Learning from Interaction for Long-Horizon AI Agents

PhD Qualifying Examination


Title: "Learning from Interaction for Long-Horizon AI Agents"

by

Mr. Xichen ZHANG


Abstract:

Despite advances in reasoning, tool use, and planning, agents built on large 
language and multimodal models still struggle to reliably complete complex 
tasks through sustained interaction with an environment. In long-horizon 
tasks, later decisions depend on earlier actions and observations. Delayed 
feedback, changing conditions, and limited context can therefore cause errors 
to accumulate, relevant evidence to be lost, and plans to become outdated. 
The central challenge is to translate competence at individual steps into 
reliable completion of the overall task.

This survey examines how agents learn from interaction and sustain reliable 
execution. It organizes the literature into four directions: training 
environments, policy learning, context and persistent memory, and execution 
control. Within this framework, it compares major mechanisms, their 
assumptions and limitations, and the evidence for their effectiveness across 
web research, software tasks, and continuous multimodal assistance.

Across these directions, a recurring challenge is to align training 
experience with the information needs and recovery decisions that arise 
during execution. Differences in supervision, information access, and 
computational budgets also affect how reported improvements should be 
interpreted. These observations motivate research on using execution failures 
to guide learning, coordinating memory with control, and generalizing to 
unfamiliar tasks and longer horizons.


Date:                   Thursday, 29 October 2026

Time:                   2:00pm - 4:00pm

Venue:                  Room 3494
                        Lifts 25/26

Committee Members:      Prof. Jiaya Jia (Supervisor)
                        Dr. Qifeng Chen (Chairperson)
                        Dr. Long Chen