Early-Signal Detector

method · first seen 4 Oct 2026

group relative policy optimization

Also: GMPO, GRPO, group-relative advantage estimation, reinforcement learning from group-relative policy optimization

Policy gradient method that estimates advantages by normalizing rewards across a group of sampled rollouts for the same prompt.

Group Relative Policy Optimization (GRPO) is a policy-gradient method that drops the learned value critic used in PPO. Instead, it samples a group of completions for each prompt and computes advantages by normalizing their rewards within that group. It was introduced by DeepSeek in 2024 and became a standard post-training recipe for reasoning models. The sources treat it as a default tool rather than a new idea. Papers apply it to video research agents, recommendation, red teaming, battery charging control and coding models. One paper ablates GRPO estimator choices at 1B scale. Repositories list it as a training stage alongside PPO, DAPO and SFT in agentic RL frameworks and distillation pipelines.

GRPO is already the common baseline for RL with verifiable rewards, so the spike in mentions reflects routine use across application papers rather than a new trend. The more useful signal to watch is the stream of variants and ablations that change its advantage estimator.

Analyst note (established), 17 Nov 2026.

Mentions per week

09119 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 90 on 14 Nov 2026 · 179 mentions in 28 days · 172 independent groups · papers → code · first seen 6 weeks ago · acceleration -0.22

Earliest sources

Latest sources

Mentioned together with

reinforcement learning with verifiable rewardscredit assignmentlow-rank adaptationon-policy self-distillationworld modelreward hackingvision-language-action modelretrieval-augmented generationflow matchingon-policy distillation