Rohan Paul 谈 Qwen3.8-27B RL post-train 的智能体表现:幻觉率从 30.4% 降至 15.1%
中文摘要
Rohan Paul 指出 Qwen3.8-27B 通过 RL 后训练将智能体幻觉率从 30.4% 降至 15.1%。
English Summary
Rohan Paul noted Qwen3.8-27B's RL post-training reduced agent hallucination rates from 30.4% to 15.1%.
原文节选
Rohan Paul 观察到今年不少亮眼 Agent 结果来自小开源模型的 post-train。