IC-509GPT-4 used as a preference labeler via prompt engineering yields alignment performance comparable to a task-specific 125M model

Leitian Tao, Yixuan Li

SourceYour Weak LLM is Secretly a Strong Teacher for Alignment

The paper compares four supervisor models for providing preference labels on unlabeled data: OPT-125M (weak), OPT-1.3B (moderate), LLaMA-3-8B (strong), and GPT-4 (very strong, via prompt engineering following Lee et al. 2023). The resulting labels are used to align an OPT-1.3B student via DPO. Gold rewards are 4.63, 4.84, 4.78, and 4.91 respectively. The performance under weak, moderate, and strong supervisors is nearly comparable, and the paper notes that in some instances the 125M model outperforms GPT-4, indicating that a task-specific weak LLM can serve as a more effective supervisor than a larger model relying solely on prompt engineering.

Evidence
correlational
Key metric
gold reward 4.63 (OPT-125M), 4.84 (OPT-1.3B), 4.78 (LLaMA-3-8B), 4.91 (GPT-4) on OPT-1.3B student
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report
Datasets
HH-RLHF / HH-RLHF-RedTeam [eval]
Methods
Direct Preference Optimization / Direct Preference Optimisation / DPO / Rafailov et al. 2023 (Direct Preference Optimization) [primary]
Related work
RLAIF [builds-on]
Related findings
IC-508
Extraction
automatic-extraction