diff --git a/src/app/docs/docs-slug-renderer.test.tsx b/src/app/docs/docs-slug-renderer.test.tsx
index 448a2f14..28c31774 100644
--- a/src/app/docs/docs-slug-renderer.test.tsx
+++ b/src/app/docs/docs-slug-renderer.test.tsx
@@ -2,6 +2,7 @@ import { describe, expect, test } from "bun:test";
import {
buildDocsPageMetadata,
renderDocsSlugPage,
+ resolveLocalDocsShellDescription,
} from "@/app/docs/docs-slug-renderer";
describe("docs slug renderer locale gating", () => {
@@ -243,4 +244,21 @@ describe("docs slug renderer locale gating", () => {
);
}
});
+
+ test("local non-glossary docs shell prefers openingSummary when present", () => {
+ const shellDescription = resolveLocalDocsShellDescription({
+ description:
+ "A post-training reinforcement-learning method that compares several sampled answers for the same prompt and updates the model from their relative quality.",
+ openingSummary:
+ "Group Relative Policy Optimization, usually shortened to GRPO, is a reinforcement-learning post-training method where the model samples several answers to one prompt, scores them as a group, and learns from which answers look better inside that local set.",
+ section: "training",
+ });
+
+ expect(shellDescription).toContain(
+ "Group Relative Policy Optimization, usually shortened to GRPO",
+ );
+ expect(shellDescription).not.toContain(
+ "A post-training reinforcement-learning method that compares several sampled answers for the same prompt and updates the model from their relative quality.",
+ );
+ });
});
diff --git a/src/app/docs/docs-slug-renderer.tsx b/src/app/docs/docs-slug-renderer.tsx
index 437a0027..9b9678bd 100644
--- a/src/app/docs/docs-slug-renderer.tsx
+++ b/src/app/docs/docs-slug-renderer.tsx
@@ -20,6 +20,19 @@ import { localizedRouteAlternates } from "@/lib/i18n/route-locale";
import { source } from "@/lib/source";
import { getMDXComponents } from "../../../mdx-components";
+export function resolveLocalDocsShellDescription(options: {
+ description: string;
+ openingSummary?: string;
+ section: string;
+}) {
+ const { description, openingSummary, section } = options;
+ if (section === "glossary") {
+ return description;
+ }
+
+ return openingSummary ?? description;
+}
+
function buildDocsPageAlternates(docsSlug: string) {
const alternates = localizedRouteAlternates({
surface: "docs-page",
@@ -53,11 +66,16 @@ async function renderLocalDocsPage(
const loadedPage = await loadLocalDocsPage(localRef, locale);
const uiMessages = await loadUiMessages(locale);
+ const shellDescriptionText = resolveLocalDocsShellDescription({
+ description: loadedPage.messages.description,
+ openingSummary: loadedPage.messages.openingSummary,
+ section: localRef.section,
+ });
const description =
localRef.section === "glossary" ? (
) : (
- loadedPage.messages.description
+ shellDescriptionText
);
return (
diff --git a/src/content/docs/training/grpo/assets.json b/src/content/docs/training/grpo/assets.json
new file mode 100644
index 00000000..04ee7b59
--- /dev/null
+++ b/src/content/docs/training/grpo/assets.json
@@ -0,0 +1,10 @@
+{
+ "trainingFlow": {
+ "type": "graph",
+ "graphId": "graph.grpo-training-flow",
+ "webRenderer": "react-flow",
+ "printRenderer": "vertical-svg",
+ "altKey": "assets.trainingFlow.alt",
+ "captionKey": "assets.trainingFlow.caption"
+ }
+}
diff --git a/src/content/docs/training/grpo/messages/en.json b/src/content/docs/training/grpo/messages/en.json
new file mode 100644
index 00000000..7998b9c8
--- /dev/null
+++ b/src/content/docs/training/grpo/messages/en.json
@@ -0,0 +1,94 @@
+{
+ "title": "Group Relative Policy Optimization",
+ "description": "A post-training reinforcement-learning method that compares several sampled answers for the same prompt and updates the model from their relative quality.",
+ "openingSummary": "Group Relative Policy Optimization, usually shortened to GRPO, is a reinforcement-learning post-training method where the model samples several answers to one prompt, scores them as a group, and learns from which answers look better inside that local set.",
+ "sections": {
+ "whatItIs": {
+ "title": "What It Is",
+ "body": "Group Relative Policy Optimization is a policy-update method used after pretraining or supervised fine-tuning. Instead of judging one answer in isolation, it samples a small group of candidate answers for the same prompt and uses their relative ranking to decide which behaviors should be reinforced."
+ },
+ "whyItExists": {
+ "title": "Why It Exists",
+ "body": "A single reward score can be noisy, and PPO-style language-model training often adds the extra cost of a learned critic to estimate a baseline. GRPO tries to keep the signal useful while simplifying the loop: the group itself supplies the local baseline, so the model can learn from which samples look better than their neighbors."
+ },
+ "howItWorks": {
+ "title": "How It Works",
+ "body": "For one prompt, the policy generates several completions, a reward function scores them, and those scores are normalized within that group. Answers above the group average get a positive learning signal, answers below it get a negative one, and the policy is updated so future samples are more likely to resemble the stronger members of the group."
+ },
+ "comparedToNearbyRegimes": {
+ "title": "Compared To Nearby Regimes",
+ "body": "Group Relative Policy Optimization still sits inside the broader alignment family, but it is narrower than Reinforcement Learning from Human Feedback as a full pipeline. Reinforcement Learning from Human Feedback often means collecting preference data, training a reward model, and then running a reinforcement-learning update such as Proximal Policy Optimization. GRPO keeps the reinforcement-learning loop, but it replaces the learned critic-style baseline with relative ranking inside one sampled group. Direct Preference Optimization moves in a different direction: it stays closer to supervised-style optimization on chosen-versus-rejected pairs, while GRPO uses rewards over several sampled answers and updates the policy from that within-group ordering instead of from one pairwise objective alone."
+ },
+ "limitationsAndFailureModes": {
+ "title": "Limitations And Failure Modes",
+ "body": "The method still depends on reward quality. If the reward function prefers shallow tricks, the whole group can drift in the wrong direction together. Relative scoring also means a weak group can still produce a misleading winner if every sampled answer is bad."
+ },
+ "related": {
+ "title": "Related To"
+ },
+ "tags": {
+ "title": "Tags"
+ },
+ "references": {
+ "title": "References"
+ }
+ },
+ "callouts": {
+ "trainingFlowGraph": {
+ "title": "GRPO training flow",
+ "body": "A visual walkthrough of one prompt, grouped sampling, relative scoring, and the policy update."
+ },
+ "trainingFlowLegend": {
+ "title": "Graph legend",
+ "body": "How to read each stage of the GRPO training flow."
+ }
+ },
+ "links": {
+ "trainingFlowLegendPrompt": "One prompt anchors the whole local comparison.",
+ "trainingFlowLegendSampling": "The policy samples several candidate answers for that same prompt.",
+ "trainingFlowLegendRelativeScoring": "Those answers are scored relative to each other inside the sampled group.",
+ "trainingFlowLegendPolicyUpdate": "The policy update reinforces answers that beat the group baseline and discourages weaker ones."
+ },
+ "assets": {
+ "trainingFlow": {
+ "alt": "A training flow from one prompt to a group of sampled answers, then to relative scoring inside the group, and finally to a policy update.",
+ "caption": "GRPO learns from which answers win inside each sampled group instead of relying on one separate critic-estimated baseline."
+ }
+ },
+ "math": {
+ "grpoAdvantage": {
+ "label": "Grouped relative advantage sketch",
+ "formula": "A_i = \\frac{r_i - \\operatorname{mean}(r_{1:G})}{\\operatorname{std}(r_{1:G})}",
+ "variableDefinitions": {
+ "advantage": {
+ "term": "A_i",
+ "definition": "normalized advantage for sampled answer i"
+ },
+ "reward": {
+ "term": "r_i",
+ "definition": "reward score assigned to sampled answer i"
+ },
+ "groupSize": {
+ "term": "G",
+ "definition": "number of sampled answers in the comparison group"
+ }
+ }
+ }
+ },
+ "graph": {
+ "nodes": {
+ "prompt": {
+ "label": "One prompt"
+ },
+ "sampleGroup": {
+ "label": "Sample a group of answers"
+ },
+ "relativeScore": {
+ "label": "Score answers relative to the group"
+ },
+ "policyUpdate": {
+ "label": "Update the policy"
+ }
+ }
+ }
+}
diff --git a/src/content/docs/training/grpo/page.mdx b/src/content/docs/training/grpo/page.mdx
new file mode 100644
index 00000000..e75d83ff
--- /dev/null
+++ b/src/content/docs/training/grpo/page.mdx
@@ -0,0 +1,124 @@
+---
+title: "Group Relative Policy Optimization"
+description: "A post-training reinforcement-learning method that compares several sampled answers for the same prompt and updates the model from their relative quality."
+kind: "training-regime"
+registryId: "training-regime.grpo"
+messageNamespace: "local"
+assetNamespace: "local"
+status: "published"
+tags:
+ - foundations
+aliases:
+ - "GRPO"
+ - "Group Relative Preference Optimization"
+updatedAt: "2026-06-19"
+---
+
+import { CitationList } from "@/features/docs/components/CitationList";
+import { BlockMath } from "@/features/docs/components/Math";
+import { RelatedDocs } from "@/features/docs/components/RelatedDocs";
+import { Section } from "@/features/docs/components/Section";
+import { T } from "@/features/docs/components/T";
+import { TagPillList } from "@/features/docs/components/TagPillList";
+import { TrainingRegimeAtAGlance } from "@/features/models/components/TrainingRegimeAtAGlance";
+import { TrainingRegimeFlow } from "@/features/models/components/TrainingRegimeFlow";
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ -
+
+
+ -
+
+
+ -
+
+
+ -
+
+
+
+
+
+ {Object.entries(definitions).map(([id, row]) => (
+
+ ))}
+
+
+ {displayLabel ? (
+
+ {displayLabel}
+
+ ) : null}
+
+ {mathId ?
: null}
+