Notes on all 13 lectures of Nathan Lambert's RLHF course: one scalar reward is the root of most complaints

le_james94 · x · 2026-09-18

James Le worked through all 13 lectures accompanying Nathan Lambert's The RLHF Book (Lambert was post-training lead at the Allen Institute for AI).

Core thesis: a reward model scores an entire response with a single scalar (hidden state at EOS through a scalar head). Nearly every complaint about RLHF—hedging, unrequested refusals, sycophancy—is downstream of that compression. Almost every post-InstructGPT method works around it: rejection sampling skips RL, DPO folds the reward model into the policy, RLVR replaces humans with a checker, judges and constitutions swap in another model or document.

Author's takeaway: verifiable rewards moved the preference problem rather than solving it. If the reward isn't verifiable and no checker exists, how do you decide what to optimize, and how do you know when you've gone too far?

Also notes Lecture 13 on character training: fine-tuning Llama 3.1 8B toward narrower personalities keeps refusals intact, but the sarcastic one guesses you want to be the next Schwarzenegger—same decision, different person.

Original post →

More from Research

Research channel →