Language-model training and adaptation · MLD-0500
GRPO
A policy-optimization method estimating response advantages relative to a group of sampled responses, avoiding a separate value-model baseline in its original design.
Original reference page · detailed expansion pending. Original definitions, examples, notes and citations are preserved below. New examples are illustrative, not benchmark results.
Original example
Train on verifiable reasoning rewards.
Original caveat
Implementations vary in normalization, clipping, and KL treatment.
Context and aliases
Level: Intermediate. Stages: Training · Adaptation · Evaluation.
Tags: DL · SLM · LLM · NLP.
Aliases: Group relative policy optimization.
Related concepts
Original references
- DeepSeekMath: Group Relative Policy Optimization ↗ · Consulted
- TRL documentation ↗ · Consulted
- Original reference URL ↗