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 11648 publications
When Android Apps Request Location with a Reason: A Developer Study and User Behavior Experiment
May Smith
Aman Jain
Proceedings on Privacy Enhancing Technologies (2027)
Preview abstract
Recent reports have highlighted how mobile apps share user location data with third parties, risking user privacy and platform trust. Although location data is highly sensitive, when users grant apps location access, they may not know the full extent to which it is used. We study how requiring Android apps to show a reason for location access could impact developers, users, and the platform.
We surveyed 323 Android app developers and found most supported such a requirement. The majority said it would have a positive impact on user privacy, trust for apps, and trust for Android, where impact on user trust for Android correlated most strongly with support. Many developers also said the intervention would increase the number of users granting location access. Yet their open-ended comments also revealed consistent concerns, such as apps providing dishonest reasons and platform verification.
To study the impact on user behavior, we conducted a randomized controlled experiment with 2579 US Android users. We tested how users' decisions to grant location access were impacted by app type, whether reasons were included in the requests, and the content of the reasons, including monetization. We did not find the reasons impacted users' decisions; decisions were instead driven by app type and demographics. Yet we did find the reasons could have a positive impact on user perception for the platform when the reasons did not include using data for ads. Our findings provide insights into developers' willingness to implement privacy-enhancing changes, and expose limits to improving user privacy by simply adding information to user interfaces.
View details
A 3D Scene Graphs Survey: Open Challenges and Future Directions
Dennis Rotondi
Francesco Argenziano
Sebastian Koch
Nathan Hughes
Martin Büchner
Johanna Wald
Lukas Schmid
Daniele Nardi
Abhinav Valada
Liam Paul
Luca Carlone
Kai Arras
Annual Review of Control, Robotics, and Autonomous Systems (ARCRAS), 10 (2027) (to appear)
Preview abstract
3D Scene Graphs (3DSGs) have emerged as a powerful representation for spatial AI by combining geometric grounding with semantic and relational abstractions of the environment. Their expressiveness has made them relevant to a broad range of problems in robotics and computer vision, including mapping, task and motion planning, scene understanding, and many others. However, the field remains fragmented: different communities adopt distinct formulations, construction pipelines, and evaluation protocols, making it difficult to compare methods, identify common assumptions, and assess remaining challenges for robust real- world deployment. This survey provides a unified and critical review of 3DSGs, with particular emphasis on open challenges and future directions. We first formalize 3DSGs under a common definition and analyze the principal modeling choices that characterize existing formulations, including node and edge attributes, hierarchical structure, dynamic scene representations, and affordance-aware extensions. We then review how 3DSGs are constructed from raw sensory observations, covering both learning-oriented and construction-oriented systems. Finally, we examine downstream applications and evaluation strategies, from intrinsic graph quality to task-level performance. To support the community, we also provide a dedicated website that organizes and extends the surveyed works.
View details
Preview abstract
Modern transformers have been scaled up in size, data, training-time compute, and more recently test-time compute based on Chain-of-Thought (CoT), but the consequences of trading off these parameters are unknown. We construct explicit tradeoffs between two of these aspects: the number of layers in a transformer and the number of CoT tokens. Our first result is general: with mild assumptions and a single-token prompt, for any integer k, a transformer using soft attention and SwiGLU MLP layers with L layers and n CoT tokens can be converted to a transformer with at most ckL layers and ⌈n/k⌉ CoT tokens, and vice versa, although with an increase to the size of other parameters. We then a multi-token prompt path following task, finding that additional preprocessing from the multi-token prompt changes the depth-CoT relationship in this case. We construct an attention-only network with L layers and 2^(L−2) heads that can traverse 2^(L−1) − 1 edges along a path with a single CoT token, demonstrating at least an exponential relationship between increases in depth and reductions in the number of CoT tokens needed to traverse the entire path. This relationship for the path following task holds empirically as well, as demonstrated by training networks with up to 5 layers, with the tradeoff between depth and the maximum number of edge traversals in a single CoT token exactly matching our theoretical constructions.
View details
Preview abstract
We introduce AMS (Activation-based Model Scanner), a tool that detects modifications to safety training in language models by measuring the geometric structure of safety-relevant concepts in activation space. Safety training creates measurable separation between harmful and benign content classes; certain safety modifications collapse or rotate this structure, while others leave it intact. We validate AMS across 14 model configurations spanning 4 architecture families (Llama, Gemma, Qwen, Mistral) and four safety-modification categories (instruction-tuned, base, abliterated, uncensored fine-tunes). Leave-one-out cross-validation of thresholds achieves 71% accuracy (10/14); bootstrap 95% confidence intervals on σ point estimates have median width 3.4σ and a substantial fraction of cells cross the PASS threshold under resampling. We further measure behavioral compliance on 20 stratified JailbreakBench prompts per model and find that σ on the harmful-content concept predicts compliance with Pearson r=−0.546 ( p=0.043 ); the rank-order Spearman correlation is weaker ( ρ=−0.423 , p=0.13 ). The structural signal predicts behavior directionally but with meaningful noise. Mechanistic analysis identifies a four-class taxonomy of safety-training modifications distinguished by activation-space signature: 1) training removal collapses cluster separation (e.g., base models, Dolphin variants: 0.5– 1.4σ ); 2) weight-orthogonalization-style abliteration both collapses separation and rotates the refusal direction (Llama-3.1-abliterated: σ=3.33 , direction cos sim 0.30); 3) rotation-without-collapse abliteration preserves cluster separation while rotating the refusal direction (Gemma-2-9b-abliterated: σ=4.54 , direction cos sim 0.84); and 4) behavioral fine-tuning that preserves both magnitude and direction (DarkIdol-1.2-Uncensored: σ=5.45 , direction preserved, 97% behavioral compliance). 1) and 2) AMS’s Tier 1 σ -threshold detects classes; 3) Tier 2 direction-similarity verification detects class; and 4) Class is undetectable by activation-only probing and represents a documented failure mode of the approach. We discuss threshold calibration, limitations of single-run measurement, and the open problem of detecting behavioral-only safety modifications.
View details
Preview abstract
Massive KV caches can cause severe memory-bandwidth bottlenecks during long-context decoding. Sparse attention methods mitigate this via selective loading, but that comes at a cost: rigid heuristics drop necessary context, leading to quality degradation. We introduce \textbf{Elastic Threshold Attention (ETA)}, an end-to-end trainable architecture that achieves hardware-accelerated decoding speed without sacrificing dense model quality. ETA predicts dynamic, contextual thresholds directly from query representations, allowing the model to allocate dense-like context to difficult retrieval or reasoning steps while pruning routine tokens. To learn this policy from scratch without representation collapse, ETA \emph{multiplicatively suppresses} sub-threshold logits toward zero during training rather than deleting them. Training against this smooth uniform attention floor provides a distributed probability reservoir that \textbf{causes localized attention sinks on initial tokens to disappear}. It also enables the model to hard-prune uninformative KV blocks at inference time and absorb incidental tokens co-admitted by coarse GPU block selection. As a result, a 1.45B pretrained ETA model rivals dense attention across language modeling, commonsense reasoning, and long-context needle retrieval at $\approx 85\%$ training sparsity and $\approx 38\%$ active decode density. At inference time, we implement a custom decode kernel in Triton that screens KV blocks in $O(1)$ time using cached geometric-probabilistic bounds, delivering up to $2.5\times$ wall-clock decode speedups over FlashAttention-2 on sequences up to 512K tokens. Finally, we introduce an offline calibration algorithm for domain-specific deployments that freezes per-head constant thresholds to eliminate predictor overhead, cutting attention compute by an additional $27\%$.
View details
Performance analysis of updated Sleep Tracking algorithms across Google and Fitbit wearable devices
Arno Charton
Linda Lei
Siddhant Swaroop
Marius Guerard
Michael Dixon
Logan Niehaus
Shao-Po Ma
Ross Wilkinson
Ryan Gillard
Conor Heneghan
Pramod Rudrapatna
Mark Malhotra
Shwetak Patel
Google, Google, 1600 Amphitheatre Parkway
Mountain View, CA 94043 (2026) (to appear)
Preview abstract
Background: The general public has increasingly adopted consumer wearables for sleep tracking over the past 15 years, but reports on performance versus gold standards such as polysomnogram (PSG), high quality sleep diaries and at-home portable EEG systems still show potential for improved performance. Two aspects in particular are worthy of consideration: (a) improved recognition of sleep sessions (times when a person is in bed and has attempted to sleep), and (b) improved accuracy on recognizing sleep stages relative to an accepted standard such as PSG.
Aims: This study aimed to: 1) provide an update on the methodology and performance of a system for correctly recognizing valid sleep sessions, and 2) detail an updated description of how sleep stages are calculated using accelerometer and inter-beat intervals
Methods: Novel machine learning algorithms were developed to recognize sleep sessions and sleep stages using accelerometer sensors and inter-beat intervals derived from the watch or tracker photoplethysmogram. Algorithms were developed on over 3000 nights of human-scored free-living sleep sessions from a representative population of 122 subjects, and then tested on an independent validation set of 47 users. Within sleep sessions, an algorithm was developed to recognize periods when the user was attempting to sleep (Time-Attempting-To-Sleep = TATS). For sleep stage estimation, an algorithm was trained on human expert-scored polysomnograms, and then tested on 50 withheld subject nights for its ability to recognize Wake, Light (N1/N2), Deep (N3) and REM sleep relative to expert scored labels.
Results: For sleep session estimation, the algorithm had at least 95% overlap on TATS with human consensus scoring for 94% of nights from healthy sleepers. For sleep stage estimation, comparing with the current Fitbit algorithm, Cohen’s kappa for four-class determination of sleep stage increased from an average of 0.56 (std 0.13) to 0.63 (std 0.12), and average accuracy increased from 71% (std 0.10) to 77% (std 0.078)
Conclusion: A set of new algorithms has been developed and tested on Fitbit and Pixel Watches and is capable of providing robust and accurate measurement of sleep in free-living environments.
View details
Visual Planning: Let’s Think Only with Images
Han Zhou
Caiqi Zhang
Anna Korhonen
Chengzu Li
Yi Xu
Ivan Vulic
International Conference on Learning Representations (ICLR) (2026)
Preview abstract
Recent advancements in Large Language Models (LLMs) and their multimodal extensions (MLLMs) have significantly enhanced machine reasoning across diverse tasks. However, these models predominantly rely on language as the medium for both expressing and structuring reasoning, even when visual information is present. In this work, we argue that language may not always be the most natural or effective modality for reasoning, particularly in tasks involving spatial, geometric, or physical dynamics. Motivated by this, we propose a new paradigm, Visual Planning, which enables planning through purely visual representations, independent of textual mediation. In this paradigm, planning is executed via sequences of images that encode step-by-step inference in the visual domain, akin to how humans sketch or visualize future actions. We then introduce a novel two-stage reinforcement learning framework empowered by GRPO for post-training large vision models, resulting in substantial improvements in planning accuracy and generalization across both seen and novel scenarios, validated in representative visual navigation tasks, FrozenLake and Maze. Our results establish Visual Planning as a viable and promising alternative to language-based reasoning, opening new avenues for tasks that benefit from intuitive, image-based inference.
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
An AI system to help scientists write expert-level empirical software
Eser Aygün
Anastasiya Belyaeva
Gheorghe Comanici
Hao Cui
Renee Johnston
Zahra Shamsi
David Smalling
James Thompson
Sarah Martinson
Lai Wei
Yuchen Zhou
Qian-Ze Zhu
Matthew Abraham
Erica Brand
Anna Bulanova
Jeffrey Cardille
Chris Co
Scott Ellsworth
Grace Joseph
Malcolm Kane
Ryan Krueger
Johan Kartiwa
Jackson Cui
Paul Raccuglia
Julie Wang
Kat Chou
James Manyika
Lizzie Dorfman
Shibl Mourad
Nature (2026)
Preview abstract
The cycle of scientific discovery is frequently bottlenecked by the slow, manual creation of software to support computational experiments. To address this, we present Empirical Research Assistance (ERA), an AI system that creates expert-level scientific software whose goal is to maximize a quality metric. The system uses a Large Language Model (LLM) and Tree Search (TS) to systematically improve the quality metric and intelligently navigate the large space of possible solutions. ERA achieves expert-level results when it explores and integrates complex research ideas from external sources. The effectiveness of tree search is demonstrated across a diverse range of tasks. In bioinformatics, ERA discovered 40 novel methods for single-cell data analysis that outperformed the top human-developed methods on a public leaderboard. In epidemiology, ERA generated 14 models that outperformed the CDC ensemble and all other individual models for forecasting COVID-19 hospitalizations. ERA also produced expert-level software for geospatial analysis, neural activity prediction in zebrafish, and numerical solution of integrals, and a novel rule-based construction for time series forecasting. By devising and implementing novel solutions to diverse tasks, ERA represents a significant step towards accelerating scientific progress.
Keywords: Tree Search, Generative AI, Scorable Scientific Tasks, Empirical Software
View details
Parametric Robustness Assessment of Ultra-Low Voltage Standard Cells via High-Sigma Verification (HSV)
Varshini Giri
Sravanth Mucharla
2026
Preview abstract
In modern digital System-on-Chip (SoC) designs, sequential standard cells—such as single-bit D-Flip-Flops (DFFs) and Multibit Flip-Flops (MBFFs)—serve as fundamental building blocks for memory arrays, registers, and pipelines. To meet ultra-low-power requirements in mobile, edge, and high-performance computing (HPC) applications, these circuits increasingly operate at near-threshold supply voltages (VDD0.45-0.50V). However, aggressive voltage scaling introduces critical failure mechanisms:
Write-Back Failures: At low supply voltages, transmission gate or pass-gate write paths fail to overpower cross-coupled feedback inverters, preventing state transitions and causing cells to retain stale values.
Hold-Time/Race Violations: Clock skew and fast transitions trigger internal races. When process variation weakens clock gating or feedback control, internal storage nodes degrade, leading to transient glitches, state corruption, or functional failure.
To verify standard cell reliability, brute-force Monte Carlo (MC) simulations are computationally prohibitive, requiring billions of SPICE iterations. This work outlines a High-Sigma Verification (HSV) methodology using advanced statistical sampling—such as Importance Sampling, Boundary Search, or machine-learning-driven MC—to evaluate rare, extreme variation tails with a fraction of the computational overhead.
By mapping the multidimensional local variation space, the framework calculates true statistical yield margins (the "sigma-metric") and identifies dominant transistor-level variation sources, including pull-up/pull-down network balance and pass-gate drive strength. Beyond single-point analyses, the workflow bridges layout-level parasitic extraction and silicon-level yield predictability by scaling high-sigma evaluation across comprehensive PVT corners and diverse input slew/output load profiles. This multi-corner, slew-aware verification captures tail-distribution failures and dynamic delay non-linearities under extreme operating conditions, delivering characterization-ready yield bounds for robust, near-threshold SoC sign-off.
View details
Preview abstract
Mid-air gestures in Extended Reality (XR) often lead to fatigue, discomfort and imprecision, limiting their suitability for extended use. Surface-based interactions offer a compelling alternative, providing improved accuracy, speed, and comfort. However, current egocentric vision-based methods struggle with reliable surface inputs due to challenges in hand tracking and surface-plane estimation from oblique and occluded viewing angles. To this extent, we introduce SurfaceXR, a novel sensor fusion approach that combines headset based hand tracking with micro-vibration data sampled from commodity smartwatch IMUs to enable precise and robust inputs on arbitrary surfaces. Our system is designed with flexibility in mind - it can function using only hand tracking, only IMU sensing, or optimally with both modalities combined. Our user study across 12 participants validates SurfaceXR's effectiveness in augmenting surface touch tracking and 8 class hand-surface gesture recognition, demonstrating significant improvements over single-modality approaches. Enabled by SurfaceXR, we demonstrate a series of interactive apps for both AR and VR, ranging from on-surface sketching, text entry and gesture based navigation.
View details
ARM-Orchestrated Dynamic Workload Partitioning for Vision Model Training on FPGA Accelerators
Ram Babu Penugonda
Swaroop Arlagadda
2026
Preview abstract
Silent performance degradation and accelerator contention remain largely unaddressed in contemporary heterogeneous training systems, where Field-Programmable Gate Array (FPGA) offloading decisions are often static, heuristic-driven, or evaluated using idealized assumptions that ignore queueing effects, transfer overheads, and shared accelerator load. This paper introduces a runtime, price-conscious co-execution integration of CPU and FPGA to train deep learning models, where execution choices are dynamically made at the training stage granularity with a discrete latency cost model. The proposed method incorporates the CPU execution latency estimation, FPGA queue depth monitoring, time-varying service rate modeling, background contention, data transfer overhead, and switching hysteresis into a single decision mechanism that is driven by an ARM-based runtime monitor. The model is realistic FPGA queueing model, which is used to simulate shared accelerator contention and avoid optimistic assumptions of performance. Experimental evaluation on a representative convolutional training pipeline demonstrates that the cost-minimizing orchestration case achieves a mean training step time of 0.0213 s (±0.0036) with a throughput proxy of 47.03 steps/s, while maintaining bounded FPGA queue behavior and 99.94% stall avoidance relative to static FPGA execution. Conversely, a different execution setup that has a greater FPGA usage shows a greater queue buildup, an average step time of 0.0330 s, a maximum FPGA queue depth of 77 units, and a high level of congestion exposure, which demonstrates how crucial a queue-aware execution configuration is. These findings underscore the fact that the acceleration capability of the FPGA is not controlled by the raw compute capability but it is controlled by runtime sensitive and contention-sensitive orchestration, creating a principled and reproducible methodology to evaluate heterogeneous training systems in realistic conditions of accelerator sharing.
View details
Preview abstract
Statistical evaluation aims to estimate the generalization performance of a model using held-out i.i.d. test data sampled from the ground-truth distribution. In supervised learning settings such as classification, performance metrics such as error rate are well-defined, and test error reliably approximates population error given sufficiently large datasets. In contrast, evaluation is more challenging for generative models due to their open-ended nature: it is unclear which metrics are appropriate and whether such metrics can be reliably evaluated from finite samples. In this work, we introduce a theoretical framework for evaluating language models and establish evaluability results for commonly used metrics. We study two categories of metrics: test-based metrics, including integral probability metrics (IPMs), and similarity-based metrics, including Rényi and KL divergences. We show that IPMs with respect to any bounded test class can be evaluated from finite samples up to multiplicative and additive approximation errors. Moreover, when the test class has finite fat-shattering dimension, IPMs can be evaluated with arbitrary precision. In contrast, similarity-based metrics, including Rényi and KL divergences, are not evaluable from finite samples, as their values can be critically determined by rare events. We also analyze the potential and limitations of perplexity as an evaluation method.
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
Working Together on Technologies: A Case Study of Collaboration in Aotearoa
Gabi Conlon
Greg Duncum
Jamie Towler
Carrie Jones
Tane Karamaina
Hautahi Kingi
Caroline Rainsford
Ngahiwi Apanui-Barr
2026
Preview abstract
This paper presents a case study in collaboration on TTS technologies between Te Taura Whiri and Google. The former is a Māori-led organisation whose mission is the revitalisation of te reo Māori. The latter is a global technology company which has a presence in Aotearoa but which is headquartered elsewhere. This case study focuses not on technical implementation details, but rather on the challenges in the partnership to mutual understanding and alignment, and the processes we co-developed to overcome those challenges, which led to the successful development of technology which was satisfactory to both partners.
View details
