PAOSC: Plug-and-play Attention Optimization for Semantic Consistency in LLMs

Attention mechanisms are essential to the success of Large Language Models (LLMs). In practice, models often overemphasize semantically low-value tokens, forming attention sinks while failing to capture truly informative tokens. Existing inference-time optimization methods mainly rely on static adjustments or attention redistribution, which often disrupt the correspondence between attention distribution and the actual semantics of the input, leading to a loss of semantic consistency and degraded performance. To address this problem, we propose PAOSC, a plug-and-play attention optimization model designed to maintain semantic consistency by dynamically adjusting attention. PAOSC employs a generator to identify informative tokens and a discriminator to optimize the generator via policy gradients based on confidence changes and loss fluctuations. Experiments on eight LLMs show up to a 9.68% improvement in the F1 score. On the constructed HTTP-RL dataset, PAOSC eliminates 18% of low-value tokens, improving inference efficiency while maintaining semantic consistency. Our code is available at https://github.com/ChangLi000/PAOSC.