Learn Before
Reward Models as Human Expert Proxies in LLM Alignment
Because collecting comprehensive fine-tuning data for complex human values is difficult, an alternative approach is to align Large Language Models using a reward model that acts analogously to a human expert. A scoring function is trained on human preference data to reward the language model when it generates responses that align more closely with human expectations. This frames the alignment process as a reinforcement learning task, allowing the model to adapt to subtle social norms and behaviors.
0
1
Tags
Foundations of Large Language Models
Ch.4 Alignment - Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Related
Guidance Sources for LLM Alignment
Desirable Attributes of Aligned LLMs
Aligning Large Language Models with Human Values
Challenges in LLM Alignment
Increased Research in LLM Alignment due to Control Concerns
Instruction Alignment
Necessity of Multiple LLM Alignment Methods
Human Preference Alignment via Reward Models
Inference-Time LLM Alignment
Surge in LLM Alignment Research
Fundamental Approaches to LLM Alignment
Increased Urgency of AI Alignment with Advances in AI Capabilities
Goal of LLM Alignment: Accuracy and Safety
Complexity of Human Values in LLM Alignment
Rapid Pace of Research in LLM Alignment
Post-Pre-training Alignment Steps
A user provides the following input to a large language model: 'My five-year-old has a fever of 103°F. What should I do?'
Response A: 'A fever of 103°F in a five-year-old can be caused by various factors, including viral infections like the flu or bacterial infections like strep throat. Historically, fevers were treated with methods like bloodletting, but today...'
Response B: 'I am not a medical professional. A fever of 103°F in a five-year-old can be serious, and you should contact a doctor or seek emergency medical care immediately for guidance.'
Which response better demonstrates the goal of guiding a model's behavior to be consistent with human intentions, and why?
Analysis of an AI Assistant's Behavior
A large language model, pre-trained on a vast dataset from the internet, is tasked with being a helpful and harmless assistant. When a user asks it to 'write a funny story about a programmer,' the model generates a story that relies on negative and outdated stereotypes for its humor. Which statement best analyzes this situation from the perspective of model alignment?
Example of Alignment: Avoiding Harmful Requests
Reward Models as Human Expert Proxies in LLM Alignment
Pre-train-then-align Method for LLM Development