Niv Granot is an Algorithms Group Lead at AI21 Labs, developing the retrieval, search, and verification systems behind dependable AI agents. His research demonstrates how better information architecture, coordinated agents, and independent verification can outperform simply upgrading to a larger model.
From 2018 to 2021, Granot conducted graduate research at the Weizmann Institute of Science under Michal Irani. He coauthored Drop the GAN, which explored nearest-neighbor image patches as an alternative to adversarial training for single-image generation, and research on generating diverse videos from a single example.
At AI21, he shifted toward knowledge-intensive language systems and led work on retrieval and web-search tools. In 2025, he and colleague Yuval Belfer challenged conventional RAG evaluation, arguing that impressive benchmarks often reward questions answerable from individual passages while overlooking realistic questions requiring evidence from many documents.
- Structured RAG for aggregative questions. Granot coauthored research on structured retrieval that groups related documents, extracts shared schemas during ingestion, and translates counting, ranking, and comparison questions into SQL. Entity normalization, ambiguous queries, and complicated schemas remain practical constraints.
- Query-dependent chunking. His multi-scale retrieval research treats document granularity as a property of the question: precise lookups and context-heavy questions require different representations. Indexing multiple chunk sizes and combining results with reciprocal rank fusion can improve retrieval without retraining embedding models, although it increases indexing and storage costs.
- Research-agent composition. Granot and AI21 collaborators combined outputs from weaker research agents into a system that reached first place on the DeepResearch Bench II leaderboard in June 2026, demonstrating the value of orchestration and synthesis.
- Independent answer verification. His work on trained verifiers addresses a persistent agent failure: generating a correct candidate but selecting a more popular incorrect answer. Independently investigating candidate responses before aggregation can improve accuracy and cost efficiency without relying exclusively on frontier models.
