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 11415 publications
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
This talk addresses the challenges of operating Google's monitoring systems at scale, handling terabytes of telemetry data and preventing overload from diverse workloads. We'll explore how Google's internal client library and Monarch, its planet-scale time-series database, work together for cost-effective data collection. Key principles include a distributed push model, dynamic client-side data reduction, centralized retention, and periodic metric analysis. The session will then bridge these concepts to the open-source world, discussing our work with OpenTelemetry's OpAMP protocol to achieve similar scalable and efficient telemetry collection. Attendees will gain insights into adapting these principles for cost savings and learn about our collaboration with the OpAMP SIG to benefit the broader community.
View details
Preview abstract
This article delves into how Google Site Reliability Engineers (SREs) leverage Gemini 3 and the Gemini CLI to aggressively reduce Mean Time to Mitigation (MTTM) during real-world outages. By focusing on the SRE motto of "Eliminate Toil," the article walks through a simulated incident, demonstrating how an agentic CLI acts as a human-in-the-loop copilot across the entire incident lifecycle: from initial paging and investigation, through safe, tool-driven mitigation and root cause analysis, to automated postmortem generation and action item filing. This direct integration of Gemini's reasoning capabilities with operational data and internal tools creates a virtuous cycle where past incident learnings continuously inform and improve future solutions.
View details
System for a Secure, Outcome-Based Synthetic Labor Market Using Trusted Execution Environments
Patent (2026)
Preview abstract
Some artificial intelligence provisioning models that function as tools for human users or rely on labor arbitrage can present challenges for organizations, such as managing personnel rather than task outcomes and introducing data security risks. An architecture is described for an outcome-based synthetic labor market in which autonomous computational agents can be compensated based on verified task completion. The framework can leverage trusted execution environments to create secure hardware enclaves for processing sensitive data, which can render the data cryptographically inaccessible to a host system or agent provider. This approach can facilitate a secure, transactional market for autonomous professional execution, which may enable a shift from managing labor resources to procuring verified outcomes from a pool of specialized agents.
View details
What does your wearable know about the festive season?
Justin Phillips
Abram Schönfeldt
YongSuk Cho
Conor Heneghan
Robert Harle
(2026)
Preview abstract
As we reach the end of the year and people look forward to spending quality time with loved ones, here at Fitbit, we wonder what our Pixel watches and Fitbit trackers can tell us about how we are spending the festive season. We looked at the data of 11.8 million of our users all over the world between January 2022 and July 2025. Here are the key stats we wanted to share with you!
View details
Reasoning-Driven Synthetic Data Generation and Evaluation
Tim R. Davidson
Benoit Seguin
Transactions on Machine Learning Research (2026)
Preview abstract
Although many AI applications of interest require specialized multi-modal models, relevant data to train such models is inherently scarce or inaccessible. Filling these gaps with human annotators is prohibitively expensive, error-prone, and time-consuming, leading model builders to increasingly consider synthetic data as a scalable alternative. However, existing synthetic data generation methods often rely on manual prompts, evolutionary algorithms, or extensive seed data from the target distribution — limiting their scalability, explainability, and control. In this paper, we introduce Simula: a novel reasoning-driven framework for data generation and evaluation. It employs a seedless, agentic approach to generate synthetic datasets at scale, allowing users to define desired dataset characteristics through an explainable and controllable process that enables fine-grained resource allocation. We show the efficacy of our approach on a variety of datasets, rigorously testing both intrinsic and downstream properties. Our work (1) offers guidelines for synthetic data mechanism design, (2) provides insights into generating and evaluating synthetic data at scale, and (3) unlocks new opportunities for developing and deploying AI in domains where data scarcity or privacy concerns are paramount.
View details
KVCIS: Activation-Based Token Importance Prediction for Intelligent KV-Cache Compression
Zenodo (2026)
Preview abstract
We introduce KVCIS (KV-Cache Importance Scoring), a novel approach to KV-cache compression that predicts token importance from intermediate-layer activations before attention is computed. Unlike existing methods (H2O, StreamingLLM, Scissorhands) that make compression decisions based on attention scores computed during generation, KVCIS enables proactive compression at cache insertion time—determining how to store each token before paying the computational cost of attention. We discover a two-level importance structure in decoder-only transformers: the beginning-of-sequence (BOS) token acts as an "attention sink" receiving ~76% of attention, while the remaining ~24% is distributed across content tokens with 10-11× importance spread. A simple linear probe achieves R² = 0.998 overall and R² = 0.68–0.79 for discriminating among content tokens. Extensive validation across 3 model families (Llama, Mistral, Gemma), 8 layer depths, context lengths from 256 to 2048 tokens, and multiple downstream tasks demonstrates: 50% memory reduction with zero degradation on NarrativeQA (F1 = 0.064 matching baseline exactly), while uniform quantization degrades by 7.8% at the same compression ratio. KVCIS consistently achieves 5–8× better quality preservation than uniform quantization across all tested context lengths. The memory savings enable increased batch sizes and longer context support; the probe itself adds minimal overhead (~16KB direction vector, 0.06ms per token). This work extends activation-based probing from safety classification to inference optimization, demonstrating that intermediate-layer activations encode predictive signals about token importance for generation.
View details
Preview abstract
This paper demonstrates that artificial intelligence can accelerate mathematical discovery by autonomously solving an open problem in theoretical physics. We present a neuro-symbolic system, combining the Gemini Deep Think large language model with a systematic Tree Search (TS) framework and automated numerical feedback, that successfully derived novel, exact analytical solutions for the power spectrum of gravitational radiation emitted by cosmic strings. Specifically, the agent evaluated the core integral for arbitrary loop geometries, directly improving upon recent AI-assisted attempts that only yielded partial asymptotic solutions. To substantiate our methodological claims regarding AI-accelerated discovery and to ensure transparency, we detail system prompts, search constraints, and intermittent feedback loops that guided the model. The agent identified a suite of 6 different analytical methods, the most elegant of which expands the kernel in Gegenbauer polynomials to naturally absorb the integrand's singularities. The methods lead to an asymptotic result for at large that both agrees with numerical results and also connects to the continuous Feynman parameterization of Quantum Field Theory. We detail both the algorithmic methodology that enabled this discovery and the resulting mathematical derivations.
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
High-volume enterprise service organizations face a persistent challenge in transitioning from reactive support models to proactive, preventative ones. This paper introduces the Agentic Trend-to-Knowledge (ATK) methodology, a novel, autonomous framework designed to address this gap. The ATK methodology employs an AI agent that operates in a recurring, closed loop. It first uses a two-stage process for the autonomous thematic analysis of recent support cases to identify the most significant recurring issue. It then leverages Retrieval-Augmented Generation (RAG) to source relevant institutional knowledge. A key innovation is the agent's adaptive, bimodal response: if relevant knowledge is found, it drafts a proactive communication for human review; if a knowledge gap is detected, it autonomously creates a content creation task for the appropriate team. This transforms the agent from an automation tool into a proactive process owner that creates a virtuous cycle of continuous improvement for both case deflection and knowledge base quality. By automating the entire workflow from insight to action, the ATK framework provides a concrete methodology for shifting from a "human-in-the-loop" to a more strategic "human-on-the-loop" operational paradigm.
View details
Preview abstract
How many T gates are needed to approximate an arbitrary n-qubit quantum state to within
a given precision ϵ? Improving prior work of Low, Kliuchnikov and Schaeffer, we show that the
optimal asymptotic scaling is Θ(sqrt{2^n log(1/ε)} + log(1/ε)) if we allow an unlimited number of ancilla qubits. We also show that this is the optimal T-count for implementing an arbitrary
diagonal n-qubit unitary to within error ϵ. We describe an application to batched synthesis of
single-qubit unitaries: we can approximate a tensor product of m = O(log log(1/ϵ)) arbitrary
single-qubit unitaries to within error ϵ with the same asymptotic T-count as is required to
approximate just one single-qubit unitary.
View details
Preview abstract
Large language models have achieved remarkable capabilities across domains, yet mechanisms underlying sophisticated reasoning continue to be explored1,2. Recent reasoning-reinforced models, including OpenAI’s o-series and DeepSeek-r1, outperform other merely instruction-tuned models on complex cognitive tasks3,4, attributed to extended test-time computation through longer chains of thought5. Here we show that enhanced reasoning emerges not from extended computation alone, but from the systematic simulation of complex, multi-agent interactions—a society of thought—which enables the deliberate diversification and debate among internal cognitive perspectives characterized by distinct personality traits and domain expertise. Through quantitative analysis using classified outputs and mechanistic interpretability methods applied to reasoning traces6–8, we find that reasoning models like DeepSeek-r1 exhibit much greater perspective diversity than baseline models, activating broader and more conflict between heterogeneous personality- and expertise-related features during reasoning. This multi-agent structure manifests in conversational behaviors including question-answering sequences, perspective shifts, and reconciliation of conflicting views, as well as in socio-emotional roles that characterize back-and-forth conversation, which together account for over 60% of the accuracy advantage in reasoning tasks through both direct and indirect facilitation of cognitive strategies9,10. Controlled reinforcement learning experiments further reveal that priming models with conversational scaffolding—even when dialogues lead to incorrect solutions—substantially accelerates reasoning improvement compared to answer-only training. These findings indicate that the social organization of thought, rather than correctness alone, enables effective exploration of solution spaces. We suggest that reasoning models establish a computational parallel to collective intelligence in human groups11–13, where diversity enables superior problem-solving when systematically structured and suggest new opportunities for agent organization to harness the wisdom of crowds.
View details
FreshBrew: A Benchmark for Evaluating AI Agents on Java Code Migration
Victor May
Diganta Misra
Yanqi Luo
Anjali Sridhar
Justine Gehring
Silvio Soares Ribeiro Junior
2026
Preview abstract
AI coding assistants are rapidly becoming integral to modern software development. A key challenge in this space is the continual need to migrate and modernize codebases in response to evolving software ecosystems. Traditionally, such migrations have relied on rule-based systems and human intervention. With the advent of powerful large language models (LLMs), AI-driven agentic frameworks offer a promising alternative—but their effectiveness remains underexplored. In this paper, we introduce FreshBrew, a novel benchmark for evaluating AI-based agentic frameworks on project-level Java migrations. We benchmark several such frameworks, powered by state-of-the-art LLMs, and compare their performance against established rule-based tools. Our evaluation of AI agents on this benchmark of 228 repositories shows that the top-performing model, Gemini 2.5 Flash, can successfully migrate 56.5% of projects to JDK 17. Our empirical analysis reveals novel insights into the critical strengths and limitations of current agentic approaches, offering actionable insights into their real-world applicability. By releasing FreshBrew publicly upon acceptance, we aim to facilitate rigorous, reproducible evaluation and catalyze progress in AI-driven codebase modernization.
View details
Preview abstract
The fundamental theorem of statistical learning establishes that binary PAC learning is governed by a single parameter---the Vapnik-Chervonenkis ($\mathtt{VC}$) dimension---which controls both learnability and sample complexity. Extending this characterization to multiclass classification has long been challenging, since the early work of Natarajan in the late 80's that proposed the Natarajan dimension ($\mathtt{Nat}$) as a natural analogue of the VC dimension.
Daniely and Shalev-Shwartz (2014) introduced the $\mathtt{DS}$ dimension, later shown by Brukhim et al.\ (2022) to characterize multiclass \emph{learnability}.
Brukhim et al.\ (2022) also demonstrated that the Natarajan and $\mathtt{DS}$ dimensions can diverge arbitrarily, so that multiclass learning appears to be governed by $\mathtt{DS}$ rather than $\mathtt{Nat}$.
We show that the agnostic multiclass PAC sample complexity is in fact governed by \emph{two distinct dimensions}. Specifically, we prove nearly tight agnostic sample complexity bounds that, up to logarithmic factors, take the form
$$
\frac{\mathtt{DS}^{1.5}}{\epsilon} + \frac{\mathtt{Nat}}{\epsilon^2}
$$
where $\epsilon$ is the excess risk. This bound is tight up to a $\sqrt{\mathtt{DS}}$ factor in the first lower-order term, nearly matching known $\mathtt{Nat}/\epsilon^2$ and $\mathtt{DS}/\epsilon$ lower bounds.
The first term reflects the DS-controlled regime, while the second reveals that the Natarajan dimension still dictates asymptotic behavior for small $\epsilon$. Thus, unlike in binary or online classification---where a single dimension (VC or Littlestone) controls both phenomena---multiclass learning inherently involves \emph{two structural parameters}.
Our technical approach departs significantly from traditional agnostic learning methods based on uniform convergence or reductions-to-realizable techniques. A key ingredient is a novel online procedure, based on a self-adaptive multiplicative-weights algorithm which performs a label-space reduction. This approach may be of independent interest and find further applications.
View details
Preview abstract
Large language models (LLMs) are trained on web-scale corpora that exhibit steep power-law distributions, in which the distribution of knowledge is highly long-tailed, with most appearing infrequently. While scaling has improved average-case performance, persistent failures on low-frequency, domain-specific, cultural, and temporal knowledge remain poorly characterized. This paper develops a structured taxonomy and analysis of long-tail knowledge in large language models, synthesizing prior work across technical and sociotechnical perspectives.
We organize the literature along four complementary axes: how long-tail knowledge is defined, the mechanisms by which it is lost or distorted during training and inference, the technical interventions proposed to mitigate these failures, and the implications of these failures for fairness, accountability, transparency, and user trust. We further examine how existing evaluation practices obscure tail behavior and complicate accountability for rare but consequential failures. The paper concludes by identifying open challenges related to privacy, sustainability, and governance that constrain long-tail knowledge representation. Taken together, this paper provides a unifying conceptual framework for understanding how long-tail knowledge is defined, lost, evaluated, and manifested in deployed language model systems.
View details
