NCP-Bench: leaderboard
NCP-Bench (Ma et al., 2026) tests narrative commitment preservation: a language model narrates an interactive story built from a movie synopsis while an adversarial player agent tries to break its established facts and commitments. The score is the average number of turns before the first confirmed fact, commitment or player-input conflict, out of a 100-turn limit, over 100 environments; higher is better. Gemini-2.5-Flash plays the player and audits every narrator response.
Metric: Average turns until termination (out of 100): mean over the environments of the turns played before the first confirmed conflict, the 100-turn limit or full commitment satisfaction; 100 movie-synopsis narrative environments, each with an initial fact ledger, reference trajectory and commitment set; the evaluated model narrates against an adversarial Gemini-2.5-Flash player, and Gemini-2.5-Flash auditors check every narrator response and end the episode at the first confirmed fact, commitment or player-input conflict, at 100 turns, or when every achievement commitment is met; temperature 0.6, top-p 0.95. Source: arxiv.org. Saturation forecast: Around 2035. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 32.92 | #176 |
| 2 | GPT-4o Mini | 24.8 | #1623 |
| 3 | Qwen 3 235B A22B | 16.76 | #623 |
| 4 | DeepSeek V3.2 | 15.88 | #406 |
| 5 | Grok 4.1 Fast | 7.87 | #430 |
| 6 | Kimi K2.5 | 2.88 | #256 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ncp-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.