arXiv:2603.22619v1 Announce Type: new Abstract: LLMs often generate seemingly valid answers to flawed or ill-posed inputs. This is not due to missing knowledge: under discriminative prompting, the same models can mostly identify such issues, yet fail to reflect this in standard generative responses. This reveals a fundamental know-act gap between discriminative recognition and generative behavior. Prior work largely characterizes this issue in narrow settings, such as math word problems or question answering, with limited focus on how to integrate these two modes. In this work, we present a comprehensive analysis using FaultyScience, a newly constructed large-scale, cross-disciplinary benchmark of faulty scientific questions. We show that the gap is pervasive and stems from token-level autoregression, which entangles task selection (validate vs. answer) with content generation, preventing discriminative knowledge from being utilized. To address this, we propose DeIllusionLLM, a task-level autoregressive framework that explicitly models this decision. Through self-distillation, the model unifies discriminative judgment and generative reasoning within a single backbone. Empirically, DeIllusionLLM substantially reduces answer-despite-error failures under natural prompting while maintaining general reasoning performance, demonstrating that self-distillation is an effective and scalable solution for bridging the discriminative-generative know-act gap

arXiv:2501.07575v2 Announce Type: replace-cross Abstract: Dataset distillation aims to synthesize a compact yet representative dataset that preserves the essential characteristics of the original data for efficient model training. Existing methods mainly focus on improving data-synthetic alignment or scaling distillation to large datasets. In this work, we propose $\textbf{C}$ommittee $\textbf{V}$oting for $\textbf{D}$ataset $\textbf{D}$istillation ($\textbf{CV-DD}$), an orthogonal approach that leverages the collective knowledge of multiple models to produce higher-quality distilled data. We first establish a strong baseline that achieves state-of-the-art performance through modern architectural and optimization choices. By integrating distributions and predictions from multiple models and generating high-quality soft labels, our method captures a broader range of data characteristics, reduces model-specific bias and the impact of distribution shifts, and significantly improves generalization. This voting-based strategy enhances diversity and robustness, alleviates overfitting, and improves post-evaluation performance. Extensive experiments across multiple datasets and IPC settings demonstrate that CV-DD consistently outperforms single- and multi-model distillation methods and generalizes well to non-training-based frameworks and challenging synthetic-to-real transfer tasks. Code is available at: https://github.com/Jiacheng8/CV-DD.