presence penalty

**Presence penalty** is the **decoding control that reduces the likelihood of tokens that have already appeared at least once in the generated output** - it encourages topic expansion and helps avoid immediate repetition loops. **What Is Presence penalty?** - **Definition**: A penalty applied to previously seen tokens regardless of how many times they appeared. - **Mechanism**: Token logits are adjusted before sampling or search so repeated reuse becomes less likely. - **Difference**: Unlike frequency penalty, presence penalty is usually binary per token occurrence. - **Use Scope**: Commonly used in chat and creative generation where novelty is desired. **Why Presence penalty Matters** - **Diversity Lift**: Encourages the model to introduce new words and ideas over time. - **Loop Prevention**: Reduces repeated short phrases that hurt readability. - **Conversation Quality**: Helps multi-turn assistants avoid echoing user wording too closely. - **Style Control**: Provides a simple lever for balancing novelty against precision. - **Operational Safety**: Can lower risk of degeneration in long outputs. **How It Is Used in Practice** - **Penalty Tuning**: Set conservative defaults and increase only when repetition appears. - **Task Profiling**: Use lower settings for factual QA and higher settings for ideation tasks. - **Metric Tracking**: Monitor repetition rate, coherence, and user preference after changes. Presence penalty is **a practical anti-repetition control in decoding stacks** - when calibrated carefully, it improves variation without breaking coherence.