Farther the Shift, Sparser the Representation: Analyzing OOD Mechanisms in LLMs

Mingyu Jin, Yutong Yin, Jingcheng Niu†, Qingcheng Zeng, Wujiang Xu, Mengnan Du, Wei Cheng, Zhaoran Wang, Tianlong Chen and Dimitris N. Metaxas.
NeurIPS 2026

arXiv Code

TL;DR

The farther the input shifts out of distribution, the sparser the LLM’s last hidden states become. This sparsity–difficulty relation holds whether difficulty comes from harder reasoning questions, longer contexts, or more answer choices, and across diverse models and domains. Controlled analyses and a learning-dynamics explanation show the sparsity is an adaptive mechanism for stabilising reasoning under OOD, not an accident. Building on this, Sparsity-Guided Curriculum ICL (SG-ICL) uses representation sparsity to schedule few-shot demonstrations, yielding considerable gains.

How to Cite

@inproceedings{jin2026farther,
  title     = {Farther the Shift, Sparser the Representation: Analyzing {OOD} Mechanisms in {LLM}s},
  author    = {Jin, Mingyu and Yin, Yutong and Niu, Jingcheng and Zeng, Qingcheng and Xu, Wujiang and Du, Mengnan and Cheng, Wei and Wang, Zhaoran and Chen, Tianlong and Metaxas, Dimitris N.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026},
}