Breaking the Likelihood Trap: Variance-Calibrated Modulation for Large Language Model Decoding

The article introduces Variance-Calibrated Modulation (VCM), a training-free pre-decoding method designed to mitigate the "likelihood trap" in large language models (LLMs). VCM employs two mechanisms: Contextual Searchlight via Pointwise Mutual Information (PMI) to enhance contextually relevant tokens while suppressing stopwords, and Adaptive Self-Debiasing for scale-invariant penalization based on logit standard deviation. This approach improves the diversity and coherence of generated text across tasks such as open-ended generation and factual question answering, with minimal computational overhead, making it a valuable tool for practitioners aiming to enhance LLM performance.