EleutherAI tests whether EK-FAC data attribution can block subliminal learning — results are 'solidly mid'

BlancheMinerva · x · 2026-09-29

A new EleutherAI preprint examines whether data attribution techniques like EK-FAC can reliably filter training data to prevent subliminal learning — the transfer of behaviors through data where the link isn't obvious. The verdict: results are "solidly mid," meaning attribution-driven filtering is not yet dependable as a source-level safety intervention. The authors tease an upcoming paper with much better results on emergent misalignment. The work connects data attribution tooling directly to one of the thornier alignment threats.

Original post →

More from Safety

Safety channel →