Decks from lab seminars: presentations of my own work to the lab, and papers by others that I presented in our NLP reading group.
Long-term memory for multi-session dialogue agents organised like an operating system: short-, mid- and long-term tiers, topic segments with a heat score that drives eviction and promotion, and retrieval across the tiers at every turn; results on LoCoMo and GVD. HTML deck.
A mini-survey of tree-based search for document QA on one Build / Traverse / Pack frame (RAPTOR, TreeRAG, MemWalker, PageIndex, RDR2), then HiKEY in depth: document trees with index cards, two-stage traversal, evidence packing under a reader budget, and its results. HTML deck.
What my two papers solved, what a practical web-scale multimodal retriever still needs (raw HTML, reactive pages, more modalities, grounding, budgets), how recent work approaches each gap, and four research directions.
A lab-internal walkthrough of LILaC: why TextRAG and VisRAG lose information, the layered component graph, LLM-driven query decomposition, edge-level late interaction, and the ablation and runtime results.
Instruction-guided retrieval across modalities: score-level versus feature-level fusion, how the M-BEIR benchmark was assembled, and what the experiments say about instructions and the two fusion designs.
Multimodal RAG with one universal encoder and decoder: the two challenges (handling mixed modalities, training them), in-batch pretraining, two-stage fine-tuning, and the error analysis. Opens with a six-slide recap of ACORN, predicate-agnostic hybrid vector search.
RAG over textual graphs: retrieving a query-relevant subgraph with a soft pruner, feeding the subgraph's text and structure to the LLM, and results on WebQSP and ExplaGraphs.
Token-level RAG built up on one diagram: why per-token retrieval is slow, then chunked retrieval, chunked cross-attention, and frozen BERT embedders, each introduced as the fix to a specific problem. Opens with a seven-slide recap of DiskANN.
LLM-based schema matching for data ingestion when the data itself cannot be shown: object-type detection, reducing the task to attribute mapping, NER and double-RAG filters, the key detector, and the ablations. Ends with my own take.
A benchmark for retrieval over semi-structured knowledge bases: how the three SKBs and the query/answer sets are generated in four steps, the analyses of the queries, and the baseline retrievers.