原帖引用 · AI 翻译
标题：“后训练中的锐化税”
Meta 的一篇新论文发现，强化学习后训练让智能体在首次尝试时表现更好，但在获得 128 次尝试机会时，最终找到解决方案的能力反而更差。
人们以为强化学习会教会智能体调用工具等新技能。
这种说法主要来自数学和编程领域，而基础模型在这些领域已经见过大量数据。
他们让 14 对基础模型和后训练模型在 3 个智能体基准测试中运行，每项任务最多尝试 128 次。他们还开发了 PTGS，这是一种采样器，会对困难提示提高温度，对简单提示降低温度。
在 WebShop 上，gemma-4-31B 基础模型在 128 次尝试内解决了超过 85% 的任务，而后训练版本解决了 56%。在 42 个模型与基准测试组合中，有 36 个的“锐化税”为正值，这意味着任务覆盖率缩小了。
为能够重试的智能体选择后训练模型之前，应衡量它最终能解决多少任务，而不只是首次尝试的准确率。
原文
New Meta paper finds Reinforcement learning post-training makes an agent better on its first attempt and worse at ever finding a solution when it gets 128 attempts. People assume reinforcement learning teaches agents new skills like tool calling. That claim comes mostly from math and coding, where base models have already seen huge amounts of data. They ran 14 base and post-trained model pairs on 3 agentic benchmarks with up to 128 tries per task. They also built PTGS, a sampler that raises temperature on hard prompts and lowers it on easy ones. On WebShop a gemma-4-31B base model solved over 85% of tasks within 128 tries against 56% for the post-trained version. Their Sharpening Tax came out positive in 36 of 42 model and benchmark pairs, meaning coverage shrank. Before picking a post-trained model for an agent that can retry, measure how many tasks it ever solves, not just first-try accuracy.