min-p sampling

**Min-p sampling** is the **probability-threshold decoding method that keeps tokens whose probability exceeds a dynamic minimum relative to the top token** - it adapts candidate set size to local confidence conditions. **What Is Min-p sampling?** - **Definition**: Adaptive filtering strategy based on a minimum probability floor tied to peak likelihood. - **Mechanism**: Tokens below threshold are removed, then remaining probabilities are renormalized for sampling. - **Adaptive Behavior**: High-confidence steps keep small candidate sets, uncertain steps keep broader sets. - **Relation**: Acts as an alternative to fixed top-k or fixed cumulative-mass truncation. **Why Min-p sampling Matters** - **Context Sensitivity**: Candidate filtering automatically adjusts to entropy changes across steps. - **Quality Control**: Suppresses extreme tail tokens that often degrade coherence. - **Diversity Preservation**: Retains multiple options when model uncertainty is genuinely high. - **Operational Simplicity**: Single threshold parameter can replace multiple manual limits. - **Robustness**: Often stabilizes generation across heterogeneous prompt types. **How It Is Used in Practice** - **Threshold Calibration**: Tune min-p values on factuality, coherence, and diversity benchmarks. - **Joint Policies**: Combine with temperature controls for finer stochastic behavior shaping. - **Live Monitoring**: Track candidate-set size distribution to detect over- or under-filtering. Min-p sampling is **an adaptive truncation strategy for stable stochastic decoding** - min-p improves control by aligning candidate breadth with model confidence.