arXiv cs.AI· Rhitabrat Pokharel, Ameeta Agrawal, Tanay Nagar·· 10 天前AI 评分25
CLAS：面向多语言大模型的跨语言激活引导方法
Cross-Lingual Activation Steering for Multilingual Language Models
AI 导读
研究者提出免训练的推理时干预方法 CLAS（Cross-Lingual Activation Steering），通过选择性调节神经元激活来缩小大模型的多语言性能差距，在分类和生成基准上分别平均提升 2.3%（Acc.）和 3.4%（F1），同时保持高资源语言性能。该方法无需修改模型权重，且性能提升与语言聚类分离度增加相关，表明有效迁移依赖功能分化而非严格对齐。
正文 · 原文
Abstract:Large language models exhibit strong multilingual capabilities, yet significant performance gaps persist between dominant and non-dominant languages. Prior work attributes this gap to imbalances between shared and language-specific neurons in multilingual representations. We propose Cross-Lingual Activation Steering (CLAS), a training-free inference-time intervention that selectively modulates neuron activations. We evaluate CLAS on classification and generation benchmarks, achieving average improvements of 2.3% (Acc.) and 3.4% (F1) respectively, while maintaining high-resource language performance. We discover that effective transfer operates through functional divergence rather than strict alignment; performance gains correlate with increased language cluster separation. Our results demonstrate that targeted activation steering can unlock latent multilingual capacity in existing models without modification to model weights.
| Comments: | Accepted to INLG 2026 |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2601.16390 [cs.CL] |
| (or arXiv:2601.16390v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2601.16390 arXiv-issued DOI via DataCite |
Submission history
From: Rhitabrat Pokharel [view email]
[v1]
Fri, 23 Jan 2026 01:41:17 UTC (461 KB)
[v2]
Tue, 6 Oct 2026 17:23:35 UTC (645 KB)
来源：arXiv cs.AI · arxiv.org