黑料福利网

Imagine an AI agent completing a task that requires hundreds of decisions. If it succeeds, it receives a reward. If it fails, it receives nothing. But neither outcome tells the agent which decisions along the way were useful and which ones caused the problem. This makes it difficult for the agent to learn from its experience.

This challenge is known as the credit assignment problem. It is particularly important for AI systems that learn through interaction, where feedback may be sparse, delayed, or difficult to interpret.

PhD researcher investigated how this delayed feedback can be turned into more useful guidance for AI agents. She developed new methods that identify which decisions, information, and agents contributed to an outcome, allowing AI systems to learn more effectively while also making their learning process easier to understand. She defended her PhD thesis at the Department of Mathematics and Computer Science on Tuesday, September 1, 2026.

From a final score to meaningful feedback

Zhang鈥檚 research first focused on reinforcement learning, a form of machine learning in which an AI agent learns by interacting with its environment and receiving rewards.

In many reinforcement-learning tasks, the reward only becomes available after a sequence of decisions has been completed. Zhang developed methods to redistribute this final reward across the individual steps that contributed to the outcome. Her approach also identifies which parts of the agent鈥檚 observations are relevant to the reward and which are not.

This provides the agent with more informative feedback during learning. At the same time, it offers insight into why the agent succeeded or failed.

Zhang extended this approach to AI agents that learn directly from images. Visual environments contain a huge amount of information, but only some of it is relevant to the task. Zhang developed a method that breaks complex visual observations down into different components, allowing the AI agent to distinguish features that influence its reward from those that are unrelated to the outcome. This helps the agent focus on the information that matters when learning how to make successful decisions.

Giving credit to the right AI agent

The problem becomes more complex when AI agents have to cooperate. If several agents work together towards a shared goal, they may receive a single team reward when the task is completed. But this does not reveal how much each individual agent contributed.

Zhang developed a method that breaks down the team鈥檚 overall result and attributes contributions to individual agents and decision steps. This gives cooperating agents more detailed feedback about their role in the team鈥檚 success or failure.

Such information can help AI agents coordinate more effectively and learn from collective outcomes. It also makes the behaviour of multi-agent systems easier to interpret.

Teaching language-model agents from experience

The research also explores how these ideas can be applied to agents powered by large language models (LLMs). LLM-based agents are increasingly being used to perform tasks that require multiple decisions and interactions rather than simply generating a single response.

Zhang developed a method that extracts understandable rules from previous experiences and uses these rules to guide future decisions. This allows an agent to reuse knowledge it has gained from earlier tasks.

In another approach, she converts the final outcome of a task into feedback for the intermediate decisions that led to it. This allows language-model agents to improve through repeated interaction, without requiring humans to provide detailed feedback for every individual step.