"Enlightenment" Finetuning Boosts Large Model Capabilities Suddenly

The quest for autonomously self-improving AI models is a significant area of research, particularly with the advent of large foundation models. Drawing an analogy from human "aha moments," this study proposes that large models possess a similar latent capacity for sudden, substantial improvements in their capabilities. To unlock this potential, the researchers developed "Enlightenment," a new post-tuning method that requires no additional training. Instead, it modifies internal shortcuts within critical modules or layers of the model, a departure from existing training-free methods that typically manipulate attention weights. The approach has been instantiated for both large language models (LLMs) and vision-language models. For LLMs, it uses attention head-mixing shortcuts with an adaptive scaling factor. For vision-language models, it applies a scalar-modulated factor to residual connections in decoder layers. Extensive experiments demonstrate that Enlightenment effectively unlocks latent potential, leading to remarkable performance gains on diverse benchmarks and models.