Improving language model answer reranking with efficient adaptation techniques
Improving JEV Entailment Reranking with Span Readouts and Selective Adaptation
Computation and Language
Summary
Choosing the best answer from multiple options is tricky for language models. This paper looks at a way to improve how models pick answers by checking each option separately and reading important parts more carefully. The authors found a method that averages attention over parts of the answers and tweaks the model in a focused, light way. This approach works better and keeps the model easier to train.
What this means in practice
- •For natural language processing teams: Improve accuracy and efficiency of models selecting best answers from multiple choices by using selective adaptation and mean-span pooling.
- •For conversational ai developers: Enhance decision-making in chatbots by better reranking candidate responses with lightweight model updates.
Authors
Haocheng Mei, Yuchuan Tian, Ye Yuan, Enhua Wu, Chao Xu, Allen Chen, Kai Han, Yunhe Wang
Abstract
JEV decision models pair a language model backbone with a lightweight scoring head to select among natural-language options. Direct selection compares candidate answers within one request, whereas entailment reranking evaluates each candidate in a separate natural language inference (NLI) request over fixed relation options. We examine whether the readout and adaptation choices used for direct selection remain effective for cross-request reranking. We combine a mean-span readout with selective backbone adaptation. Mean-span pooling replaces the option-ending pointer with an average of the contextualized states inside each NLI option, without adding head parameters. Selective adaptation retains low-rank updates in attention and gated DeltaNet projections while removing feed-forward adapters. On MC11, which averages accuracy over eleven reranking configurations from seven sources, selective adaptation improves NeoHorse-1-4B accuracy over the all-module baseline while cutting trainable parameters by more than half, and mean-span pooling yields further consistent gains at the same parameter count.