arXiv cs.CL· Mingyu Jin, Yutong Yin, Jingcheng Niu, Qingcheng Zeng, Wujiang Xu, Mengnan Du, Wei Cheng, Zhaoran Wang, Tianlong Chen, Dimitris N. Metaxas·· 12 小时前AI 评分34
任务越难表示越稀疏：分析 LLM 的 OOD 机制
Farther the Shift, Sparser the Representation: Analyzing OOD Mechanisms in LLMs
AI 导读
研究发现 LLM 面对更难输入时，最后一层隐藏状态会系统性变稀疏，即“任务越难，表示越稀疏”，该规律在多种模型和领域上均可观察到。受控实验表明表示密度是模型习得的数据熟悉度属性：已掌握的模式形成丰富分布式表示，陌生输入则默认更稀疏激活。基于此提出的 SG-ICL 用稀疏度匹配查询难度的少样本示例，在 MATH-500 上优于标准 CoT 和 Auto-CoT 基线。
正文 · 原文
Abstract:In this work, we investigate how the internal representations of Large Language Models (LLMs) change with inputs of increasing difficulty. Although sparsity changes may not be obvious for individual samples, dataset-level results reveal a consistent statistical trend: as task difficulty increases (e.g., harder questions, longer contexts, or more answer choices), the last hidden states of LLMs become systematically sparser. In short, \textbf{the harder the task, the sparser the representation}. This sparsity--difficulty relation is observable across diverse models and domains, suggesting that harder inputs drive more concentrated activation patterns in the last hidden state. Through a series of controlled analyses and learning-dynamics experiments, we show that representational density is a learned property of data familiarity: models develop rich, distributed representations for mastered patterns, while unfamiliar inputs default to sparser activations. Our finding is also actionable. We illustrate this with Sparsity-Guided Curriculum In-Context Learning (SG-ICL), which uses sparsity to select few-shot demonstrations matched to the query's difficulty, outperforming standard CoT and Auto-CoT baselines on MATH-500. Our study provides new insights into how LLM representations reflect task difficulty and how this signal can be leveraged for inference. The source code is available at the URL: this https URL.
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2603.03415 [cs.CL] |
| (or arXiv:2603.03415v4 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2603.03415 arXiv-issued DOI via DataCite |
Submission history
From: Yutong Yin [view email]
[v1]
Tue, 3 Mar 2026 18:48:15 UTC (427 KB)
[v2]
Thu, 19 Mar 2026 17:36:27 UTC (429 KB)
[v3]
Sat, 26 Sep 2026 05:28:48 UTC (1,266 KB)
[v4]
Wed, 4 Nov 2026 05:06:30 UTC (1,258 KB)
来源：arXiv cs.CL · arxiv.org