Yuval Belfer

At AI21, Belfer connects research questions with the engineering of deployed systems. The company identifies him as a senior developer advocate and announced his retrieval and inference-debugging talks alongside engineer Asaf Gardin. He earned an MSc at the Weizmann Institute of Science in 2022; its graduate program lists his thesis on residual-network neural tangent kernels.[1][5]

Investigate confident but incorrect inference

Belfer and Gardin’s joint debugging talk follows two failures involving Jamba’s Mamba state cache in vLLM. One allowed a new request to enter decoding before its prefill had established state. Another used an index type that overflowed under larger workloads. The presentation uses comparisons of token log probabilities against another inference framework to locate the divergence. Belfer frames the cases and their lessons; Gardin explains much of the detailed first investigation.[4]

Study what residual networks learn

With Amnon Geifman, Meirav Galun and Ronen Basri, Belfer analyzed residual networks through their neural tangent kernels. The paper studies how eigenvalues vary with frequency and how the balance between residual and skip connections affects the kernel. Belfer and Geifman are marked equal contributors. This is theoretical work on a specified overparameterized network setting, supported by experiments.[3]

Belfer’s verifiability article proposes choosing an agent architecture around the task’s available checks. When a candidate answer can be tested against explicit conditions, verification can help select among attempts. An open research report presents a different problem because a system cannot easily establish completeness. He distinguishes checking the correctness of one candidate from knowing that every relevant candidate has been found, and argues that more attempts alone can leave a weak selection process unchanged.[2]