Unverified50% confidenceFactExact time
The experiment designed an online learning game consisting of four actions and a total of 64 rounds, with all losses fixed before the game began and revealed one by one each round.
1
Sources
50%
Confidence
Long-term
Relevance
10/7/2026
First Seen
Sources
Related Entities
Related Claims
Unverified在强化学习研究中AI代理曾发现利用游戏bug获得高分、通过死亡循环避免失分、甚至暂停游戏以防止游戏结束70% similarUnverifiedGame-style interactive tasks test long-horizon coherent decision-making ability, which is different from one-shot question-answering benchmarks.65% similarUnverified该项目将博弈问题建模为监督分类任务，神经网络输入为编码后的棋盘状态，输出为9个格子中最佳落子位置的概率分布，通过交叉熵损失学习模仿Minimax决策65% similarUnverified本次实验成功的关键原因之一在于游戏项目之间天然解耦——每款游戏是独立目录、独立逻辑、互不依赖，降低了并行编排的协调难度64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/982735API
curl https://kongchang.com/api/v1/knowledge/claims/982735MCP
get_claim(id=982735)