PERSON DIRECTORY
Iman Mirzadeh
Iman Mirzadeh appears in 1 indexed conversation across Machine Learning Street Talk. This directory brings every appearance, source, TL;DR, digest, and transcript into one searchable feed.
1 EPISODE1 SHOW
Language
Moving Beyond Surface Statistics (Apple researcher) [Iman Mirzadeh]
GSM-Symbolic exposes 14%-20% accuracy gaps and large variance across template variants, while GSM-NoOp’s irrelevant clause causes a substantial drop.These failures suggest distribution matching and prompt conditioning can mimic reasoning without stable concepts, even when tool use improves task completion.Novel-task learning speed, agency, and transfer—not saturated benchmark scores—remain the unresolved tests of durable autonomy.