Tag
This paper introduces Weightless Fine-Tuning (WFT), a training-free decoding-time method that approximates supervised fine-tuning effects via logit-space transport, achieving competitive personalization performance with less than 7% of the computation.
This paper introduces Decoding-Level Taboo, a runtime logit-space stress test that forces LLMs off their nominal generation paths to evaluate off-path robustness, showing that robustness improves with model scale and instruction alignment.