Conceptpolicy entropy

Policy Entropy

A statistic in reinforcement learning that measures the uncertainty or randomness of policy outputs. For LLMs it reflects the dispersion of generated responses and is often used as a diagnostic indicator or control signal for the degree of exploration.

Timeline (last 90 days)

Nov 5

Prior research attempted to use policy entropy as both a diagnostic tool and a control signal, but experiments found that merely increasing entropy cannot reliably change the eventual plateau.

Unverified50%

All Facts (1)

Unverified

Prior research attempted to use policy entropy as both a diagnostic tool and a control signal, but experiments found that merely increasing entropy cannot reliably change the eventual plateau.

50%

Source Articles

Reward Budgeting Mechanism Solves Premature Convergence in LLM Reasoning RLNov 5