Talks and presentations

From Tools to Taste: AI for Research in the Age of Agents

October 24, 2026

Invited Talk, Zhihui Chuanliu Academic Forum (No. 20), College of Computer Science, Sichuan University, Chengdu,

With the development of large language models and agent systems, AI is moving from an auxiliary tool toward an agentic system that can take part in the research process. Organized around three capabilities—scientific taste, tool execution, and evidence verification—this talk presents early explorations by the Fudan NLP Lab: training AI to judge and conceive promising research ideas from research-community feedback; using SciAgentGym to evaluate and improve models’ long-horizon use of scientific tools; and using OpenNovelty to provide traceable, verifiable novelty evidence for peer review. The talk argues for a cautious path: let AI take on more of the exploration and execution, while leaving value judgments and final responsibility to scientists.

Toward Autonomous Agents: The Evolution of Enterprise Services from “Saying” to “Doing”

October 23, 2026

Invited Talk, CNCC 2026, How Agents Reshape Enterprise Services: From Technical Breakthroughs to Practical Deployment, Chengdu,

Large language models are moving from content generation toward task execution, and agents are gradually entering real business workflows. This talk focuses on the key conditions for enterprise service agents to do tasks, do them reliably, and keep doing them over time. Drawing on financial tool use, real-environment evaluation, and reinforcement fine-tuning, it discusses domain data, structured evaluation, and cross-environment generalization. It further analyzes the role of execution assurance layers in tool scheduling, permission control, state management, auditing, and error recovery, and discusses persistent state updates in long-term service across tasks. Enterprise deployment should start from small closed loops with clear boundaries and verifiable outcomes, while keeping human review in high-risk steps.

GUI Agents: From Environment Interaction to Autonomous Planning

October 22, 2026

Invited Talk, CNCC 2026, Youth Power · Intelligent Frontiers: AI Innovation and Xiaomi Technology Collaboration Ecosystem, Chengdu,

Large language models are moving artificial intelligence from language generation toward goal-oriented task execution. Graphical user interfaces are the common interaction entry point for mobile apps, web systems, and desktop software, and they also concentrate the core capabilities agents need in real environments: perceiving state, choosing actions, using feedback, and handling exceptions. This talk introduces the perception, decision-making, action, and feedback loop of GUI agents. Drawing on Magic-GUI, ToolCUA, and MagicAgent, it discusses interactive environments, GUI–tool collaboration, autonomous planning, and reinforcement learning methods, and analyzes the roles of harness engineering, action alignment, and multi-agent collaboration in improving system reliability, safety, and controllability.

Towards General Long-Horizon Agents: Challenges and Innovations

December 26, 2025

Large language models are driving a paradigm revolution in artificial intelligence. From ChatGPT to Artificial General Intelligence (AGI), LLM-based agents have emerged as a highly promising technical direction. This talk first introduces the fundamental concepts of large models and agents, including the definition of AI agents, the core capabilities of LLM-based agents (perception and understanding, reasoning and planning, memory storage and retrieval, generation and action), as well as practical applications such as AI research assistants (WisPaper) and automated task execution. The talk then focuses on three major scaling challenges in building general long-horizon agents: Scaling Environments, Scaling Goals, and Scaling Interactions. Finally, we present our explorations and innovations on AgentGym, a cross-environment self-evolution framework, and AgentGym-RL, a reinforcement learning approach for long-horizon decision making.

Turbocharging LLMs for Scientific Discovery

December 07, 2025

AI is revolutionizing scientific discovery, as highlighted by recent Nobel Prizes in Physics and Chemistry awarded to AI-related breakthroughs. This talk explores how large language models (LLMs) can turbocharge scientific research across the full discovery pipeline — from data analysis and hypothesis generation to experimental design. We begin with an overview of LLMs and their emergent abilities, then examine their growing capabilities in reading comprehension, coding, and complex reasoning. A key focus is on enhancing LLM reasoning for science, covering techniques such as Chain-of-Thought prompting, self-consistency, process supervision, and critique models. We further discuss agent-based modeling and simulation (ABMS) as a powerful paradigm for studying complex systems, and present recent advances in multimodal multi-agent systems for scientific tasks, including PhysicsMinions for physics olympiad problem solving and AtomAgents for AI-driven materials discovery.

Opening the Black Box: Interpretability from Natural Language Processing to Large Language Models

November 09, 2025

Invited Talk, The 23rd China Symposium on Machine Learning and Applications (MLA 2025), Southeast University, Nanjing,

Interpretability is central to opening the “black box” of AI: explaining the structure and behavior of models and improving their trustworthiness and safety. It is key to the compliant use and continued improvement of AI technology. This talk reviews the course of interpretability research for large language models. It first revisits interpretability methods from traditional natural language processing and their applications in dialogue systems and question answering. It then turns to the capability boundaries of large language models and the new interpretability challenges they raise, and presents a framework for mechanistic interpretability, focusing on sparse autoencoders for disentangling superposed features, circuit analysis for uncovering causal relations among features, and sparse replacement models for constructing essential circuits. Finally, it looks ahead to scalable interpretability research as a theoretical and technical foundation for trustworthy large models.

Alignment Techniques for Large Language Models

May 09, 2025

Invited Talk, China Conference on Image and Graphics (CCIG 2025),

While large language models demonstrate remarkable capabilities, they also pose safety and ethical risks that necessitate alignment with human values. This talk provides a systematic overview of recent advances in LLM alignment techniques. We first motivate the need for alignment by examining safety and ethical challenges, along with the core alignment principles of helpfulness, honesty, and harmlessness. We then delve into human preference modeling, covering key issues such as reward model training, generalization, and online updating. Building on this foundation, we present RLHF-based alignment techniques including the PPO-MAX algorithm for stable training, Direct Preference Optimization (DPO), Linear Alignment for inference-time alignment, and multi-path feedback fusion methods. The talk also discusses post-alignment evaluation approaches for both safety/value alignment and capability alignment. Finally, we explore future directions including Self-Play multi-policy adversarial learning and reinforcement learning-centric reasoning models such as O1.

From Large Language Models to World Models

October 22, 2024

From weak AI to Artificial General Intelligence (AGI), large language models (LLMs) are driving landmark breakthroughs in AI. However, LLMs still face notable limitations, such as the inability to predict future events, handle extremely long reasoning chains, or access information beyond their training data. This talk explores the evolutionary path from LLMs to world models. We first discuss how large models construct world knowledge through language, and present our group’s research on large language models (MOSS), speech models (SpeechGPT), and multimodal models (AnyGPT). We then introduce the concept of world models — AI systems capable of simulating and understanding environments to make decisions and predictions — and discuss the integration of embodied intelligence with world models, including cutting-edge directions such as Video-Language-Action models. The talk also covers multimodal alignment techniques (LLaVA, ChatBridge) and multimodal expansion (AnyGPT), envisioning the future roadmap from LLMs toward embodied intelligence and world models.