Farther the Shift, Sparser the Representation: Analyzing OOD Mechanisms in LLMs
Mingyu Jin, Yutong Yin, Jingcheng Niu, Qingcheng Zeng · 2026
Read Farther the Shift, Sparser the Representation: Analyzing OOD Mechanisms in LLMs by Mingyu Jin, Yutong Yin, Jingcheng Niu, Qingcheng Zeng (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-03415.txt.
Description
Mingyu Jin and three others find a simple pattern in how language models handle unfamiliar inputs. As tasks drift further from training territory, through harder reasoning, longer contexts, or extra answer choices, the model's final internal representations grow sparser. Fewer units carry the signal. Their 2026 study measures this shift across many models and domains and finds it consistent and quantifiable.
Controlled analyses argue the sparsity is no accident. It works as an adaptive mechanism, concentrating computation into specialized subspaces to stabilize reasoning under strain. A learning-dynamics account supports that reading. The team then turns the effect into a tool: a curriculum method that uses measured sparsity to schedule practice cases in context. That scheduling brought considerable performance gains, and the code is public.
Catalog record
| CREATOR | Mingyu Jin, Yutong Yin, Jingcheng Niu, Qingcheng Zeng |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 57212 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.03415