Definition
Plain language
A small training reward for keeping the model's options open and uncertain, rather than locking onto one choice too early.
As stated in the literature
A reinforcement-learning shaping term that rewards higher output entropy to encourage exploration; long treated as a hand-tuned folk-wisdom hack, but shown in rollout-informativeness work to fall out directly as a term in the greedy marginal gain of a submodular informativeness objective.
Why it matters: It keeps a learning system exploring options long enough to find good strategies, rather than prematurely locking onto a mediocre one.
For example, the training nudges a model to keep considering several plausible next moves instead of fixating on the first one it happens to favor.
Heard on the show
“The regularizers — the KL divergence penalty pulling you toward a reference model, the entropy bonus keeping you from collapsing — are computed without reference to reward at all.”Episode 010 — When Reward Climbs But Reasoning Goes Generic: Diagnosing Template Collapse in Agentic RL