NOPD Boosts Vision-Language Models Without External Supervision

Researchers have introduced NOPD (Noisy Student On-Policy Self-Distillation), a straightforward yet highly effective self-distillation technique designed to enhance Vision-Language Models (VLMs). This method addresses the limitation of current post-training approaches that typically rely on human annotations, external model distillation, reinforcement learning with human feedback, or verifiable answers. NOPD allows VLMs to improve without any external supervision. The core idea behind NOPD is to leverage the natural discrepancies that arise when a model processes both clean and corrupted versions of the same input, using these differences to generate a self-supervision signal. Specifically, the model learns from corrupted inputs while simultaneously using its own predictions on clean inputs as token-level supervision. NOPD has demonstrated its effectiveness across five visual reasoning tasks, matching and even surpassing the performance of reinforcement learning and external distillation methods. Notably, training with just 2.1K samples from Geometry3K led to a 20-point improvement for Qwen2.5-VL-7B on its validation set and a 7.4-point gain on the out-of-distribution MathVista benchmark. This approach is generalizable, showing improvements across three different VLM models on 12 benchmarks.