跳到正文
@kimmonismus· @kimmonismus · X·· 20 天前AI 评分34
AI 导读

Pocket FM 正在研发让 AI 更好保持长篇故事一致性的技术,其研究项目 Sherpa 引入叙事世界模型,追踪角色已知信息、关系变化和未解决的故事线。在团队测试中,它正确回答了 176 道需跨多章节信息的题目中的 89.8%,另一 AI 记忆系统 Graphiti 得分为 57.4%,两者使用相同 AI 模型作答。在另一组 576 道公开题目上,两者得分分别为 62.5% 和 51.6%。

正文

Pocket FM is working on making AI better at keeping long stories consistent. And this is super interesting:

The research behind Sherpa introduces a Narrative World Model that tracks what characters know, how their relationships change, and which storylines are still unresolved.

In the team’s tests, it answered 89.8% of 176 questions requiring information from multiple chapters correctly. Graphiti, another AI memory system, scored 57.4%. Both used the same AI model to answer, with different systems supplying the story information.

On a separate public set of 576 questions, the scores were 62.5% versus 51.6%.

These results show better recall of story details. Whether that translates into better writing still needs to be tested.

来源:@kimmonismus · x.com