Corti (research lab and infrastructure for healthcare) has developed a new interpretability method, Gradient Interaction Modification (GIM), and it ranked #1 on the global Hugging Face MIB benchmark, outperforming approaches from Meta, Harvard, Cambridge, and DeepMind-affiliated researchers.
Why this matters: Why it matters: This is one of the first methods that can reliably inspect the internal “circuits” behind a model’s decisions at modern scale. Regulators, enterprise safety teams, and researchers have flagged the lack of visibility into model reasoning as a mounting risk. (Gartner recently noted that many enterprises are delaying broad AI deployment until these transparency gaps are resolved.)
What’s new:
GIM reveals causal neural pathways in seconds instead of months
Works on real-world, billion-parameter models
Identifies the interactions behind a model’s behavior — where traditional tools miss key mechanisms
Why this matters: Why it matters: This is one of the first methods that can reliably inspect the internal “circuits” behind a model’s decisions at modern scale. Regulators, enterprise safety teams, and researchers have flagged the lack of visibility into model reasoning as a mounting risk. (Gartner recently noted that many enterprises are delaying broad AI deployment until these transparency gaps are resolved.)
What’s new:
GIM reveals causal neural pathways in seconds instead of months
Works on real-world, billion-parameter models
Identifies the interactions behind a model’s behavior — where traditional tools miss key mechanisms
Corti is open-sourcing the method