resulting quantization error into the optimizer's momentum buffer, creating an error-feedback loop with zero additional memory overhead for quantization. Beyond its practical efficiency, ECO comes with theoretical guarantees. Specifically, under standard assumptions, naive master weight removal can lead to unbounded drift from the ideal parameter trajectory, whereas ECO provably bounds this drift, ensuring stable convergence. We validate ECO across a range of models, including small transformers (30M--800M), Gemma-3 1B, and an SMoE 2.1B model, using FP8 quantization. In all cases, ECO achieves near-lossless accuracy compared to high-precision baselines. For large SMoE models, ECO reduces memory usage by up to 25\%, establishing a new Pareto frontier for the trade-off between static memory and training loss.
View details
Differential Sensitivity of Impedance Plethysmography and Photoplethysmography Sensors to Temperature-Induced Peripheral Vasoconstriction
Seobin Jung
Seamus Thomson
Alexandros Pantelopoulos
Lindsey Sunden
Pete Richards
Shwetak Patel
Sam Sheng
Scientific Reports (2026)
Preview abstract
Impedance plethysmography (IPG) and photoplethysmography (PPG) are non-invasive techniques for measuring blood volume changes. This study investigated the differential responses of IPG and PPG to temperature-mediated vasoconstriction induced by localized cooling. Twenty-one participants underwent control and treatment conditions, with fake or real ice cubes applied to the forearm. Blood pressure remained stable, while heart rate decreased. PPG signal amplitude significantly decreased with cooling (p_adj = 0.004), indicating sensitivity to superficial blood flow changes. In contrast, IPG signal amplitude remained stable (p_adj = 1.0). No statistically significant differences were observed in timing-derived metrics. These findings suggest IPG is less sensitive to superficial changes in blood flow than PPG, and may be more suitable for monitoring deeper blood flow. This study provides insights into the distinct sensitivities of IPG and PPG, with implications for wearable device development and cardiovascular monitoring.
View details
Preview abstract
Here’s a thought experiment. Say I wave a magic wand across a codebase and an entire class of technical debt, poof, goes away and immediately evaporates if introduced in the future. For example, maybe I make it so that dead feature flags are simply no longer a problem: they just delete themselves as soon as the engineer wills it. Or maybe large-scale migrations just migrate themselves. Maybe we magically have 100% test coverage, without an engineer lifting a finger.
What will happen to developer productivity?
Surely, developer productivity increases overall.
But will the productivity metrics that we all use as a proxy for “developer productivity” move up and to the right. Let’s explore this idea.
View details
Preview abstract
This defensive publication describes a framework for multi-artificial intelligence (AI) orchestration that can be used to address potential limitations associated with reliance on single AI models, such as correlated systemic failures or cognitive blind spots. The described system is a cognitive orchestration framework that can function as a middleware layer to manage tasks across a heterogeneous ensemble of AI models. An orchestrator node can decompose a user request into a sequence of sub-tasks, which an arbitrage engine may then dynamically assign to suitable AI models based on certain factors, such as capability, cost, and latency. For certain tasks, such as those designated as high-risk, a byzantine consensus layer can route the task to multiple diverse models in parallel and may trigger a process, for example a 'cognitive debate,' which could be adjudicated by a third-party judge model to help resolve conflicting outputs. This framework can facilitate a more resilient system that may improve the accuracy and reliability of outputs when compared to some single-model architectures.
View details
Uncovering Relationships between Android Developers, User Privacy, and Developer Willingness to Reduce Fingerprinting Risks
Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26) (2026)
Preview abstract
The major mobile platforms, Android and iOS, have introduced changes that restrict user tracking to improve user privacy, yet apps continue to covertly track users via device fingerprinting. We study the opportunity to improve this dynamic with a case study on mobile fingerprinting that evaluates developers’ perceptions of how well platforms protect user privacy and how developers perceive platform privacy interventions. Specifically, we study developers’ willingness to make changes to protect users from fingerprinting and how developers consider trade-offs between user privacy and developer effort. We do this via a survey of 246 Android developers, presented with a hypothetical Android change that protects users from fingerprinting at the cost of additional developer effort.
We find developers overwhelmingly (89%) support this change, even when they anticipate significant effort, yet prefer the change be optional versus required. Surprisingly, developers who use fingerprinting are six times more likely to support the change, despite being most impacted by it. We also find developers are most concerned about compliance and enforcement. In addition, our results show that while most rank iOS above Android for protecting user privacy, this distinction significantly reduces among developers very familiar with fingerprinting. Thus there is an important opportunity for platforms and developers to collaboratively build privacy protections, and we present actionable ways platforms can facilitate this.
View details
Preview abstract
In "Elephants, Goldfish and the New Golden Age of Software Engineering," the author discusses how AI is changing knowledge work, especially software development. Written from the perspective of April 2026, the article points out that while AI speeds up coding, it can also quickly generate a lot of mistakes and messy code if it isn't carefully managed by human oversight and clear processes.
The paper outlines a practical approach to working with AI, broken down into three main sections:
Using AI as a Tool, Not a Toy:
The author notes that people often get poor results by asking AI to do everything in a single prompt. Instead, users should have back-and-forth conversations with AI to question assumptions, set clear grading rules, and guide the research. The main point is that humans must still provide the final judgment; AI is simply a way to speed up and record that thinking.
The Elephant-Goldfish Model:
As AI creates more code than humans can easily read, written design documents become more important than the code itself. To keep AI on track, the author suggests a two-part method:
* The Elephant: A long chat session where the human and AI discuss ideas and write a detailed design document *before* any code is written. This session holds all of the project's background information and decisions.
* The Goldfish: A brand-new AI chat session with no memory. The human asks this "goldfish" to read the design document. If the goldfish cannot understand the plan based only on that document, the document needs more details.
* Only after the design document is clear enough for the goldfish to understand does the human ask the AI to write the code based on those strict instructions.
* Managing AI and the Future of Work:
The author expects that regular employees will soon act like managers, overseeing multiple AI helpers. Because of this, workers need to learn basic management skills, like how to delegate tasks and set clear boundaries. Also, since AI will handle routine chores, humans will need to practice focusing for longer periods to do deeper, harder thinking. Ultimately, a worker's value will come from their planning and decision-making skills, rather than their ability to type code.
View details
Preview abstract
Generative AI (GenAI) is evolving from standalone tools to interconnected ecosystems that integrate chatbots, cloud platforms, and third-party services. While this ecosystem model enables personalization and extended services, it also introduces complex information flows and amplifies privacy risks. Existing solutions focus on system-level protections, offering little support for users to make meaningful privacy choices. To address this gap, we conducted two vignette-based survey studies with 486 participants and a followup interview study with 16 participants. We also explored users’ needs and preferences for privacy choice design across both GenAI personalization and data-sharing. Our results reveal paradoxical patterns: participants sometimes trusted third-party ecosystems more for personalization but perceived greater control in first-party ecosystems when data was shared externally. We discuss design implications for privacy choice interfaces that enhance transparency, control, and trust in GenAI ecosystems.
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
