Tomas Pfister
Tomas Pfister is the Head of Cloud AI Research. He came to Google from Apple where he cofounded Apple's central AI research group and published Apple’s first research paper that won the Best Paper Award at CVPR’17. Tomas’ key scientific achievements have been proposing a method to improve the realism of synthetic images; developing the first automated method to detect facial micro-expressions; and inventing a new way for neural networks to exploit spatiotemporal structure. He is currently exploring learning from small amount of labeled data (using techniques such as generative models, few-shot learning, transfer learning) and explainability/interpretability of deep learning models, and is particularly excited about the potential of AI in healthcare & education. His research has laid the foundation for several applications such as Face ID in iPhone X, autonomous driving, human pose estimation, detecting facial micro-expressions & translating sign language. Tomas did his PhD in deep learning with Prof Andrew Zisserman at Oxford University and bachelor’s degree in computer science at Cambridge University. He is the recipient of the Forbes 30 Under 30 award, and has received over 40 research awards, including 3 best paper awards, with numerous publications in top AI research venues. His work has been frequently featured in mainstream media, including Forbes, BusinessInsider & Wired.
Authored Publications
Sort By
PI-Hunter: Automated Red-Teaming for Exposing and Localizing Prompt Injections
Pengfei He
Vishesh Sharma
Ash Fox
George Lee
Jiliang Tang
2026
Preview abstract
Large Language Models (LLMs) are rapidly evolving into agentic systems that interact with external tools and dynamic environments, but this also introduces severe security risks. In particular, indirect prompt injection attacks can compromise agents through malicious instructions hidden in external sources such as web pages, emails, and retrieved documents. Existing defenses are largely reactive, while current automated red-teaming methods mainly optimize attack success rather than systematically uncovering hidden vulnerabilities within the agent pipeline. In this work, we propose PI-Hunter, an automated agentic red-teaming framework that shifts the focus from attack optimization to vulnerability exposure. By combining static attack-surface analysis, source-aware seeding, trajectory evaluation, and feedback-guided exploration, PI-Hunter proactively discovers vulnerable ingestion paths and localizes how malicious instructions propagate through agent reasoning. Extensive experiments across multiple benchmarks, agent architectures, attacks, and defenses show that \method~substantially improves vulnerability exposure and attack-surface coverage compared with existing automated red-teaming baselines, while remaining effective even under strong prompt injection defenses.
View details
Co-RedTeam: Orchestrated Security Discovery and Exploitation with LLM Agents
Pengfei He
Ash Fox
Stefan Friedli
Daniel Fabian
Burak Gokturk
Jiliang Tang
ICML 2026
Preview abstract
Large language models (LLMs) have shown promise in assisting cybersecurity tasks, yet existing approaches struggle with automatic vulnerability discovery and exploitation due to limited interaction, weak execution grounding, and a lack of experience reuse. We propose Code-RedTeam, a security-aware multi-agent framework designed to mirror real-world red-teaming workflows by integrating security-domain knowledge, code-aware analysis, execution-grounded iterative reasoning, and long-term memory. Code-RedTeam decomposes vulnerability analysis into coordinated discovery and exploitation stages, enabling agents to plan, execute, validate, and refine actions based on real execution feedback while learning from prior trajectories. Extensive evaluations on challenging security benchmarks demonstrate that Code-RedTeam consistently outperforms strong baselines across diverse backbone models, achieving over 60% attack success rate in vulnerability exploitation and up to 10% absolute improvement in vulnerability detection. Ablation and iteration studies further confirm the critical role of execution feedback, structured interaction, and memory for building robust and generalizable cybersecurity agents.
View details
Budget-Aware Tool Use Enables Effective Agent Scaling
Tengxiao Liu
Jin Miao
I-Hung Hsu
Samira Daruki
COLM 2026 (2026)
Preview abstract
Scaling test-time computation improves performance across different tasks on large language models (LLMs), yet mainstream scaling methods remain challenging for tool-augmented LLM agents. Sequential scaling tends to yield shallow tool use and under-exploration, whereas parallel scaling inflates cost through repeated tool calls. The dual costs of tokens and tool calls further complicate cost accounting and hinder fair comparison. In this work, we study the test-time scaling of widely used, tool-reliant search agents under resource constraints, analyzing performance with a unified cost metric that incorporates both tokens and tool calls. To this end, we propose Cost-effective Agent Test-Time Scaling (CATS), a budget-aware framework designed to support more cost-effective scaling by guiding resource allocation between sequential and parallel exploration. Experiments across search-intensive benchmarks show that CATS produces more favorable scaling curves, attaining higher accuracy with fewer tool calls and lower overall cost. Our work introduces a cost-conscious design for agent test-time scaling and contributes empirical insights that enable a more transparent and principled understanding of scaling in tool-augmented agents.
View details
Preview abstract
Temporally consistent long video generation remains a fundamental challenge. Existing methods suffer from feature drift, where entities and environments gradually change unintentionally, or content collapse, where narratives fail to progress meaningfully. We introduce A${^2}$RD, an agentic autoregressive video generation architecture that decouples creative synthesis from consistency by modeling consistency as a test-time objective. A${^2}$RD features segment-by-segment generation augmented with a novel Multimodal Video Memory (\memory{}) that tracks segment contexts and dynamics and Test-Time Scaling algorithms that verify and refine generation. For each segment, it operates in a Retrieve--Synthesize--Refine--Update (RSRU) loop: the agent retrieves relevant contexts, determines the segment generation mode (extrapolation or interpolation) adaptively, synthesizes boundary frames then video segment with refinements applied at both frame and video levels, and updates \memory{} for subsequent generation. We further develop LVbench-C, a challenging benchmark measuring long-horizon entities and environments evolving in non-linear transitions. Extensive experiments on public and LVbench-C benchmarks across one-, three-, and five-minute video generation demonstrate that A${^2}$RD generates significantly more consistent, meaningful videos than existing baselines. Human evaluations confirm strong consistency in characters, objects, and environments, with smooth motion and meaningful narrative progression.
View details
LiSA: Lifelong Safety Adaptation via Conservative Policy Induction
Minbeom Kim
Phil Wallis
Bharath Chandrasekhar
Kyomin Jung
NeurIPS (2026)
Preview abstract
Large language model agents increasingly act in deployment environments where failures are contextual, user-specific, and costly. In such settings, a \emph{static general-purpose guardrail is often insufficient}: whether an action should be allowed may depend on local privacy norms, organizational rules, or evolving user expectations that are difficult to enumerate fully in advance. We study \emph{lifelong deployment-time guardrail adaptation}, where a fixed base guardrail improves over time from sparse, noisy user-reported failures without repeated fine-tuning. We propose a conservative policy induction framework organized as an online--offline loop. Online, the deployed guardrail uses structured policy memory to guide runtime decisions. Offline, newly accumulated reports are converted into reusable policy items and folded back into memory through periodic refresh. The method combines three ingredients: \emph{broad policy abstraction} for sparse failure generalization, \emph{conflict-aware local policies} for mixed-label regions where broad reuse becomes too coarse, and \emph{confidence-gated reuse} based on conservative posterior lower bounds so that weakly supported memory does not influence inference too early. Across PrivacyLens+, ConFaide+, and AgentHarm, the resulting system consistently improves over a lightweight base guardrail and strong memory-based baselines in sparse-feedback regimes, remains robust to noisy feedback, traces a better cost--performance frontier than scaling the base model alone, and jointly reduces over-refusal and over-acceptance without an explicit balance knob.
View details
Preview abstract
Understanding long visual documents, where information is distributed across extensive pages of text and graphics, remains a critical challenge for modern Vision-Language Models (VLMs). This difficulty is rooted in two fundamental obstacles: poor evidence localization and a high tendency for model hallucination. To address these issues, we propose DocLens, a multi-agent framework that decomposes the task into two specialized stages.
First, a Lens Module leverages document parsing tools for fine-grained, hierarchical evidence localization at both the page and element level. Second, a Reasoning Module employs a sampling-adjudication mechanism to systematically analyze the localized evidence, mitigating hallucination and synthesizing reliable answers.
Paired with Gemini-2.5-Pro, DocLens achieves state-of-the-art performance on MMLongBench-Doc and FinRAGBench-V, even surpassing human experts. Furthermore, our framework offers a highly cost-effective variant that delivers comparable performance to strong baselines at a five-fold reduction in cost.
View details
PaperBanana: Automating Academic Illustration for AI Scientists
Dawei Zhu
Sujian Li
International Conference on Machine Learning (2026)
Preview abstract
While AI scientists increasingly automate research tasks through advanced language models, generating publication-ready illustrations remains a labor-intensive bottleneck in the scientific workflow. To lift this burden, we introduce PaperBanana, an agentic framework for automated generation of publication-ready academic diagrams. Powered by Nano-Banana-Pro and Gemini-3-Pro, PaperBanana orchestrates a team of specialized agents to retrieve reference examples, devise detailed plans for content and style, render the image, and perform iterative refinement based on self-critique. To rigorously evaluate our framework and address the absence of dedicated benchmarks for automated academic illustration, we introduce PaperBananaBench, comprising 292 test cases for methodology diagrams curated from NeurIPS 2025 publications.
Comprehensive experiments demonstrate that PaperBanana consistently outperforms vanilla Nano-Banana-Pro across all four dimensions—faithfulness, conciseness, readability, and aesthetics—achieving human-level performance. We further show that PaperBanana seamlessly extends to statistical plots through targeted adaptations. Collectively, PaperBanana enables AI scientists to fully automate the generation of publication-ready academic illustrations.
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
TFRBench: A Reasoning Benchmark for Evaluating Forecasting Systems
MD ATIK AHAMED
Qiang Cheng
Forty-third International Conference on Machine Learning (2026)
Preview abstract
Time-series forecasting has traditionally been evaluated solely on numerical accuracy, treating models as "black boxes'' that fail to capture the underlying reasoning. To address this gap, we introduce TFRBench, a novel benchmark designed to evaluate the reasoning capabilities of forecasting systems alongside their numerical accuracy. Unlike existing benchmarks, TFRBench requires models to generate verifiable natural language reasoning by analyzing cross-channel dependencies, identifying strategic trends, and justifying significant events using external context. To construct this benchmark, we propose a systematic multi-agent framework comprising Reasoning, Search, Verifier, Forecasting, and Summary agents. Our benchmark spans five diverse domains including Energy, Sales, Web/CloudOps, Transportation, and Finance, covering 10 distinct datasets. Qualitative evaluation confirms that our generated reasoning is highly faithful and effective; specifically, Large Language Models (LLMs) prompted with our generated reasoning demonstrate significantly improved forecasting accuracy compared to direct forecasting with LLMs. Conversely, benchmarking experiments reveal that off-the-shelf LLMs consistently struggle with both reasoning (shows lower LLM-as-Judge scores) and direct numerical forecasting (MAE and MASE), frequently failing to capture domain-specific dynamics. TFRBench thus establishes a new standard for interpretable, reasoning-based evaluation in time-series forecasting.
View details
VeriGuard: Enhancing LLM Agent Safety via Verified Code Generation
Dj Dvijotham
Mirko Montanari
ACL 2026
Preview abstract
Artificial intelligence is rapidly evolving, marked by the emergence of Large Language Model (LLM) agents – systems capable of complex reasoning, planning, and interaction with digital and physical environments. These agents, powered by advancements in LLMs, demonstrate remarkable capabilities across diverse domains, including finance, healthcare, web navigation, software development, and daily task assistance. Unlike traditional AI systems, LLM agents can perceive their surroundings, formulate multi-step plans, utilize external tools and APIs, access memory or knowledge bases, and execute actions to achieve specified goals. This ability to act upon the world, however, introduces significant safety and security challenges.
The safety paradigms developed for traditional LLMs, primarily focused on mitigating harmful textual outputs (e.g., toxicity, bias), are insufficient for safeguarding LLM agents. Agents interacting with dynamic environments and executing actions present a broader attack surface and new categories of risk. These include performing unsafe operations, violating privacy constraints through improper data handling or access control failures, deviating from user objectives (task misalignment), and susceptibility to novel manipulation techniques like indirect prompt injection and memory poisoning. Ensuring the trustworthy operation of these powerful agents is paramount, especially as they are integrated into high-stakes applications. To address this critical challenge, we introduce VeriGuard, a novel framework designed to enhance the safety and reliability of LLM agents by interactively verifying their policies and the actions. VeriGuard integrates a verification module that intercepts code-based actions proposed by the agent. In the first step, VeriGuard will generates and verifies the policies. The policies are rigorously checked against a set of predefined safety and security specifications Then each action will be verified to make sure it will align with the agent specification. This interactive verification loop ensures that the agent's behavior remains within safe operational bounds, effectively preventing the execution of harmful or unintended operations. By verifying each step, VeriGuard provides a robust safeguard, substantially improving the trustworthiness of LLM agents in complex, real-world environments.
View details
