Yuval Belfer

Yuval Belfer is a senior developer advocate at AI21 Labs, a machine-learning researcher and educator, and the host of Yet Another AI Podcast. He specializes in making generative-AI systems dependable under real-world constraints: retrieving relevant information, following complex instructions, checking answers, and balancing accuracy against cost and latency.

Before joining AI21, Belfer designed chips at Amazon. He earned an MSc in computer science from the Weizmann Institute of Science and coauthored a 2024 Journal of Machine Learning Research paper on neural tangent kernels in deep residual networks. His work at AI21 progressed from a 2022 tutorial on generating professional profiles from résumés to the 2023 introduction of Contextual Answers, a document-grounded question-answering product. He also teaches generative-AI and language-model development at Reichman University, Google & Reichman Tech School, and Nebius Academy, and interviews AI practitioners through Yet Another AI Podcast.

• RAG evaluation must test realistic questions. Belfer and Niv Granot argue that benchmarks built around answers contained in individual passages conceal failures on counting, comparison, and aggregation across documents. Their critique of conventional retrieval evaluation describes structured RAG, which extracts document schemas and answers suitable aggregate queries with SQL. Their shared conference presentation also identifies practical limitations, including inconsistent entity names and corpora that resist relational structure.

• Retrieval should adapt its chunk size to the question. Belfer and Granot’s multi-scale retrieval research indexes documents at several passage lengths and combines results through reciprocal rank fusion. Their experiments reported retrieval gains of 1% to 37% across benchmarks without retraining the underlying model.

• Complex agents need dynamic planning and replanning. Belfer distinguishes ordinary routing, tool calling, and ReAct loops from systems that evaluate candidate actions, revise plans, validate intermediate outputs, and execute independent steps concurrently. His account of planning-centered agent architectures uses AI21 Maestro to illustrate decisions informed by expected cost, latency, and success probability. His practical recommendation is to start with the simplest architecture that meets the task.

• Independent answer verification improves agent reliability. In research with Guy Freund, Eran Goldstein, and Niv Granot, Belfer investigates agents that generate correct candidate answers but select incorrect results. An independent verifier can investigate competing responses and reject unsupported claims. His conversation with AI21 systems lead Eran Gat adds another operational concern: standardized tool access does not provide the workspace isolation for concurrent agents needed when multiple processes modify shared files or environments.