Transfer learning for weak-to-strong generalization
summary
Claim: weak-to-strong generalization is possible by eliciting latent knowledge from pre‑trained LLMs. Method: apply refinement via in‑context learning—use weak teacher labels to build local prompt examples, query a strong pre‑trained model (GPT‑3.5‑Turbo / GPT‑4o MINI) to produce refined labels, then fine‑tune on those. Theory frames this as a latent concept shift deconvolution problem and proves naive fine‑tuning can fail; experiments on GSM8K and harder math sets show consistent gains but notable failures (e.g., LLaMA‑7B) and an irreducible bias remain.
AI-written summary · about
chapters (7)
- 0:00Introduction to superalignment
- 2:58Latent concept shift framework
- 10:15Limitations of naive fine-tuning
- 12:16Refinement as a solution
- 19:47Theoretical identification
- 20:25Experiments and evaluation
- 37:57Q&A and concluding remarks
related talks
43mWeak-to-Strong Generalization
01h 27mChallenges in Making LLMs Safe and Robust
58mIn-Context Principle Learning from Mistakes
08mLeveraging large-scale models for efficient learning
22mSpecializing Small Language Models with less data
01h 04mRobust Generalization in the Era of LLMs: Jailbreaking Attacks and Defenses
29mLocal LLM fine-tuning: a practical example
01h 30mToward Understanding In-context Learning