Reward Shaping to Mitigate Reward Hacking in RLHF

Name: Reward Shaping to Mitigate Reward Hacking in RLHF
Uploaded: 2026-09-06T02:01:39.686132+00:00
Duration: 403 s
Description: The episode discusses a paper titled "Reward Shaping to Mitigate Reward Hacking in RLHF," which introduces Preference as Reward (PAR). The hosts analyze how PAR addresses reward hacking by transforming raw scores into bounded, preference-based signals. They conclude that this method improves reliability, consistency, and performance across various AI models.

The episode discusses a paper titled "Reward Shaping to Mitigate Reward Hacking in RLHF," which introduces Preference as Reward (PAR). The hosts analyze how PAR addresses reward hacking by transforming raw scores into bounded, preference-based signals. They conclude that this method improves reliability, consistency, and performance across various AI models.

Open the full episode