Concept · 3 episode(s)

Policy Gradient

Definition

Policy gradient methods train a policy directly by following the gradient of expected reward with respect to its parameters, rather than learning a value function and acting greedy. REINFORCE, PPO, and GRPO are all variants tuned for different variance/bias trade-offs.

Episodes covering this

060
When Splitting One Model Across Three Agents Doubles Its Accuracy
Lu, Fang, Zhong et al. · University of Georgia·26 min·May 20, 2026
028
Teaching a Model to Hire Copies of Itself: Recursive Agent Optimization
Gandhi, Chakraborty, Wang et al. · Carnegie Mellon University·23 min·May 08, 2026
025
The Missing Gradient Term That Predicts Sycophancy in RLHF
Gauthier, Bach, Jordan · Inria·22 min·May 07, 2026

Worth reading next

Papers we haven't done a deep dive on yet, but would recommend on this topic.

Scaling LLM Test-Time Compute Optimally Can Be More Effective than Scaling Model Parameters