A DAI 2026 WORKSHOP

Adaptive
Decision Making

Foundations, Algorithms,
and Agentic Systems

Learning through interactionA branching decision path connects observation, learning, and a chosen action, with feedback informing future decisions.feedbackObserveLearnDecide
WhenSunday, November 29, 202609:00–12:00 · HKT (UTC+8)
WhereCity University of Hong KongHong Kong · Room to be announced
Add to calendar

The Chinese University of Hong Kong

Talk details
09:00 — 12:00November 29 · Hong Kong time
–Opening Remarks
–Bo An · Nanyang Technological UniversityAgentic Reinforcement Learning
Abstract

LLM-based autonomous agents are playing an increasingly important role in many fields, including general computer control, software engineering, scientific discovery, and social simulation. Early implementations of autonomous agents largely relied on manually constructed workflows based on prompt engineering. However, such static workflows often struggle to generalize effectively to out-of-domain tasks, lack the ability to maintain high-quality, continuous interaction with environments and users, and are unable to achieve self-improvement during long-term operation. As a result, a growing body of research has begun to introduce reinforcement learning, enabling agents to continuously optimize their policies through interaction and thereby systematically improve their generalization, interaction, and tool-use capabilities. This talk will discuss major challenges encountered in agentic RL and some recent progress.

–Zhongyu Wei · Fudan University | Shanghai Innovation InstituteHuman Behavior Foundation Models for Social Simulation
Abstract

Simulating human social behavior requires capturing individual differences and understanding how they shape collective outcomes. This talk introduces SocioVerse, a general social simulation framework grounded in a pool of 10 million real-world users, integrating four engines for social environments, users, scenarios, and behavior. The focus is on developing human behavior foundation models to power its Behavior Engine, enabling personalized and context-sensitive simulation. The speaker will present two complementary advances: CURP, which combines a user encoder with a shared prototype codebook to capture multidimensional individual characteristics and enable efficient personalization; and Socio-Foundation, which integrates task-specific expertise through hierarchical capability distillation to build a unified model for generalizable individual behavior simulation. Together, these efforts address how individuals are represented and how their behavior can be simulated consistently across diverse social contexts.

–Ying Jin · Wharton School, University of PennsylvaniaAI agents for automated validation and generation of scientific hypotheses
Abstract

AI agents powered by Large Language Models (LLMs) can now automate data analysis and propose scientific hypotheses from data. However, a key challenge is to determine whether the generated hypotheses are worth costly follow-up validation, since the sheer volume of hypotheses LLMs can generate makes manual validation impractical.

We argue the role of statistical evidence for AI-in-the-loop scientific discovery. In this talk, we present recent efforts on developing agentic frameworks that enable automated validation and generation of hypotheses with valid statistical evidence. We design high-level statistical principles for the exploration of AI agent to guarantee error control while allowing flexibility and diversity. First, we present POPPER, an agentic framework for rigorous automated validation of free-form, abstract, and high-level hypotheses. Guided by Karl Popper's principle of falsification, Popper validates a hypothesis using LLM agents that design and execute falsification experiments targeting its measurable implications. A sequential testing framework ensures strict Type-I error control while actively gathering evidence from diverse observations. We demonstrate that Popper delivers robust error control, high power, and scalability in a curated functional genomics dataset. Compared to human scientists, Popper achieved comparable performance in validating complex biological hypotheses while reducing time by 10 folds. Second, we present ongoing work on agentic frameworks to orchestrate hypotheses generation and validation to produce candidate scientific hypotheses with Type-I error control.

–Yikun Ban · Beihang UniversityRethinking Advantage Estimation: From Group-Relative Bias to Heterogeneous Agent Collaborative Reinforcement Learning
Abstract

Advantage estimation is central to policy optimization, shaping how agents learn from their own experience and the experience of others. This talk examines its theoretical foundations and its role in enabling collaborative reinforcement learning across heterogeneous agents. I first present a theoretical analysis showing that group-relative advantage estimation exhibits systematic bias when conditioned on groups with nonzero reward variance. This bias depends on task difficulty, leading to underestimated advantages on difficult tasks and overestimated advantages on easy tasks under binary rewards. I then introduce a collaborative reinforcement learning framework that incorporates capability-aware advantage estimation and addresses policy differences to enable effective experience sharing across heterogeneous agents. Experiments on reasoning tasks demonstrate reciprocal improvements across models with different capabilities. Notably, stronger models benefit from both the successful and unsuccessful experiences of weaker models, highlighting that an agent’s contribution to collaborative learning can extend beyond its individual performance.

–Congyuan Duan · The Chinese University of Hong KongRegret Minimization and Statistical Inference in Online Decision Making with High-dimensional Covariates
Abstract

We investigate regret minimization, statistical inference, and their interplay in high-dimensional online decision-making based on the sparse linear contextual bandit model. We integrate the ε-greedy policy for decision-making with an online hard thresholding algorithm for estimating sparse bandit parameters, and introduce an inference framework based on inverse probability weighted debiasing. We show that, when the algorithm achieves a regret upper bound of O(T1−γ), the asymptotic variance of the debiased estimator scales as O(1/T1−γ), indicating an unavoidable trade-off between regret performance and inference efficiency. However, if an additional covariate diversity condition holds, we demonstrate that a pure-greedy bandit algorithm, i.e., an exploration-free algorithm, combined with an average weighted debiasing method can achieve optimal O(log T) regret and O(√T)-consistent inference simultaneously. We also show that a simple sample mean estimator can provide valid inference for the optimal policy's value. Numerical simulations and experiments on Warfarin dosing data further validate the effectiveness of our methods.

Fang Kong

Southern University of Science and Technology

Zhiheng Zhang

Shanghai University of Finance and Economics

Personal website