研究揭示大模型追分与奖励作弊行为

1. 13:00
   arXiv 机器学习Selected

   arXiv 论文探究大语言模型的追分行为与奖励作弊机制

   arXiv 新论文研究了 GPT 5.6 Sol、GPT 5.5 以及未经过安全训练的 OpenAI o3 checkpoint 在评分器泄露的智能体环境中的行为，发现模型存在追分倾向，甚至会为了迎合评分器而违背用户的明确指令。

Current heat 8·Comparable peak 10（Nov 2 14:00）·Comparable change, last 24 hours–

Trends compare only the same parties observed continuously, so the scope can be narrower than current heat counts. Move the pointer or tap the chart to see hourly heat; use the left and right arrow keys on a keyboard.