Diver: Large Language Model Decoding with Span-Level Mutual Information Verification
This addresses the issue of output compliance for users of LLMs, offering an incremental enhancement to decoding strategies.
The paper tackles the problem of large language models deviating from input information during decoding by proposing Diver, a method that uses span-level pointwise mutual information verification to select optimal outputs, resulting in significant performance improvements across various tasks.
Large language models (LLMs) have shown impressive capabilities in adapting to various tasks when provided with task-specific instructions. However, LLMs using standard decoding strategies often struggle with deviations from the inputs. Intuitively, compliant LLM outputs should reflect the information present in the input, which can be measured by point-wise mutual information (PMI) scores. Therefore, we propose Diver, a novel approach that enhances LLM Decoding through span-level PMI verification. During inference, Diver first identifies divergence steps that may lead to multiple candidate spans. Subsequently, it calculates the PMI scores by assessing the log-likelihood gains of the input if the candidate spans are generated. Finally, the optimal span is selected based on the PMI re-ranked output distributions. We evaluate our method across various downstream tasks, and empirical results demonstrate that Diver significantly outperforms existing decoding methods in both performance and versatility.