AI Algorithm Engineer, DICT Center, China Mobile (Shaanxi)
I build reliability into multi-step AI systems: detecting, correcting, and calibrating errors in multi-hop retrieval and agent memory.
I'm an AI algorithm engineer working on multi-hop retrieval, cross-modal retrieval, and agent systems that run in production. My research asks what happens when one step of a long reasoning chain goes wrong, and how a system can notice and recover.
Research in one picture
Long reasoning chains fail one step at a time. My work is about catching that step and repairing it, without throwing the whole chain away.
Q Were the directors of Film A and Film B born in the same country?
01Decompose — The question becomes a graph of sub-questions with explicit dependencies.
Decompose hard questions into dependent hops and retrieve evidence for each.
Keep long-lived memory accurate: what to store, revise, and retract.
Detect, localise, and repair the step that went wrong instead of starting over.
01 / Featured work
Projects at the intersection of retrieval, structured reasoning, and memory.
Placeholder: one or two sentences on the problem and the idea. For example: decompose a question into a dependency graph of sub-questions, retrieve per hop, and repair only the hop that fails.
Placeholder: one or two sentences on the problem and the idea. For example: keep a long-lived agent memory consistent when facts are revised, retracted, or contradicted.
02 / Research index
Selected papers. The full list is on the publications page.
Xuhang Shi, Second Author, Third Author
Placeholder abstract. Two or three sentences: the problem, the key idea, and the headline result.
Retrieval
First Author, Xuhang Shi, Third Author
Placeholder abstract. Two or three sentences: the problem, the key idea, and the headline result.
Memory
Xuhang Shi, Second Author
Placeholder abstract. Two or three sentences: the problem, the key idea, and the headline result.
Reliability