From a359b8c45ff2093ca299debcf618c83198613d62 Mon Sep 17 00:00:00 2001 From: Le-Brown Date: Wed, 5 Aug 2026 10:36:52 +0800 Subject: [PATCH] Drop \! spacing command from the KL formula Same kramdown-escaping issue as the earlier \{ \} and * fixes: \! (negative thin space) has its backslash stripped by kramdown before MathJax sees it, leaving a stray literal "!" rendered next to KL. Removing it is a no-op typographically (KL(...) reads fine without the extra spacing tweak) and sidesteps the bug. Co-Authored-By: Claude Sonnet 5 --- _posts/2026-08-04-rombrl-post.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/_posts/2026-08-04-rombrl-post.md b/_posts/2026-08-04-rombrl-post.md index 0405abda16..090fdd5364 100644 --- a/_posts/2026-08-04-rombrl-post.md +++ b/_posts/2026-08-04-rombrl-post.md @@ -44,7 +44,7 @@ $$\max_\theta J(\theta, \phi') \quad \text{s.t.} \quad \phi' \in \arg\min_{\phi where $\Phi$ is an uncertainty set of world models $\phi \in \mathcal{M}$ anchored to the maximum-likelihood estimate $\hat\phi$ — a KL trust region around what the offline data actually supports: -$$\mathbb{E}_{(s,a)\sim\mathcal{D}}\left[ \mathrm{KL}\!\left(P_{\hat\phi}(\cdot|s,a) \,\|\, P_\phi(\cdot|s,a)\right) \right] \le \epsilon$$ +$$\mathbb{E}_{(s,a)\sim\mathcal{D}}\left[ \mathrm{KL}\left(P_{\hat\phi}(\cdot|s,a) \,\|\, P_\phi(\cdot|s,a)\right) \right] \le \epsilon$$ (conservatism still matters offline, we're not letting the model run wild). The policy maximizes its return against the *worst* model in that set; the model adversarially minimizes it.