PhD Thesis Defense: Hansi Zeng, Generative Information Retrieval for the Real World
Content
Speaker:
Abstract:
Recent advances in large generative models have reshaped information retrieval. Generative Information Retrieval (GIR) reformulates document retrieval as a conditional generation problem: each document is represented by a unique document identifier (DocID), and a generative model is trained to generate relevant DocIDs given a query. By aligning retrieval with the generative objective of language models, GIR offers a conceptually unified and differentiable framework that can benefit from large-scale pre-training, advanced decoding, and joint optimization with downstream generation.
Despite its appeal, existing generative retrieval models exhibit significant practical limitations. They struggle to scale to corpora containing millions of documents and often underperform lexical and dense retrieval baselines on benchmarks such as MSMARCO Moreover, GIR relies on autoregressive decoding with constrained beam search, making inference expensive and introducing a fundamental trade-off between retrieval effectiveness and efficiency. These challenges raise critical questions about deploying GIR in real-world large-scale retrieval systems.
This dissertation makes generative retrieval scalable and practical for systems operating over millions of documents through optimization objectives and decoding strategies that jointly improve effectiveness and efficiency at scale. We first introduce RIPOR, which constructs DocIDs via residual quantization of document embeddings and trains the model with a rank-oriented objective considering the sequential decoding nature of GIR; RIPOR scales stably to the full 8.8-million-passage MSMARCO corpus and reaches effectiveness competitive with strong dense retrieval models. We then propose PAG, a hybrid simultaneous-and-autoregressive decoding strategy that plans ahead using document-level score estimates, improving MRR@10 by 15.6% over RIPOR while using a 10x smaller beam and reducing query latency by 22x.
Building on these advances, we integrate generative retrieval as a trainable component within large language model reasoning systems for agentic search, where a model interleaves reasoning and retrieval to answer complex questions. Rather than treating retrieval as a fixed tool, we present CoSearch, which jointly trains a multi-step reasoning agent and a generative document ranker via reinforcement learning, shifting the ranking objective from human relevance labels toward downstream answer utility through a semantic grouping strategy and a composite reward. Across seven question-answering benchmarks, joint training consistently outperforms strong agentic-search baselines.
Taken together, this dissertation advances generative retrieval from a conceptually appealing but impractical paradigm into a system that is competitive at scale, efficient at inference, and able to serve as a trainable retrieval component in modern LLM-based applications for real-world search and knowledge-intensive AI systems.
Advisor:
Hamed Zamani