Alignment

Papers on alignment techniques and fine-tuning methods for large language models.

Overview

This section contains 6 papers covering:

• LIMA - Minimal alignment with 1000 examples
• Large Reasoning Models - Learning alignment from flawed thinking
• RECAP - Mitigating capabilities forgetting during RL
• PPO-max - Stabilizing RLHF with careful PPO engineering
• Reducing Sycophancy - Synthetic data to reduce agreement bias
• TÜLU 2 - Advanced instruction-tuning dataset mixture