Tag
一个跨 4 个产品团队的实践观察：在 12800 个行程预订会话（每会话 5 轮，共 64000 轮）中，单轮相关性高达 97.6%，但约 7.4%（947 个会话）在第 4 轮前因压缩对话状态丢失用户拒绝记录，重新推荐已被拒绝的行程。作者讨论如何划分会话级评分与行为规范，并询问如何区分真正的偏好反转与丢失的对话状态。
This paper evaluates whether open-weight instruction-tuned language models maintain safe behavior during long adversarial conversations, finding that safe-response rates drop from 85-100% on the first turn to 15-44% by depth 101, demonstrating that strong single-turn safety does not persist across sustained interaction.
This paper introduces MT-InfoSeek, a controlled evaluation suite for assessing LLMs' multi-turn information seeking capabilities, revealing that models underestimate missing information and perform poorly with increased underspecification across domains.
Design Arena announces Agentic Game Development, a new multi-file, multi-turn evaluation with access to asset catalogs, built-in libraries like Howler and Tween.js, and expanded tool calls for sprite generation and asset discovery.