概念低语而语法高呼:频谱反集中与Transformer表示的双几何
Concepts Whisper: Spectral Anti-Concentration and the Dual Geometry of Transformer Representations
AI总结:
研究通过跨语言概念传输测试,发现因果内积在频谱正则化下无显著差异,揭示了残差流中反集中现象及静态解嵌行对比的高方差集中特性,支持语法优先编码于高方差子空间。
摘要:
我们测试因果内积是否能实现跨语言概念传输。在17个模型和4个语言对中,匹配频谱随机化测试发现白化因果对齐与频谱正则化无显著差异(p=0.95)。然而,这一失败揭示了更广泛的现象:反集中在五个架构家族的残差流均值差向量中被观察到(p<10^-33),并通过SAE特征(如p=4.5×10^-19)和Gemma和Llama的线性探测支持。我们发现双几何:激活空间的概念方向在频谱尾部反集中,而静态解嵌行对比在高方差方向集中(p<10^-4)。Split-injection因果干预支持Gemma和Llama的功能基础(Cohen's d最高达1.80),POS标签探测显示语法优先编码于高方差子空间(p<0.013),Qwen~2.5家族显示显著反转,符合架构特定频谱结构。这些结果表明,Transformer可能在上下文化处理中将语义内容旋转到频谱安静区域,编码概念以减少语法干扰。
英文摘要:
We find that transformer concept representations systematically anti-concentrate in the spectral tail of the unembedding covariance, encoding word-level concepts in low-variance directions across a 17-model core suite and an expanded set of 22 semantic concept categories, with convergent replications from three independent extraction methods. Residual-stream difference-of-means vectors anti-concentrate in all 17 models (model-level one-sample t-test, p = 3.8e-9), and remain more tail-aligned than norm-matched random directions in 13 of 17; convergent support comes from sparse autoencoder (SAE) features (p = 4.5e-19 across concepts within a model) and linear probes on Llama and Qwen. We identify a dual geometry: activation-space concept directions anti-concentrate while static unembedding-row contrasts concentrate in high-variance directions (p < 10^-4). This investigation arose from testing whether the causal inner product of Park et al. (2024) aids cross-lingual concept transport; a matched-spectrum randomization across 17 models and four language pairs finds no evidence that Whitened Causal Alignment improves over spectral regularization alone (p = 0.95). Split-injection interventions, restricted to steering strengths at which both arms remain interpretable, show the predicted interference asymmetry in four of five models (paired Cohen's d_z up to 1.19) with no significant reversal inside that regime, and POS-tag probing across eight models shows syntax preferentially encoded in the high-variance subspace in six of eight architectures, with a significant reversal in the Qwen 2.5 family. These results suggest transformers rotate semantic content into spectrally quiet regions during contextualized processing, where, in some architectures, interventions may reduce grammatical disruption relative to high-variance steering.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。