← Publications / Safety Post-Training & Alignment

Targeted Multilingual and Cybersecurity Safety Repair for Gemma 4 31B

Lexsi Labs Team

Technical Report 20262026Technical report
Contrast, attribute, localise and verify pipeline for circuit-restricted weight edits on Gemma 4 31B

Targeted weight interventions for multilingual and cybersecurity safety failures, reporting improved safety and multi-turn jailbreak resistance while largely preserving broader capabilities. Part of the Lexsi Alignment and Safety Stack (CuratorKIT, AlignTune, SafeTune, CircuitKIT, AuditKIT, MANSU).

Multilingual checkpoint · Cybersecurity checkpoint · Technical report

Cite

@misc{team2026targeted,
  title         = {{Targeted Multilingual and Cybersecurity Safety Repair for Gemma 4 31B}},
  author        = {Lexsi Labs Team},
  note          = {Technical Report 2026},
  year          = {2026},
  url           = {https://openreview.net/forum?id=Mgnk5R9Z0l}
}
All →