Large Language Models (LLMs) require high quality preference datasets to align with human preferences. However, conventional methods for constructing such datasets face significant challenges: reliance on pre-collected instructions often leads to distribution mismatches with target models, while the need for sampling multiple stochastic responses introduces substantial computational overhead. In this work, we explore a paradigm shift by leveraging inherent regulation of LLMs’ representation space for efficient and tailored preference dataset construction, named Icon². Specifically, it first extracts layer-wise direction vectors to encode sophisticated human preferences and then uses these vectors to filter self-synthesized instructions based on their inherent consistency. During decoding, bidirectional inherent control is applied to steer token representations, enabling the precise generation of response pairs with clear alignment distinctions. Experimental results demonstrate significant improvements in both alignment and efficiency. Llama3-8B and Qwen2-7B achieve an average win rate improvement of 13.89% on AlpacaEval 2.0 and 13.45% on Arena-Hard, while reducing computational costs by up to 48.1%.

Large language models (LLMs) have achieved remarkable advancements in natural language processing. However, the massive scale and computational demands of these models present formidable challenges when considering their practical deployment in resource-constrained environments. While techniques such as chain-of-thought (CoT) distillation have displayed promise in distilling LLMs into small language models (SLMs), there is a risk that distilled SLMs may still inherit flawed reasoning and hallucinations from LLMs. To address these issues, we propose a twofold methodology: First, we introduce a novel method for distilling the self-evaluation capability from LLMs into SLMs, aiming to mitigate the adverse effects of flawed reasoning and hallucinations inherited from LLMs. Second, we advocate for distilling more comprehensive thinking by incorporating multiple distinct CoTs and self-evaluation outputs, to ensure a more thorough and robust knowledge transfer into SLMs. Experiments on three NLP benchmarks demonstrate that our method significantly improves the performance of distilled SLMs, offering a new perspective for developing more effective and efficient SLMs in resource-constrained environments.