Skip to main content
Back

Tsinghua's CROPI Trains Smarter AI Reasoning 2.16x Faster on 10% of Data

A new paper from Tsinghua and Z.AI proposes CROPI, using influence functions to pick the most useful prompts for reasoning RL, cutting training data to 10% per stage.

  • Tsinghua CoAI and Z.AI propose CROPI, an influence-function-based data selector for RLVR, accepted at ACL 2026.
  • On a 1.5B model, CROPI reports 2.66x step-level speedup using only 10% of data per stage.
  • Wall-clock speedup after selection overhead is about 2.16x versus full-data GRPO.
  • Off-policy gradients from base-model trajectories avoid the 9x cost of fresh rollouts.
  • Sparse random projection drops 90% of gradient dims and improves rank precision from 13% to 80%.
  • Code and pipeline available at github.com/thu-coai/CROPI.

Reinforcement learning with verifiable rewards (RLVR) has become the default recipe for improving reasoning in language models, but it burns compute on prompts that barely move the policy. A new paper from Tsinghua's CoAI Group and Z.AI, accepted to ACL 2026, proposes CROPI, a curriculum RL framework that uses influence functions to score which training prompts actually matter for the current policy, then trains only on the top slice.

CROPI stands for Curriculum RL with Off-Policy Influence guidance. Rather than relying on repeated online trial-and-error to find useful training data, it estimates which prompts are most helpful for the current policy, selects a compact high-value subset, and runs RL on that subset. Less wasted compute, more targeted learning, cleaner loop.

Why influence functions were stuck outside RLVR

Influence functions are a classic statistical tool that estimate how much a single training example contributes to a model's objective. They have solid theoretical grounding and a track record in pretraining and supervised fine-tuning. Existing data selection methods for RLVR, by contrast, are mostly heuristic, with no guarantees and limited generalization. CROPI applies influence functions to estimate each prompt's contribution to the learning objective, and sidesteps the prohibitive cost of online rollouts by approximating influence from pre-collected offline trajectories.