LIMR: Less is More for RL Scaling

Free

Less is More for RL Scaling

FreeFree tier

Type

Open Source

Visit Website GitHub (40 stars)

About LIMR: Less is More for RL Scaling

LIMR (Less is More for RL Scaling) is an open-source research method from a 2025 arXiv paper that challenges the assumption that scaling up reinforcement learning training data inherently improves language model reasoning. It introduces Learning Impact Measurement (LIM), an automated technique to evaluate and prioritize training samples based on alignment with model learning trajectories. Using just 1,389 strategically selected samples, LIMR outperforms the full 8,523-sample dataset, achieving 16.7% higher accuracy on AIME24 and surpassing other data-efficient approaches like LIMO and s1 by 13.0% and 22.2% on MATH500. The method is applied directly to base models without distillation and is fully open-sourced, including implementation, training/evaluation code, curated datasets, and trained models.

Key Features

Learning Impact Measurement (LIM) for automated training sample selection

Uses only 1,389 strategically selected samples vs. full 8,523-sample dataset

Achieves 16.7% higher accuracy on AIME24 and outperforms LIMO and s1 on MATH500

Applied directly to base models without distillation from larger models

Fully open-sourced: implementation, code, datasets, and trained models available

Pros & Cons

Pros

Demonstrates that precise sample selection can outperform larger datasets in RL
Automated, reproducible method for evaluating training sample impact
Significant performance gains on standard reasoning benchmarks
Open-sourced for community use and further innovation

Best For

Enhancing reasoning capabilities of large language models through reinforcement learningData-efficient RL training for math reasoning benchmarks (AIME, MATH)Selecting high-impact training samples to improve model performance with fewer resources

FAQ

What is LIMR?

LIMR stands for Less is More for RL Scaling, a method that uses Learning Impact Measurement (LIM) to select a small, strategic subset of training data for reinforcement learning, achieving better reasoning performance than using the full dataset.

Is LIMR open source?

Yes, the authors have open-sourced the implementation of LIM, training and evaluation code, curated datasets, and trained models for reproducible research.

On what benchmarks was LIMR evaluated?

LIMR was evaluated on AIME24 and MATH500, achieving higher accuracy compared to using the full 8,523-sample dataset and outperforming methods like LIMO and s1.