Blaise Aguera y Arcas
Blaise Agüera y Arcas is a VP and Fellow at Google, where he is the CTO of Technology & Society and founder of Paradigms of Intelligence (Pi). Pi is an organization working on basic research in AI and related fields, especially the foundations of neural computing, active inference, sociality, evolution, and Artificial Life.
In 2008, Blaise was awarded MIT’s TR35 prize. During his tenure at Google, Blaise has innovated on-device machine learning for Android and Pixel; invented Federated Learning, an approach to decentralized model training that avoids sharing private data; and founded the Artists + Machine Intelligence program.
Prior to Google, Blaise was a Distinguished Engineer at Microsoft, where he worked in a variety of roles, from inventor to strategist, and led teams with strengths in experience design, prototyping, machine vision, augmented reality, wearable computing and mapping.
Authored Publications
Sort By
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
Multi-agent cooperation through in-context co-player inference
Rajai Nasser
Alexander Meulemans
Marissa Weis
João Sacramento
Maciej Wołczyk
Rif A. Saurous
2026
Preview abstract
Achieving cooperation among self-interested agents remains a fundamental challenge in multi-agent reinforcement learning. Promising recent work has shown that cooperation can be established between ``learning-aware'' agents that explicitly account for and shape the learning dynamics of their co-players. However, existing approaches typically rely on hardcoded, often inconsistent, assumptions about co-player learning rules or enforce a strict separation between ``naive learners'' updating on fast timescales and ``meta-learners'' observing these updates. Here, we demonstrate that the in-context learning capabilities of sequence models allow for co-player learning awareness without requiring hardcoded assumptions or explicit timescale separation. We show that training sequence model agents against a diverse distribution of co-players naturally induces \textit{in-context best-response} strategies, effectively functioning as learning algorithms on the fast intra-episode timescale. We find that the cooperative mechanism identified in prior work—where vulnerability to extortion drives mutual shaping—emerges naturally in this setting: in-context adaptation renders agents vulnerable to extortion, and the resulting mutual pressure to shape the opponent's in-context learning dynamics resolves into the learning of cooperative behavior.
View details
A game theory for foundation models shows new paths to rational cooperation
Rajai Nasser
Marcus Hutter
Guillaume Lajoie
Alexander Meulemans
Seijin Kobayashi
Angelika Steger
Marissa Weis
João Sacramento
James Manyika
Blake Richardson
Maciej Wołczyk
Roberta Rocca
Rif A. Saurous
arXiv (2026)
Preview abstract
As autonomous agents powered by foundation models are increasingly integrated into social and economic systems, understanding the principles governing their collective behavior is essential for ensuring safety and cooperation. Classical game theory, the dominant framework for modeling rational interaction, is built upon the assumption of `decoupled agency,' where agents treat their own decision-making as independent of the environment and other actors. Modern AI agents, however, jointly predict their own future actions alongside external observations. Here, we report a striking finding: when interacting in stylized social dilemmas, foundation model agents engaging in optimal planning consistently converge to stable cooperation, directly contradicting classical game-theoretic predictions of mutual defection. To understand this phenomenon, we introduce the `embedded Bayesian agent,' a theoretical model for foundation model agents. By shifting from decoupled to embedded agency, these agents model themselves as part of the universe they inhabit, maintaining epistemic uncertainty about their own decision-making algorithms. We show that by inferring whether others are behaviorally similar, an embedded agent treats its own deliberation during planning as evidence: a decision to cooperate predicts a similar decision by a similar partner. We formalize this mechanism of similarity inference through the `embedded equilibrium,' a novel solution concept replacing the Nash equilibrium to provide a foundational game theory for the social behavior of modern AI agents.
View details
LLMs achieve adult human performance on higher-order theory of mind tasks
Oliver Siy
Geoff Keeling
Benjamin Barnett
Michael McKibben
Tatenda Kanyere
Robin I.M. Dunbar
Frontiers in Human Neuroscience (2025)
Preview abstract
This paper examines the extent to which large language models (LLMs) are able to perform tasks which require higher-order theory of mind (ToM)–the human ability to reason about multiple mental and emotional states in a recursive manner (e.g. I think that you believe that she knows). This paper builds on prior work by introducing a handwritten test suite – Multi-Order Theory of Mind Q&A – and using it to compare the performance of five LLMs of varying sizes and training paradigms to a newly gathered adult human benchmark. We find that GPT-4 and Flan-PaLM reach adult-level and near adult-level performance on our ToM tasks overall, and that GPT-4 exceeds adult performance on 6th order inferences. Our results suggest that there is an interplay between model size and finetuning for higher-order ToM performance, and that the linguistic abilities of large models may support more complex ToM inferences. Given the important role that higher-order ToM plays in group social interaction and relationships, these findings have significant implications for the development of a broad range of social, educational and assistive LLM applications.
View details
Toward expert-level medical question answering with large language models
Karan Singhal
Juro Gottweis
Le Hou
Kevin Clark
Heather Cole-Lewis
Amy Wang
Sami Lachgar
Philip Mansfield
Sushant Prakash
Bradley Green
Ewa Dominowska
Nenad Tomašev
Renee Wong
Sara Mahdavi
Joelle Barral
Nature Medicine (2025)
Preview abstract
Large language models (LLMs) have shown promise in medical question answering, with Med-PaLM being the first to exceed a ‘passing’ score in United States Medical Licensing Examination style questions. However, challenges remain in long-form medical question answering and handling real-world workflows. Here, we present Med-PaLM 2, which bridges these gaps with a combination of base LLM improvements, medical domain fine-tuning and new strategies for improving reasoning and grounding through ensemble refinement and chain of retrieval. Med-PaLM 2 scores up to 86.5% on the MedQA dataset, improving upon Med-PaLM by over 19%, and demonstrates dramatic performance increases across MedMCQA, PubMedQA and MMLU clinical topics datasets. Our detailed human evaluations framework shows that physicians prefer Med-PaLM 2 answers to those from other physicians on eight of nine clinical axes. Med-PaLM 2 also demonstrates significant improvements over its predecessor across all evaluation metrics, particularly on new adversarial datasets designed to probe LLM limitations (P < 0.001). In a pilot study using real-world medical questions, specialists preferred Med-PaLM 2 answers to generalist physician answers 65% of the time. While specialist answers were still preferred overall, both specialists and generalists rated Med-PaLM 2 to be as safe as physician answers, demonstrating its growing potential in real-world medical applications.
View details
MesaNet: Sequence Modelling by Locally Optimal test-Time Training
Songlin Yang
Razvan Pascanu
Alexander Meulemans
Seijin Kobayashi
Yanick Schimpf
João Sacramento
Maximilian Schlegel
Luca Versari
Oliver Sieberling
Rif A. Saurous
2025
Preview abstract
Sequence modeling is currently dominated by causal transformer architectures that
use softmax self-attention. Although widely adopted, transformers require scaling
memory and compute linearly during inference. A recent stream of work linearized
the softmax operation, resulting in powerful recurrent neural network (RNN)
models with constant memory and compute costs such as DeltaNet, Mamba or
xLSTM. These models can be unified by noting that their recurrent layer dynamics
can all be derived from an in-context regression objective, approximately optimized
through an online learning rule. Here, we join this line of work and introduce a
numerically stable, chunkwise parallelizable version of the recently proposed Mesa
layer (von Oswald et al., 2024), which could only run sequentially in time and was
therefore not scalable. This layer again stems from an in-context loss, but which is
now minimized to optimality at every time point using a fast conjugate gradient
solver. Through an extensive suite of experiments study up to the billion-parameter
scale, we show that optimal test-time training enables reaching lower language
modeling perplexity and higher downstream benchmark performance than previous
RNNs, especially on tasks requiring long context understanding. This performance
gain comes at the cost of additional flops spent during inference time. Our results
are therefore intriguingly related to recent trends of increasing test-time compute to
improve performance – here by spending compute to solve sequential optimization
problems within the neural network itself.
View details
Towards Generalist Biomedical AI
Danny Driess
Andrew Carroll
Chuck Lau
Ryutaro Tanno
Ira Ktena
Basil Mustafa
Aakanksha Chowdhery
Simon Kornblith
Philip Mansfield
Sushant Prakash
Renee Wong
Sara Mahdavi
Bradley Green
Ewa Dominowska
Joelle Barral
Karan Singhal
Pete Florence
NEJM AI (2024)
Preview abstract
BACKGROUND: Medicine is inherently multimodal, requiring the simultaneous interpretation and integration of insights between many data modalities spanning text, imaging, genomics, and more. Generalist biomedical artificial intelligence systems that flexibly encode, integrate, and interpret these data might better enable impactful applications ranging from scientific discovery to care delivery.
METHODS: To catalyze development of these models, we curated MultiMedBench, a new multimodal biomedical benchmark. MultiMedBench encompasses 14 diverse tasks, such as medical question answering, mammography and dermatology image interpretation, radiology report generation and summarization, and genomic variant calling. We then introduced Med-PaLM Multimodal (Med-PaLM M), our proof of concept for a generalist biomedical AI system that flexibly encodes and interprets biomedical data including clinical language, imaging, and genomics with the same set of model weights. To further probe the capabilities and limitations of Med-PaLM M, we conducted a radiologist evaluation of model-generated (and human) chest x-ray reports.
RESULTS: We observed encouraging performance across model scales. Med-PaLM M reached performance competitive with or exceeding the state of the art on all MultiMedBench tasks, often surpassing specialist models by a wide margin. In a side-by-side ranking on 246 retrospective chest x-rays, clinicians expressed a pairwise preference for Med-PaLM Multimodal reports over those produced by radiologists in up to 40.50% of cases, suggesting potential clinical utility.
CONCLUSIONS: Although considerable work is needed to validate these models in real-world cases and understand if cross-modality generalization is possible, our results represent a milestone toward the development of generalist biomedical artificial intelligence systems.
View details
Computational Life: How Well-formed, Self-replicating Programs Emerge from Simple Interaction
Ettore Randazzo
Luca Versari
Jyrki Alakuijala
James Evans
arxiv (2024)
Preview abstract
The fields of Origin of Life and Artificial Life both question what life is and how it emerges from a distinct set of "pre-life" dynamics. One common feature of most substrates where life emerges is a marked shift in dynamics when self-replication appears. While there are some hypotheses regarding how self-replicators arose in nature, we know very little about the general dynamics, computational principles, and necessary conditions for self-replicators to emerge. This is especially true on "computational substrates" where interactions involve logical, mathematical, or programming rules. In this paper we take a step towards understanding how self-replicators arise by studying several computational substrates based on various simple programming languages and machine instruction sets. We show that when random, non self-replicating programs are placed in an environment lacking any explicit fitness landscape, self-replicators tend to arise. We demonstrate how this occurs due to random interactions and self-modification, and can happen with and without background random mutations. We also show how increasingly complex dynamics continue to emerge following the rise of self-replicators. Finally, we show a counterexample of a minimalistic programming language where self-replicators are possible, but so far have not been observed to arise.
View details
Preview abstract
Cascades are a common type of machine learning system where a larger, remote model can be queried if a local model is not able to handle a user’s query by itself. They are becoming an increasingly popular choice of a design for Large Language Models (LLMs) serving stacks due to their ability to preserve task performance, while dramatically reducing inference costs. However, applying cascade systems in situations where the local model has access to sensitive data constitutes a significant privacy risk for users since any such data could be forwarded to the remote model. In this work, we show the feasibility of applying cascade systems in such setups, equipping the local model with privacy-preserving techniques that reduce the risk of leaking private information when querying the remote model. To analyze the privacy of such a setup, we introduce a novel privacy measure that quantifies sensitive information leakage. We then propose a system that leverages the recently introduced social learning paradigm in which LLMs collaboratively learn from each other by exchanging natural language and demonstrate on several datasets that our methods minimize the privacy loss while at the same time improving task performance compared to a non-cascade baseline.
View details
Large Language Models Encode Clinical Knowledge
Karan Singhal
Sara Mahdavi
Jason Wei
Hyung Won Chung
Nathan Scales
Ajay Tanwani
Heather Cole-Lewis
Perry Payne
Martin Seneviratne
Paul Gamble
Christopher Kelly
Abubakr Abdelrazig Hassan Babiker
Nathanael Schaerli
Aakanksha Chowdhery
Philip Mansfield
Dina Demner-Fushman
Katherine Chou
Juraj Gottweis
Nenad Tomašev
Alvin Rajkomar
Joelle Barral
Nature (2023)
Preview abstract
Large language models (LLMs) have demonstrated impressive capabilities, but the bar for clinical applications is high. Attempts to assess the clinical knowledge of models typically rely on automated evaluations based on limited benchmarks. Here, to address these limitations, we present MultiMedQA, a benchmark combining six existing medical question answering datasets spanning professional medicine, research and consumer queries and a new dataset of medical questions searched online, HealthSearchQA. We propose a human evaluation framework for model answers along multiple axes including factuality, comprehension, reasoning, possible harm and bias. In addition, we evaluate Pathways Language Model (PaLM, a 540-billion parameter LLM) and its instruction-tuned variant, Flan-PaLM on MultiMedQA. Using a combination of prompting strategies, Flan-PaLM achieves state-of-the-art accuracy on every MultiMedQA multiple-choice dataset (MedQA, MedMCQA, PubMedQA and Measuring Massive Multitask Language Understanding (MMLU) clinical topics), including 67.6% accuracy on MedQA (US Medical Licensing Exam-style questions), surpassing the prior state of the art by more than 17%. However, human evaluation reveals key gaps. To resolve this, we introduce instruction prompt tuning, a parameter-efficient approach for aligning LLMs to new domains using a few exemplars. The resulting model, Med-PaLM, performs encouragingly, but remains inferior to clinicians. We show that comprehension, knowledge recall and reasoning improve with model scale and instruction prompt tuning, suggesting the potential utility of LLMs in medicine. Our human evaluations reveal limitations of today’s models, reinforcing the importance of both evaluation frameworks and method development in creating safe, helpful LLMs for clinical applications.
View details
