This research introduces a novel calibration-free LLM compression technique that employs reinforcement learning to determine optimal pruning strategies without external datasets.

• Achieves up to 3× compression ratio without significant performance loss
• Uses a neural policy network to identify which parameters to prune
• Adapts automatically to different compression requirements without retraining
• Outperforms existing methods across multiple benchmarks with faster runtime

Engineering Impact: This approach significantly reduces deployment barriers for resource-constrained environments by enabling more efficient AI systems while maintaining performance integrity.

You Only Prune Once: Designing Calibration-Free Model Compression With Policy Learning