XTC Sampling Paper Accepted to NeurIPS 2026, Targets Creative Blandness
ziv_ravid · x · 2026-09-26
The team's paper on XTC (Exclude Top Choices) sampling has been accepted to NeurIPS 2026.
Key observation: most samplers either cut the tail (top-p, min-p) or flatten everything (temperature), but much of creative blandness lives in the head of the distribution — the model sees several good continuations and still picks the most generic one. XTC is designed to fix exactly that.
Related event: XTC sampling paper accepted to NeurIPS, tackling creative mediocrity(2 posts)→
More from Research
- GPT-6 Luna uses fewer reasoning tokens than 5.6 on ARC-AGI-2, hard tasks stymie both — mhmazur · 2026-09-26
- LLM Clustering Could Separate ICLR Slop Papers Into Their Own Cluster, Devs Argue — joelbot3000 · 2026-09-26
- Contrastive World Models: latent-space world models without pixel prediction — bonniesjli · 2026-09-26
- Researchers including Google build first complete brain map of a male fruit fly, 166,000+ neurons — burny_tech · 2026-09-26
- Joel Leibo: future LLM clustering will reveal a separable cluster of zero-human AI research papers — joelbot3000 · 2026-09-26
- LLM clustering of OpenReview papers will expose a separable AI-slop cluster, researcher predicts — joelbot3000 · 2026-09-26