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 11377 publications
Preview abstract
While non-verbal behaviors and expressive movements are essential for natural human-robot interaction, existing methods often overlook a crucial element: the human’s internal cognitive state. Consequently, proactive multi-agent systems frequently interrupt humans at inopportune moments, leading to cognitive overload and decreased task performance. This paper introduces a framework for generating “cognitively aligned” multi-agent interactions, enhancing the ability of robotic systems to contextually defer communications during moments of high human mental workload. We present the design and implementation of a closed-loop architecture that explores the interplay between autonomous task execution and real-time neurophysiological focus. Utilizing a consumer-grade Brain-Computer Interface (BCI), our approach continuously monitors Electroencephalography (EEG) spectral band powers while a human performs a cognitive-load-inducing task. We propose a workload-driven pipeline where an HTTP-based signaling mechanism places a primary agent’s sensory inputs and audio outputs into a holding state upon detecting high cognitive load. This allows secondary agents to seamlessly process complex, delegated tasks in the background. Once the human’s cognitive state returns to a baseline, the primary agent releases the queued agent message. Our preliminary results demonstrate the feasibility of leveraging real-time signal processing, Large Language Models (LLMs), and physical robotic embodiments to create interrupt-aware, non-intrusive multi-agent systems.
View details
Preview abstract
As artificial intelligence (AI) is rapidly integrated into healthcare, ensuring that this innovation helps to combat health inequities requires engaging marginalized communities in health AI futuring. However, little research has examined Black populations’ perspectives on the use of AI in health contexts, despite the widespread health inequities they experience–inequities that are already perpetuated by AI. Addressing this research gap, through qualitative workshops with 18 Black adults, we characterize participants’ cautious optimism for health AI addressing structural well-being barriers (e.g., by providing second opinions that introduce fairness into an unjust healthcare system), and their concerns that AI will worsen health inequities (e.g., through health AI biases they deemed inevitable and the problematic reality of having to trust healthcare providers to use AI equitably). We advance health AI research by articulating previously-unreported health AI perspectives from a population experiencing significant health inequities, and presenting key considerations for future work.
View details
Reinforcement Learning with Discrete Diffusion Policies for Combinatorial Action-Spaces
Haitong Ma
Ofir Nabati
Na Li
Shie Mannor
Guy Tennenholtz
Proceedings of the 43rd International Conference on Machine Learning (ICML-26), Seoul, South Korea (2026)
Preview abstract
Reinforcement learning (RL) algorithms have achieved superhuman performance
on many sequential decision-making tasks, but often struggle in domains with
large, combinatorial action spaces. To address this, we introduce a practical and
stable algorithm for training discrete diffusion models to represent policies in
such environments. We formulate a policy mirror descent algorithm that enhances
training stability by reframing policy optimization as an inference problem, which
naturally aligns with the learning objective of discrete diffusion models. Through
extensive experiments on a suite of challenging benchmark tasks, we demonstrate
that our approach achieves significant improvements over existing methods in both
performance and sample efficiency. This work opens a promising new direction
for applying discrete diffusion models in RL to tackle long-standing challenges in
large-scale combinatorial action spaces.
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
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
JAXBench: Benchmarking Autonomous TPU Kernel Optimization
Vijay Janapa Reddi
Charles Hong
Arya Tschand
Julian Walker
Suvinay Subramanian
Shangkun Wang
Sethu Sankaran
Nina Cai
2026
Preview abstract
Evaluation benchmarks have driven progress in automated kernel optimization, yet existing suites target GPUs exclusively. We present JAXBench, a TPU-native benchmark for AI-generated kernel optimization on Google Cloud TPUs. JAXBench comprises 50 JAX workloads, including 17 production LLM operators extracted from architectures in the public MaxText library such as Llama-3.1, DeepSeek-V3, Mixtral, Mamba-2, and AlphaFold2, and 33 fused operator sequences adapted from KernelBench. Eight of the 17 production operators ship with hand-optimized Pallas TPU kernels from the public Tokamax library, whose block sizes we tune via grid search, establishing strong reference baselines. We evaluate one-shot generation, iterative coding agents, the same iterative loop with TPU documentation injected, and a TPU-enabled Autocomp configuration augmented with the same TPU-specific documentation. Across the full 50-benchmark suite with Gemini 3 Flash, best-of-N solves 13/50 benchmarks at 1.01x geomean and iterative refinement reaches 32/50 at 1.18x. Injecting TPU documentation lifts iterative refinement to 48/50 at 1.28x and raises per-sample correctness from 5.8% to 37.3%. Autocomp solves 45/50 but converts those correct kernels into 1.36x geomean with 76% of benchmarks beating XLA. On the 8 hand-tuned references, Autocomp reaches 1.60x against the XLA baseline, showing strong performance against the 2.08x Tokamax geomean but trailing on the specialized paged and ragged attention operators. A Gemini 3.1 Pro ablation lifts Autocomp to 49.1% per-sample correctness and a 3.13x geomean. High-quality TPU kernel generation remains open, and we release the benchmark, profiling harness, and baseline results to support reproducible research.
View details
The Ontic-Epistemic Distinction: Implications for General Intelligence
Master's Thesis (2026) (to appear)
Preview abstract
The current pursuit of robust machine intelligence is largely predicated on a substrate independent, computational functionalist view of cognition, where sufficiently complex computational processing is expected to eventually yield generalized reasoning. This paper explores the ontological distinctions between these computational frameworks and biological cognition, specifically how these differences impact the capacity for semantic understanding. By analyzing phenomena such as the "reversal curse" where models fail to generalize the symmetry in identity relations (A=B implies B=A), and performance on novel reasoning benchmarks (e.g., ARC-AGI), this paper examines whether current model limitations are transient artifacts of scale or indicative of a distinct architectural category. Integrating Stevan Harnad’s “symbol grounding problem” with Evan Thompson’s biological model of “intrinsic normativity,” I investigate whether robust general intelligence might require sense-making: a process distinct from information processing, whereby an agent’s internal states are causally coupled with its environment via survival or system-wide stakes which grounds symbols in meaning. Current Large Language Models (LLMs) appear to lack this intrinsic normativity, and consequently may operate primarily as epistemic instruments rather than ontic agents. By introducing the concept of “ontic grounding”, this paper presents a potential framework for distinguishing between the simulation of reasoning and true understanding, which could have implications for AI safety and governance.
View details
Preview abstract
In today's AI landscape, success depends not just on prompting large language models but on orchestrating them into intelligent systems that are scalable, compliant, and cost-effective. GenAI on Google Cloud is your hands-on guide to bridging that gap. Whether you're an ML engineer or an enterprise leader, this book offers a practical game plan for taking agentic systems from prototype to production.
Written by practitioners with deep experience in AgentOps, data engineering, and GenAI infrastructure, this guide takes you through real-world workflows from data prep and deployment to orchestration and integration. With concrete examples, field-tested frameworks, and honest insights, you'll learn how to build agentic systems that deliver measurable business value.
> Bridge the production gap that stalls 90% of vertical AI initiatives using systematic deployment frameworks
> Navigate AgentOps complexities through practical guidance on orchestration, evaluation, and responsible AI practices
> Build robust multimodal systems for text, images, and video using proven agent architectures
> Optimize for scale with strategies for cost management, performance tuning, and production monitoring
View details
VISTA: A Test-Time Self-Improving Video Generation Agent
Xuan Long Do
Hootan Nakhost
The IEEE/CVF Conference on Computer Vision and Pattern Recognition (to appear) (2026)
Preview abstract
Despite rapid advances in text-to-video (T2V) synthesis, generated video quality remains critically dependent on precise user prompts. Existing test-time optimization methods, successful in other domains, struggle with the multi-faceted nature of video. To address this, we introduce VISTA, a novel multi-agent system that autonomously refines prompts to improve video generation. VISTA operates in an iterative loop, first decomposing a user's idea into a structured temporal plan. After generation, the best video is identified through a robust pairwise tournament. This winning video is then critiqued by a trio of specialized agents focusing on visual, audio, and contextual fidelity. Finally, a reasoning agent synthesizes this feedback to introspectively rewrite and enhance the prompt for the next generation cycle. To rigorously evaluate our proposed approach, we introduce MovieGen-Bench, a new benchmark of diverse single- and multi-scene video generation tasks. Experiments show that while prior methods yield inconsistent gains, VISTA consistently improves video quality, achieving up to 60% pairwise win rate against state-of-the-art baselines. Human evaluators concur, preferring VISTA's outputs in 68% of comparisons.
View details
Agentic AI Infrastructure in Practice: Learn These Key Hurdles to Deploy Production AI Agents Efficiently
https://swisscognitive.ch/ (2026)
Preview abstract
The emergence of Agentic AI—autonomous systems capable of reasoning, decision-making, and multi-step execution—represents a paradigm shift in enterprise technology. Moving beyond simple generative tasks, these agents offer the potential to solve long-standing industry pain points, with over 90% of enterprises planning integration within the next three years. However, the transition from successful proof-of-concept (PoC) to a resilient, production-grade system presents significant hurdles.
This article categorizes these challenges into three primary domains:
Technical and Engineering Hurdles: Issues such as "entangled workflows" that complicate debugging, the struggle to maintain output quality and mitigate hallucinations, and the unpredictability caused by shifting underlying models or data sources.
People, Process, and Ecosystem Hurdles: The high operational costs and unclear ROI of large models, the necessity of a new "Agent Ops" skillset, the complexity of integrating agents with disparate enterprise systems, and a rapidly evolving regulatory landscape.
The Pace of Change and Security risks: The technical debt incurred by shifting software frameworks and the expanded attack surface created by autonomous agents.
The article concludes that successful deployment requires a shift from informal "vibe-testing" to rigorous engineering discipline. By adopting code-first frameworks, establishing robust evaluation metrics (KPIs), and prioritizing functional deployment over theoretical optimization, organizations can effectively manage the lifecycle of Agentic AI and realize its transformative business value.
View details
Dynamic Cogeneration of Bug Reproduction Test in Agentic Program Repair
Michele Tufano
José Cambronero
Renyao Wei
Grant Uy
34th ACM International Conference on the Foundations of Software Engineering (FSE) (2026)
Preview abstract
Bug Reproduction Tests (BRTs) have been used in many Automated Program Repair (APR) systems, primarily for validating fixes and aiding fix generation. In practice, when developers submit a patch, they often implement the BRT alongside the fix. Our experience deploying agentic APR reveals that developers desire a BRT within AI-generated patches to increase their confidence. However, canonical APR systems tend to generate BRTs and fixes separately, and focus on producing only the fix in the final patch. In this paper, we study agentic APR in the context of cogeneration, where the APR agent is instructed to generate both a fix and a BRT in the same patch. We evaluate the effectiveness of different cogeneration strategies on 120 human-reported bugs at Google and characterize different cogeneration strategies by their influence on APR agent behavior. We develop and evaluate patch selectors that account for test change to select patches with plausible fixes (and plausible BRTs). Finally, we analyze the root causes of failed cogeneration trajectories. We show that cogeneration allows the APR agent to generate BRTs for at least as many bugs as a dedicated BRT agent, without compromising the generation rate of plausible fixes, thereby reducing engineering effort in maintaining and coordinating separate generation pipelines for fix and BRT at scale.
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
Diffusion Controller: Framework, Algorithms and Parameterization
Tong Yang
Moonkyung Ryu
Guy Tennenholtz
Yuejie Chi
Proceedings of the 43rd International Conference on Machine Learning (ICML-26), Seoul, South Korea (2026)
Preview abstract
Controllable generation with diffusion models is often treated as a collection of heuristics rather than a unified optimization problem. We propose a principled control formulation by viewing the diffusion reverse process as an instance of a (generalized) linearly-solvable Markov decision process (LS-MDP). This perspective turns controllable generation into regularized optimal control around a pretrained diffusion policy, yielding tractable objectives and algorithmic updates.
Under this framework, we study two practical finetuning regimes. When paired target data are available, we obtain a supervised finetuning (SFT) objective. When only a terminal reward model is available, we derive reinforcement-learning finetuning (RLFT) methods from the LS-MDP solution structure, including (i) a reward-weighted regression loss and (ii) a policy-gradient approach (with standard extensions such as PPO).
Crucially, the LS-MDP optimality conditions imply an explicit relationship between the optimal and pretrained score functions. We leverage this to derive a new score-function parameterization that isolates the control signal and enables “gray-box” finetuning with substantially fewer trainable parameters. Experiments across SFT and RLFT show this parameterization improves over existing finetuning baselines while achieving stronger sample/parameter efficiency.
View details
Preview abstract
The management of a hybrid workforce comprising human and autonomous computational agents may be challenged by the use of separate systems for human capital and software assets, which can create a governance gap. A system can provide a unified framework for managing a hybrid workforce. For example, the system may utilize a labor service mesh to analyze and route tasks to either a human intent tier or an agentic execution tier. A potential principle of the system is structural symmetry, where computational agents can be assigned digital identities and managed through a lifecycle process that may parallel human resource functions, such as onboarding, performance evaluation, and structured offboarding. This integrated approach can facilitate a unified system of record and governance model for an organization's intelligence capacity.
View details
Towards AI as a Collaborative Partner: A Taxonomy of AI Agent Behavior in Software Engineering
Proceedings of the 3rd ACM International Conference on AI-Powered Software (AIware '26), ACM, Montreal, QC, Canada (2026) (to appear)
Preview abstract
The ongoing transition of Large Language Models (LLMs) in software engineering from one-shot code generators into agentic partners requires a shift in how we define and measure success. While models are becoming more capable, the industry lacks a clear understanding of the behavioral norms that make an interactive software engineering (SWE) agent effective in collaborative software development in the enterprise. This work addresses this gap by presenting a taxonomy of desirable SWE agent behaviors, synthesized from 91 sets of developer-defined rules for SWE agents and validated through interviewing 15 experienced professional developers. In this taxonomy, we identify four core expectations: Adhere to Standards and Processes, Ensure Code Quality and Reliability, Solve Problems Effectively, and Collaborate with the Developer. These findings offer a concrete vocabulary for aligning SWE agent behavior with developer preferences, enabling researchers and practitioners to move beyond correctness-only benchmarks and start designing evaluations that reflect the socio-technical nature of professional software development in enterprises.
View details
