Cascading Adversarial Bias from Injection to Distillation in Language Models

Name: Cascading Adversarial Bias from Injection to Distillation in Language Models
Uploaded: 2026-01-27T00:00:00.000Z
Duration: 38 min 13 s
Description: Harsh's talk, based on work he did at Google, shows that an attacker who poisons only 0.25–0.5% of a teacher model's instruction-tuning data can plant a subtle bias that a distilled student model inherits and often am...

Google TechTalks · 38:13

Harsh's talk, based on work he did at Google, shows that an attacker who poisons only 0.25–0.5% of a teacher model's instruction-tuning data can plant a subtle bias that a distilled student model inherits and often am...

Read the full summary on tuber