NCP-Bench: Can LLM agents keep a story consistent over 100 turns? A new ICML 2026 benchmark tests long-horizon narrative commitment preservation across 100 movie-based environments. Best model…

📉 Key finding: long-horizon interaction remains very hard for current models. • Best model only 42% of runs stay conflict-free after 20 turns • 40–68% of runs end in factual conflicts • Within 100…