Truth Initiative

most citedEmergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs

22 citations

researchers with a paper here

Owain Evans2 profiles2 · h 24
Anna Sztyber-Betley1 · h 6
Daniel Tan1 · h 3
Jan Betley1 · h 9
Mart'in Soto1 · h 3
Nathan Labenz1 · h 2
Niels Warncke1 · h 3
Xuchan Bao1 · h 3

collaborating institutions

Aisin (United States)US1 paper
Berkeley CollegeUS1 paper
Bexley HallUS1 paper
Risk Management Solutions (United Kingdom)GB1 paper
Stanford UniversityUS1 paper
University College LondonGB1 paper
University of California, BerkeleyUS1 paper
University of TorontoCA1 paper
Warsaw University of TechnologyPL1 paper

Showing cs.CLShow all

1 paper · 1 filter

cs.CL2026★ 22 cited

Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs

Jan Betley, Daniel Tan, Niels Warncke +5

We present a surprising result regarding LLMs and alignment. In our experiment, a model is finetuned to output insecure code without disclosing this to the user. The resulting mode…