arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2406.05883cs.LGcs.ITmath.ITstat.ML
收藏

大语言模型中策略对齐的信息论保证

Information Theoretic Guarantees For Policy Alignment In Large Language Models

IBM Research(IBM研究院)

Youssef Mroueh

AI总结:

本文研究大语言模型策略对齐的奖励改进上界,证明在次高斯尾条件下√KL信息论上界成立,并利用顺序统计量Rényi表示与数据处理不等式为n中最佳策略给出任意f-散度的KL上界,同时分析代理奖励到黄金奖励的转移退化。

摘要:

大语言模型的策略对齐指的是约束策略优化,其中策略被优化以最大化奖励,同时在诸如 $\mathsf{KL}$ 散度等 $f$-散度方面保持接近参考策略。$n$ 中最佳对齐策略从参考策略中选取在 $n$ 个独立样本中具有最大奖励的样本。对于这两种情况(策略对齐和 $n$ 中最佳),近期工作通过实验表明,对齐策略相对于参考策略的奖励改进大致按 $\sqrt{\mathsf{KL}}$ 缩放,并且对于 $n$ 中最佳策略的 $\mathsf{KL}$ 有一个关于 $n$ 的显式界。我们在本文中证明,如果参考策略下的奖励具有次高斯尾部,则 $\sqrt{\mathsf{KL}}$ 信息论上界成立。此外,我们针对 $n$ 中最佳策略证明,借助顺序统计量的 Rényi 表示和数据处理不等式,可以通过归约到指数顺序统计量为任意 $f$-散度获得 $\mathsf{KL}$ 上界。如果已知对齐策略尾部的额外信息,我们证明可以通过 Rényi 散度获得对奖励改进的更紧控制。最后,我们展示了这些上界如何从代理奖励转移到黄金奖励,由于代理奖励的高估和近似误差,这会导致黄金奖励改进的下降。

英文摘要:

Policy alignment of large language models refers to constrained policy optimization, where the policy is optimized to maximize a reward while staying close to a reference policy with respect to an $f$-divergence such as the $\mathsf{KL}$ divergence. The best of $n$ alignment policy selects a sample from the reference policy that has the maximum reward among $n$ independent samples. For both cases (policy alignment and best of $n$), recent works showed empirically that the reward improvement of the aligned policy on the reference one scales like $\sqrt{\mathsf{KL}}$, with an explicit bound in $n$ on the $\mathsf{KL}$ for the best of $n$ policy. We show in this paper that the $\sqrt{\mathsf{KL}}$ information theoretic upper bound holds if the reward under the reference policy has sub-gaussian tails. Moreover, we prove for the best of $n$ policy, that the $\mathsf{KL}$ upper bound can be obtained for any $f$-divergence via a reduction to exponential order statistics owing to the Rényi representation of order statistics, and a data processing inequality. If additional information is known on the tails of the aligned policy we show that tighter control on the reward improvement can be obtained via the Rényi divergence. Finally we demonstrate how these upper bounds transfer from proxy rewards to golden rewards which results in a decrease in the golden reward improvement due to overestimation and approximation errors of the proxy reward.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑