Latent Space·Dec 31, 2025

Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?

Name: Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
Uploaded: 2025-12-31T01:11:46+00:00
Description: Reinforcement Learning with Verifiable Rewards (RLVR) improves sampling efficiency by sharpening probabilities of existing paths within a base model, but does not expand the model's reasoning capabili