Latent Reward Steering: An Adaptive Inference-Time Framework that Implicitly Promotes Cognitive Behaviors in Reasoning LLMs
The episode discusses a paper titled "Latent Reward Steering," which proposes an adaptive inference-time framework to implicitly promote cognitive behaviors in reasoning LLMs. Hosts discuss how this method learns a latent reward model on reasoning traces to guide the AI toward correct paths by applying targeted corrections at fragile states, leading to consistent performance gains across different model sizes.