Reproduces CoSER’s literary role-play evaluation. Scores varied substantially with the judge, conversation length, and evaluation run. The leaderboard was withdrawn. This report documents the findings, reproduction differences, and limitations.
Evaluates 77 characters from Chinese fiction across 12 criteria using CharacterRM. This report traces the implementation to the authors’ code and data, documents deployment checks, and examines numerical differences from the published results.
Reproduces MiniMax’s evaluation protocol and tests the official judge’s consistency and ranking stability. Documents verified settings, implementation choices, version changes, and test coverage.
Documents the CharacterBench setup, from the prompts in Table 11 to context construction and CharacterJudge scoring. Each setting is traced to the paper or official code.
RESEARCH DIRECTIONS
What we’re exploring
We study model architecture, data, and systems through real tasks.
Social intelligence
We study how identity, emotion, beliefs, and relationships shape dialogue. Our work connects social state and event context with character consistency across long conversations.
We explore generation from text, images, video, and audio, including causal chunking and production-grade 4K. We study character performance, scene structure, and complex motion together.
We trace data, prompts, judges, and statistical settings to their sources. Reproduction studies document implementation choices, deviations, and limitations so others can review the results.
Social intelligence
We study how identity, emotion, beliefs, and relationships shape dialogue. Our work connects social state and event context with character consistency across long conversations.
Explore OmniSoul