Skip to content
RLRLTasksby ulam.ai
AtlasBenchmarksModelsPost-trainingArticlesCompare models
Post-training / CPPO

METHOD NOTEBOOK / CPPO

CPPO (Completion Pruning Policy Optimization)

Accelerates GRPO training by up to 4–8× by dynamically pruning low-contribution completions inside each group G before the expensive backward pass (keeping only completions with large absolute relative advantages |A_i|).

Mechanism

Accelerates GRPO training by up to 4–8× by dynamically pruning low-contribution completions inside each group GGG before the expensive backward pass (keeping only completions with large absolute relative advantages ∣A^i∣|\hat{A}_i|∣A^i​∣).

References

  1. CPPO: Accelerating the Training of Group Relative Policy Optimization-Based Reasoning Models ↗

    Zhihang Lin, Mingbao Lin, Yuan Xie, Rongrong Ji · 2025

CONTINUE READING

RLVR →Pure Zero-SFT RLVR →GRPO →Dr.GRPO →DAPO →VAPO →GSPO →CISPO →Find an evaluation →
RLTasks / The intelligence atlas

LLM benchmarks, model comparisons, and post-training.

Methodology & sourcesModel directoryTraining liftEpoch AI collectionBenchmark referencesPost-training referenceUlam benchmarksPowered by ulam.ai ↗Previous purchases

Independent editorial explanations. Ulam evaluations are identified throughout.
Curated snapshot · 26 October 2026 · LLMs and language-model agents.