Science advances by formulating and testing hypotheses, collecting and analyzing data, and drawing conclusions—yet much of a scientist’s time is spent in tasks such as coding analyses, writing and revising text, reviewing the literature, and learning new concepts. Can recent advances in AI help reclaim some of that time? In this talk, I will show how large language models and AI agents may help scientists with these tasks. I will present Denario, a complex, publicly available, multi-AI-agent system designed to function as a research assistant. Developed and evaluated by a diverse team of scientists, mathematicians, and philosophers, Denario is an interdisciplinary tool capable of generating ideas, searching the literature, developing research plans, writing and executing code, crafting plots, drafting and reviewing scientific papers. I will present and discuss some of the documents generated by Denario in disciplines such as astrophysics, biology, biophysics, biomedical informatics, chemistry, machine learning, material science, mathematical physics, medicine, neuroscience, planetary physics, and quantum physics. I'll finalize discussing the benefits and perils of automating (parts) of science.

Traditionally, communication has been treated as the reliable transfer of data between entities, with the goal of reproducing messages as faithfully as possible, independent of meaning or downstream utility. While effective in classical systems, this data-centric paradigm becomes limiting in modern intelligent environments involving distributed decision-making, multi-agent interaction, and large-scale coordination, where what matters is not only what is transmitted, but what is meant and how it is used. In this talk, we focus on semantic communication protocols as a general paradigm for information exchange in intelligent systems, with applications spanning wireless networks and beyond. These protocols go beyond symbol fidelity to enable meaning-aware, task-driven interaction under strict resource constraints, operating over shared latent structure rather than raw data. We discuss how such protocols can be learned and optimized, and how they naturally extend across a broad range of applications, including agentic AI systems, where coordination, reasoning, and adaptation emerge through interaction rather than explicit design. We present a unified perspective grounded in artificial intelligence, causal reasoning, game theory, and theory of mind, where communication, inference, and control are jointly learned processes. Within this framework, we highlight three directions: emergent semantic communication via signaling games and neuro-symbolic reasoning, where efficient protocols arise from interaction; causal semantic communication in digital twin and wireless network systems that exploits invariances for robust control; and multi-agent world models with metacognitive reasoning, where agents infer and adapt to others’ beliefs and intentions without explicit communication, leading to emergent collective behavior. Collectively, these results point toward a shift from engineered communication protocols to learned, adaptive, and emergent semantic interaction systems, where meaning and reasoning become first-class design principles.

AI is rapidly transitioning from a passive analytical assistant to an active, self-improving partner in scientific discovery. In the material world, this shift means developing systems that not only recognize patterns but also reason, hypothesize, and autonomously explore new ideas for design, discovery and manufacturing. This talk presents emerging approaches toward 'superintelligent' discovery engines -integrating reinforcement learning, graph-based reasoning, and physics-informed neural architectures with generative models capable of cross-domain synthesis. We explore multi-agent swarm systems inspired by collective intelligence in nature, enabling continuous self-evolution as they solve problems. Case studies from materials science, engineering and biology illustrate how these systems can uncover hidden structure-property relationships, design novel materials, and accelerate innovations in biology and beyond. These advances chart a path toward AI that actively expands the boundaries of human knowledge.

A central challenge in agentic scientific computing is equipping agents with world models that can accurately simulate the physical consequences of actions. Recent advances in video generation models have sparked interest in using video to simulate such physical causes and effects. In this talk, we investigate using physical forces as a conditioning signal for video generation, which enable users and agents to interact with images and videos through both localized point forces, such as poking a plant, and global force fields, such as wind blowing on fabric. We demonstrate that these "force prompts" can enable videos to respond realistically to physical control signals by leveraging the visual and motion priors in the original pretrained model, without using any 3D asset or physics simulator at inference time. The primary challenge of force prompting is the difficulty in obtaining high-quality paired force-video training data, both in real-world data, due to the difficulty of obtaining force signals, and in synthetic data, due to limitations in the visual quality and domain diversity of physics simulators. Our key finding is that video generation models can generalize remarkably well when adapted to follow physical force conditioning using only synthetic videos. Building on this, we explore the role of such force-conditioned world models in robot planning and decision-making.

The talk introduces an agentic AI framework centered on how role-specialized LLM agents can support decision-making, adaptation under non-stationary environment, and system validation and verification (V&V) for trustworthy autonomous system. Our framework provides two complementary directions. First, ATLAS proposes agentic self-evolution, where an evolving agent within the core algorithm EvoDPO is improved over a long-horizon by task-distributed supporter LLM agents responsible for exploration, fine-tuning strategy, and adaptive reference-policy gating mechanism. This project targets progressive agent development under drifting environment and demonstrates domain adaptation for sustained improvement in changing domains, such as non-stationary bandits and scientific machine learning. Second, AIVV conducts agentic verification and validation, where mathematically grounded anomaly detection is strictly validated by role-specialized LLM agents that verify failure modes of the system again natural-language requirements. This framework distinguishes nuisance faults from true faults and produces actionable engineering protocols, such as gain-tuning parameter proposals. Together, these projects define a common paradigm: Agentic AI as a structured orchestration that combines statistical or optimization-based inner loops with role-based LLM oversight to achieve scalable adaptation, reliability, and safety. This perspective moves beyond viewing LLM agents as standalone assistants, instead positioning them as coordinating components for progressive self-improving, and trustworthy system V&V intelligent systems.

We introduce AgenticSciML, a collaborative multi-agent system in which over 10 specialized AI agents collaborate to propose, critique, and refine SciML solutions through structured reasoning and iterative evolution. Across physics-informed learning and operator learning tasks, the framework discovers solution methods that outperform single-agent and human-designed baselines by up to four orders of magnitude in error reduction. The agents produce novel strategies -- including adaptive mixture-of-expert architectures, decomposition-based PINNs, and physics-informed operator learning models -- that do not appear explicitly in the curated knowledge base. These results show that collaborative reasoning among AI agents can yield emergent methodological innovation, suggesting a path toward scalable, transparent, and autonomous discovery in scientific computing.

Progress in computational science depends on complex numerical workflows that must faithfully encode physical laws, yet translating conceptual insight into reliable code remains a major bottleneck. Although large language models can generate isolated code fragments, they lack the structured reasoning required to design, verify, and iteratively refine complete scientific pipelines. Here we introduce ATHENA, an agentic framework explicitly designed to emulate scientific research modeled as a knowledge-driven contextual bandit process. Its core loop separates conceptual policy from numerical realization through expert-derived conceptual scaffolding, enabling principled diagnosis, reformulation, and repair of computational strategies. Across scientific computing and scientific machine learning tasks, ATHENA autonomously discovers exact analytical solutions, constructs stable numerical solvers, diagnoses ill-posed formulations, and orchestrates hybrid symbolic–numeric workflows. Quantitatively, ATHENA exceeds expert-designed workflows, achieving superhuman accuracy and robustness on canonical benchmarks. By reframing computation as an object of agentic reasoning, our framework enables autonomous orchestration of heterogeneous algorithms across scientific domains.

I'll discuss how multi-agent systems are effective tools for scientific discovery. I'll first present a series of ways to build multi-agent language model systems, such as having systems that can debate with each other, verify each others responses, or define a search objective to optimize across agents. I'll illustrate some of the benefits of multi-agent systems and how they allow us to build a more self-improving overall system.