Persistent Pre-Training Poisoning of LLMs

Name: Persistent Pre-Training Poisoning of LLMs
Uploaded: 2026-01-27T00:00:00.000Z
Duration: 33 min 10 s
Description: Poisoning just 0.1% of an LLM's pre-training data can plant behaviours that are still there after standard safety and chat training (SFT and DPO). The talk tested four attacks. Making the model output gibberish when i...

Google TechTalks · 33:10 · Watch on YouTube

Poisoning just 0.1% of an LLM's pre-training data can plant behaviours that are still there after standard safety and chat training (SFT and DPO). The talk tested four attacks. Making the model output gibberish when i...

Read the full summary on retube

한국어