Retrieval, RAG & Document AI

Probabilistic Information Retrieval & LM-based IR

Probabilistic models of IR: BM25, language models, and relevance models for document retrieval.

Learning-to-Rank

Consistency and calibration of learning-to-rank: pairwise, listwise losses, and surrogate analysis.

Dense vs Sparse Retrieval

Theory of dense and sparse neural retrieval: representation, training, and fusion strategies.

Metric Learning & Approximate Nearest Neighbor

Theory of metric learning losses and ANN data structures for embedding-based retrieval.

RAG Error Decomposition & Performance Bounds

Analyze RAG system errors: retrieval failures, generation hallucinations, and end-to-end performance bounds.

Evaluation Theory in IR/NLP

Rigorous evaluation methodology: inter-annotator agreement, statistical testing, and replicability in IR/NLP.

Tokenization & Subword Models

Information-theoretic analysis of tokenization: BPE, Unigram, and their impact on downstream performance.

Fact Verification & Hallucination Testing

Methods for automated fact checking, hallucination detection, and faithfulness evaluation in LLMs.

Document Structure as Graphs

Model document structure—sections, tables, references—as graphs for enhanced understanding.

Provenance & Verifiable Retrieval

Track and verify the provenance of retrieved information for trustworthy RAG systems.

Cross-Lingual Retrieval & Alignment Theory

Theory of cross-lingual information retrieval: multilingual embeddings, alignment, and zero-shot transfer.

Knowledge Editing & Consistency Constraints

Edit knowledge in language models while maintaining consistency: ROME, MEND, and constraint propagation.