OPERNEX LEARNING

Learn Agentic AI

A structured, practical learning path from language model foundations to intelligent agentic systems.

6 Learning parts
453 Subtopics
415 Published lessons
RL Methods for LLMs / GRPO - Group Relative Policy Optimization / GRPO Variants and Extensions
Lesson 285 of 415 5 min read
7.5.2.2

Asymmetric Clipping: Clip-Higher

Learn how DAPO uses a wider upper clipping bound for positive advantages while retaining conservative clipping for negative advantages.

What is it?

DAPO uses asymmetric clipping, called Clip-Higher. For positive advantages, it permits a wider upper importance-ratio bound. For non-positive advantages, it retains the usual conservative clipping range.

Key points

Standard clipping: [1 − ε, 1 + ε]. DAPO positive advantage: [1 − ε, 1 + εhigh]. DAPO non-positive advantage: [1 − ε, 1 + ε]. εhigh is greater than ε. Typical values: ε = 0.2 and εhigh = 0.28.

Clip-Higher Example

With ε = 0.2 and εhigh = 0.28, the positive-advantage upper bound becomes 1.28 instead of 1.20. The policy can move further toward a good token. For a non-positive advantage, the upper bound remains 1.20, preserving the usual conservative clipping.

Real-world application

Use asymmetric clipping when you want the policy to have more room to reinforce positively advantaged tokens without relaxing the negative-advantage constraint.
Key takeaway

DAPO allows more upward movement for good tokens while retaining conservative clipping for non-positive advantages.