Benchmark RadarAI BENCHMARK PROFILE
WSE-bench
WSE-bench evaluates LLM storytelling in open-ended world simulations, assessing three capacities: Generation Coverage (proportion of planned narrative steps), Consistency (canon coherence), and Richness (meaningful development). Uses a process benchmark to compare agent architectures.
- Released
- 2026-08-16
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Storytelling evaluation has focused on finished stories, but open-ended narratives require sustained generation, coherence, and development. WSE-bench makes these dynamics visible and shows they are distinct capacities.
Motivation
Large language models can write fluent stories, but open-ended storytelling requires more than local fluency.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.