Event ·Multi-source

RetroChessBench 基准发布:GPT 6 Astra 在国际象棋回溯推理中大面积失败

研究者推出 RetroChessBench,一个包含 12 项任务的回溯国际象棋基准,要求模型重建合法棋史并区分必然与可能事件。在每任务两次试验中,GPT 6 Astra High 无工具时 24 次尝试中仅 2 次给出正确最终答案,使用代码执行后也只改善到 6/24。终端 RL 奖励均值分别为 14.75% 和 27.19%。失败案例显示模型会拒绝合法棋史、把临时阻碍当作永久阻碍,研究者推测是网络训练数据缺乏此类详细推导文本所致。

Current read

GPT 6 Astra 做数学题很厉害,但给它回溯国际象棋题,24 次只对 2 次,这个新基准专门测模型的纯推理能力,失败原因分析很有意思。

11 sources34° AI heatLast updated 2026/10/29 04:32:11

Evidence

11 sources
Primary sourcearXiv cs.LG
View source
Related source 1rohanpaul_ai
View source
Related source 2kimmonismus
View source
Related source 3TestingCatalog
View source
Related source 4Yuchen Jin
View source
Related source 6Artificial Analysis
View source
Related source 8量子位
View source
Related source 9Ethan Mollick
View source
Related source 10IT之家
View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.