Challenges in Making LLMs Safe and Robust
Talkat Special Year on Large Language Models and Transformers, Part 1 Boot Camp · 2024-09-04..09-06·
summary
Shows that transformers trained from scratch can implement both ridge vs discrete regression but fine-tuning on one task quickly ruins performance on the other; simple prompt transforms (scaling labels, switching language) recover the lost ability. For vision CLIP features, learn linear classifier first then fine-tune preserves pre-trained structure and gives better OOD robustness than naive end-to-end tuning; weight-averaging pre-trained and fine-tuned weights also helps. Against attackers, nonsensical adversarial suffix/jailbreak strings transfer from open models to closed systems and erasing a token (ban a word) proves stubborn.
AI-written summary · about
chapters (25)
- 0:00Introduction
- 1:22LLM Safety Challenges
- 3:34Distribution Shifts
- 4:52Catastrophic Forgetting
- 7:46Jailbreaking Mechanisms
- 8:38Implicit Inference Model
- 11:13Forgetting Simulation
- 14:38Solving Forgetting
- 17:17Scaling and Prompting
- 20:27Reverting Behaviors
- 22:51Jailbreak Examples
- 27:35Extracting Data
- 30:42Generalization Issues
- 33:39Robust Fine-tuning
- 38:42Linear Probing
- 45:36LP-FT Method
- 50:23Weight Averaging
- 52:51Finetuning Loss
- 55:21Distortion and Robustness
- 56:58Factual Knowledge
- 1:01:59Knowledge Extraction
- 1:07:03Adversarial Challenges
- 1:11:12LLM Attack Vectors
- 1:13:49Adversarial Examples
- 1:16:46Conclusion
related talks
01h 00mUnderstanding the Robustness of Deep Learning
25mFinetuning Large Language Models (LLMs) for Security Log Detections
46mLearning Theory of Transformers: Generalization and Optimization of In-Context Learning
29mLocal LLM fine-tuning: a practical example
31mFine-tuning LLMs: A Cost-Benefit Analysis for Businesses
42mExploring Pillars of Trustworthy AI: Robustness and Fairness
25mWhy LLMs Still Fail: Reasoning, Robustness, and Safety
25mTricking Neural Networks : Explore Adversarial Attacks