XTC sampling paper accepted to NeurIPS, already available in llama.cpp and other local engines

The XTC (Exclude Top Choices) sampling method proposed by @zivravid's team has been accepted by NeurIPS 2026, and has already been built into open-source inference engines such as llama.cpp, ExLlamaV2, and text-generation-webui, so users running local models can use it right now.

Confirmed

• The paper was accepted at NeurIPS 2026, providing XTC with a formal definition and systematic evaluation for the first time, filling a previous gap
• Mechanism: when two or more tokens have probabilities above a threshold τ, it removes all candidates except the least likely one with probability ρ and renormalizes; when the model is already confident, it doesn't intervene
• Key insight: mainstream samplers either truncate the long tail (top-p, min-p) or flatten the distribution overall (temperature), while mediocre creative output largely stems from high-probability head tokens being repeatedly selected—exactly what XTC targets
• Cross-model-family experiments show XTC makes creative output more diverse and less repetitive, with human raters preferring it on creativity dimensions at no cost to fluency; at equal diversity levels it hurts instruction-following far less than raising temperature, and can be combined with temperature and min-p
• Acknowledged contributors: XTC originators Philipp Weidmann, Judah Goldfeder, and Sanjay Basu, plus @Hellisotherpe10 who actually ran the experiments

Why it matters

• XTC is the first paper to systematically formalize the "exclude top choices" creative sampling idea, offering a new tool for tackling bland, repetitive model output
• It's already in mainstream open-source inference engines, so ordinary local users can try it immediately—remarkably fast practical adoption
• Compared to simply raising temperature, it strikes a better balance between diversity and instruction-following

2026-09-26 ~ 2026-09-26 · 5 related posts

• [source] XTC Sampling Paper Accepted to NeurIPS 2026, Targets Creative Blandness — ziv_ravid · 2026-09-26
• [source] How XTC Works: Probabilistically Excluding Top Token Choices to Boost Diversity — ziv_ravid · 2026-09-26
• XTC Sampling Makes Creative Outputs More Diverse Without Hurting Fluency — ziv_ravid · 2026-09-26
• [source] XTC Sampling Is Already Built Into llama.cpp and Other Open-Source Inference Engines — ziv_ravid · 2026-09-26

1 near-duplicate retellings: ziv_ravid