Recent Talks
Agentic Information Access
This talk presents a framework for “agentic information access” in a future with millions of specialized large language models. It explores how user agents can identify which models have the right expertise for a query, how to rank and aggregate their answers, and how to ensure reliability, robustness, and safety in this multi-LLM ecosystem. The talk introduces a large-scale experimental setup built on an 80M-document corpus (ClueWeb22) and 20,000 queries to simulate expertise across 1,000 domains. It discusses core research questions on expert selection, uncertainty estimation, adversarial robustness, response aggregation, and evaluation methodologies, illustrating why new principled frameworks are needed as AI shifts from general-purpose models to vast networks of specialized LLMs.
Broken Telephone
This talk explores why information access becomes fragile in the era of LLMs and how retrieval systems break down under real-world noise—typos, ASR errors, false memories, and conversational ambiguity. Using a series of studies on dense retrieval, speech-based search, adversarial robustness, and “tip-of-my-tongue” queries, the talk shows how small distortions propagate through the retrieval pipeline like a broken telephone. It presents methods such as data augmentation, contrastive learning, uncertainty estimation, domain transfer, and multimodal retrieval to make models more robust. The talk closes with insights from the TREC Tip-of-the-Tongue track, highlighting the challenges of complex, memory-based, and conversational queries.