NCP-Bench - Trajectory Progress: leaderboard
Part of NCP-Bench (Ma et al., 2026): how far an interactive story narrated by the model advances along the reference trajectory of its movie synopsis before the episode ends at the first narrative conflict under an adversarial player. Percentage of the reference trajectory reached, averaged over 100 environments; higher is better.
Metric: Trajectory progress (%; share of the reference trajectory reached at termination, averaged over the environments; 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: Not forecast. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.2 | 15.4 | #406 |
| 2 | Qwen 3 235B A22B | 13.16 | #623 |
| 3 | Grok 4.1 Fast | 12.07 | #430 |
| 4 | GPT-4o Mini | 10.57 | #1623 |
| 5 | Kimi K2.5 | 10.26 | #256 |
| 6 | GPT-5.2 | 9.94 | #176 |
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-trajectory-progress · How It Works · Data refreshed daily, snapshot 2026-11-13.