Meimingwei Li

The Logit Refiner is introduced, a lightweight autoregressive module that restores intra-scale dependencies by sequentially sampling tokens conditioned on frozen backbone features and generalizes to text-to-image generation, confirming that the mean-field bottleneck persists across VAR variants and is effectively alleviated by the method.

VCM reshapes LLM output distributions to enhance diversity and coherence, effectively breaking the cycle of repetitive degeneration in text generation.

Hyperfitting's surprising generation improvements aren't just temperature scaling – they stem from a "Terminal Expansion" in the final transformer block that dynamically reorders token ranks.

Forget temperature tuning: Min-$k$ sampling finds the "semantic cliff" in your LLM's logits, delivering robust and high-quality text even when other methods fall apart.