Why LLMs loop on long outputs: top-p ends up picking the #1 token every step
ziv_ravid · x · 2026-09-28
Explanation of the mechanism: at each step you can check where the chosen word ranks. Normal writing picks from a handful of good options, but with top-p sampling, by the end of a long generation the model takes its #1 choice every single time — which is exactly what a loop is. Adaptive samplers never reach that state.
Related event: Long-text repetition loops traced to sampler, not model(3 posts)→
More from Research
- Microbots Built on 20-Year-Old 55nm Process Float Free After Silicon Etch — ctjlewis · 2026-09-28
- Colosseum paper finds most off-the-shelf LLM agents prone to collusion — nandofioretto · 2026-09-28
- OpenPhysicsAI: open challenge to make AI predict 13 unsolved physics experiments — sebasmtl · 2026-09-28
- Pre-training a Foundation Model Whose Tokenizer Is PTX, Not Natural Language — rickasaurus · 2026-09-28
- Backtesting an LLM Is Hard: virattt Hides Tickers and Dates to Curb Data Leakage — virattt · 2026-09-28
- Humor benchmark lolbench survives kill test: 87 obscure jokes show the gap isn't memory — AffectionateGas9544 · 2026-09-28