Schedule

Keynote 1: Architecting Controllable Parametric Memory in Language Models
(Speaker: Aditi Raghunathan, CMU)

• Abstract: Parametric memory in language models is difficult to control because the same weights that enable generalization also absorb source-specific memorization. We argue this difficulty is structural: standard gradient descent entangles memorization with general capabilities, making targeted interventions fundamentally hard. We introduce Memorization Sinks (MemSinks), which routes repeated sequences into dedicated sparse neuron sets, isolating memorization by design. Building on this, Natively Unlearnable LLMs (NULLs) assigns each data source a deterministic sparse mask over a shared neuron pool, reducing knowledge control to a deployment-time masking operation with no gradient updates required. Through combinatorial masking, a single pool of sink neurons supports millions of independently addressable sources. Experiments on Wikipedia and Harry Potter show selective suppression tracking the gold standard of retraining from scratch, with resistance to adversarial relearning attacks. We argue for a broader shift: rather than intervening on monolithic parametric memory post-hoc, we can structure memory during training so that individual sources remain addressable throughout deployment.

Keynote 2: Scaling Text Optimization for Reasoning and Agentic LLMs
(Speaker: Chelsea Finn, Stanford)

Keynote 3: Toward the Next Generation of Personalized, Proactive, and Self-Evolving Agents
(Speaker: Weiwen Liu, SJTU)

• Abstract: AI agents are rapidly moving toward practical deployment, demonstrating enormous potential to reshape personal digital life. A key enabler of this transition is agent memory, namely the ability to accumulate, organize, and continually update knowledge about users, tasks, and past interactions over time. Such memory provides the foundation for long-term personalization, contextual continuity, and adaptive decision-making, making it essential for agents that aim to serve as persistent companions rather than one-shot assistants. However, traditional passive response paradigms are insufficient to support users’ sustained and in-depth needs, while the cost and burden of high-frequency interaction continue to hinder their integration into everyday life. Enabling agents to provide long-term companionship, proactive service, and personalized adaptation has therefore become a central challenge in artificial intelligence. This talk will present a new perspective on agent development by shifting the focus from performance-centered design to value-centered design. From this perspective, the talk will systematically examine three key pathways toward all-day proactive understanding and long-term continual evolution, including sleep-time scaling, multi-step capabilities, and proactive interaction. Building on this foundation, the talk will further explore the evolution toward an open agentic web. The next generation of general-purpose agents will be capable of cross-organization and cross-platform collaboration and will be equipped with persistent identity and continual evolution mechanisms. Finally, the talk will summarize the core trajectory of general-purpose agent evolution and outline a foundational framework and future vision for its development.

Keynote 4: Data-Centric Memory: Improving LLM Agents Through Better Data
(Speaker: Fred Sala, Snorkel.AI)

• Abstract: LLM/agent memory is often framed as a systems problem, addressing questions such as how to build better storage layers, design retrieval mechanisms, and improve context pipelines. In this talk, I will argue for a complementary view: memory as a fundamental data problem. Many memory failures result from poor choices about what trajectories are recorded, how they are represented, when they are retrieved, and how their usefulness is evaluated. We will discuss data-centric perspectives on agent memory, where memory is treated as a continuously constructed dataset. This framing highlights familiar issues, such as selection bias, noise, redundancy, compression, etc. It also suggests new directions for memory research. In particular, it points toward memory systems that are optimized not just for scale or latency, but for data quality and downstream utility. In other words, better agent memory will require not only better infrastructure, but better data. 

Keynote 5: Does your LLM Agent Have a Self?
(Speaker: Mengye Ren, NYU)

Keynote 7: The AI Scientific Revolution will be Driven by Open-Ended Algorithms
(Speaker: Jeff Clune, UBC)

Keynote 8: From Context to Continuity: Memory for Self-Managed Agents
(Speaker: Jeff Z. Pan, University of Edinburgh)

Oral Presentation Session 1 (11:50-12:20)

• Episodic Memory from Compression Boundaries in Latent Representation Space (David Ferreira ⋅ Priscila Ribeiro ⋅ EMANUEL PASSINATO ⋅ Diogo Costa Silva ⋅ Arlindo Galvão Filho)

• SimpleMem: Efficient Lifelong Memory for LLM Agents (Jiaqi Liu ⋅ Yaofeng Su ⋅ Peng Xia ⋅ Siwei Han ⋅ Zeyu Zheng ⋅ Cihang Xie ⋅ Mingyu Ding ⋅ Huaxiu Yao)

• SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning (Peng Xia ⋅ Jianwen Chen ⋅ Hanyang Wang ⋅ Jiaqi Liu ⋅ Kaide Zeng ⋅ Yu Wang ⋅ Siwei Han ⋅ Yiyang Zhou ⋅ Xujiang Zhao ⋅ Haifeng Chen ⋅ Zeyu Zheng ⋅ Cihang Xie ⋅ Huaxiu Yao)

• Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models (Qizheng Zhang ⋅ Changran Hu ⋅ Shubhangi Upasani ⋅ Boyuan Ma ⋅ Fenglu Hong ⋅ Vamsidhar Kamanuru ⋅ Jay Rainton ⋅ Chen Wu ⋅ Mengmeng Ji ⋅ Hanchen Li ⋅ Urmish Thakker ⋅ James Y Zou ⋅ Kunle Olukotun)

• Toward a Theory of Hierarchical Memory for Language Agents (Yashar Talebirad ⋅ Ali Parsaee ⋅ Csongor Szepesvari ⋅ Amirhossein Nadiri ⋅ Osmar Zaiane)

Oral Presentation Session 2 (13:30-14:00)

• Chow–Liu Ordering for Long-Context Reasoning in Chain-of-Agents (Naman Gupta ⋅ Vaibhav Singh ⋅ Arun Iyer ⋅ Kirankumar Shiragur ⋅ Pratham Grover ⋅ Ramakrishna Bairi ⋅ Ritabrata Maiti ⋅ Sankarshan Damle ⋅ Shachee Gupta ⋅ Rishikesh Maurya ⋅ Vageesh C)

• Evaluating Memory Structure in LLM Agents (Alina Shutova ⋅ Alexandra Olenina ⋅ Vinogradov Ivan ⋅ Anton Sinitsin)

• Learning to Continually Learn via Meta-learning Agentic Memory Designs (Yiming Xiong ⋅ Shengran Hu ⋅ Jeff Clune)

• Spectral Attention Steering for Prompt Highlighting (Waylon Li ⋅ Yuchen Niu ⋅ Yongxin Yang ⋅ Keshuang Li ⋅ Tiejun Ma ⋅ Shay B Cohen)

• AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (Yujie Zhao ⋅ Boqin Yuan ⋅ Junbo Huang ⋅ Haocheng Yuan ⋅ Zhongming Yu ⋅ Lanxiang Hu ⋅ Haozhou Xu ⋅ Abhilash Shankarampeta ⋅ Zimeng Huang ⋅ Wentao Ni ⋅ Yuandong Tian ⋅ Jishen Zhao)

Oral Presentation Session 3 (16:15-16:45)

• INFMEM: Learning System-2 Memory Control for Long-Context Agent (Xinyu Wang ⋅ Mingze Li ⋅ Peng Lu ⋅ Xiao-Wen Chang ⋅ Lifeng Shang ⋅ Jinpeng Li ⋅ Fei Mi ⋅ Prasanna Parthasarathi ⋅ Yufei CUI)

• Just-In-Time Reinforcement Learning: Continual Learning in LLM Agents Without Gradient Updates (Yibo Li ⋅ Zijie Lin ⋅ Ailin Deng ⋅ Xuan Zhang ⋅ Yufei He ⋅ Shuo Ji ⋅ Tri Cao ⋅ Bryan Hooi)

• Log-Augmented Generation: Scaling Test-Time Reasoning with Reusable Computation (Peter Baile Chen ⋅ Yi Zhang ⋅ Dan Roth ⋅ Samuel Madden ⋅ Jacob Andreas ⋅ Mike Cafarella)

• Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents? (Thibaud Gloaguen ⋅ Niels Mündler ⋅ Mark Mueller ⋅ Veselin Raychev ⋅ Martin Vechev)

• GAM: Hierarchical Graph Memory for LLM-based Agents (Zhaofen Wu ⋅ Hanrong Zhang ⋅ Fulin Lin ⋅ Wujiang Xu ⋅ Xinran Xu ⋅ Yankai Chen ⋅ Henry Peng Zou ⋅ shaowen chen ⋅ Weizhi Zhang ⋅ Xue Liu ⋅ Philip Yu ⋅ Hongwei Wang)