Generative language models (GLMs) excel at text generation and instruction following, while discriminative language models (DLMs) provide strong semantic representations for understanding tasks. Under the manifold hypothesis, natural language semantics can be regarded as lying on a shared latent semantic manifold, while DLMs and GLMs learn different representations of this manifold in their respective latent spaces. However, due to differences in training objectives and modeling strategies, these representation spaces are not naturally aligned, preventing discriminative representations from being directly utilized by GLMs through simple concatenation or linear transformation. To address this issue, a semantic alignment framework through discriminative conditioning for generation (DisCon4Gen) is proposed. From a manifold-based perspective, DisCon4Gen transforms DLM semantics into continuous conditions for GLMs. It further incorporates both global semantic signals and fine-grained discriminative cues to guide generation, enabling GLMs to better exploit external discriminative semantics. Experiments on five downstream datasets show that DisCon4Gen improves over strong supervised fine-tuning baselines and achieves competitive performance across both semantic understanding and generative reasoning tasks.

Multiple-choice questions (MCQs) are a widely used and vital assessment format for evaluating large language models (LLMs). This study reveals that LLMs are susceptible to “cognitive load” caused by distractor options in MCQs, leading to excessive attention to distractors and consequent vacillation between correct and incorrect options. To mitigate this cognitive burden, we introduce a novel reasoning prompt strategy, called EoT, which effectively reduces cognitive load by steering the model’s attention away from erroneous options. This enables the model to focus more effectively on reasonable answers. Additionally, by documenting the elimination process, EoT enhances the transparency and interpretability of the model’s reasoning. Experimental results demonstrate that EoT, as a plug-and-play approach, significantly reduces cognitive load and improves performance, showcasing its potential to enhance both the accuracy and interpretability of LLMs.