提出轨迹追踪方法,精准定位概念在模型中的稳定表示层。
Geometric Evolution Maps: Extracting Stable Concept Probes from Transformer Residual Streams

- 通过记录概念方向在装配区的演化路径,找到其稳定后的输出层。
- 稳定方向能更有效抑制概念分离,且效果优于随机方向。
- 适合需要精确定位概念表示的研究者,如模型可解释性分析。
概念探测器的可靠性取决于所取层级。传统在固定深层或分离曲线峰值处探测,忽略了概念形成过程中的关键结构特征:探测方向在概念装配区(CAZ)内持续旋转,直至该区段末尾才趋于稳定。本文提出几何演化图(GEM),记录概念方向在单个CAZ段内的轨迹,选取该段末层的稳定方向,并确定其后首个可独立评估该方向的交接层(手办层)。该交接层由段边界决定,而非旋转标准检测。一个概念通常占据多个段,这些段构成其‘地图集’(atlas)。方向旋转现象显著且跨架构、概念类型一致,是表征本身的特性,非估计误差。所提取方向具有因果性:消融它比消融随机方向更显著削弱概念分离,但无单一位置承载全部效应。在旋转完成后探测精度更高,深度匹配对照表明优势源于更深探测,而非交接边界特有。少数结构性例外被系统记录而非忽略。因此,该方法交付的单位是概念的完整地图集,而非单一方向。GEM为本项目中CAZ验证论文提供消融靶点方法。
原文摘要 · Abstract (English)
A concept probe is only as reliable as the layer it is taken from. Probing at a fixed late layer, or at the peak of a separation curve, ignores a structural feature of how concepts form: the probe direction rotates substantially during assembly and does not settle until after the Concept Allocation Zone (CAZ) in which it forms. We introduce Geometric Evolution Maps (GEMs). A GEM records a concept's directional trajectory across one CAZ segment, takes the settled direction at that segment's final layer, and locates the handoff layer immediately beyond it - the first layer at which that direction can be evaluated outside the window it was estimated on. The handoff layer is derived from the segment boundary, not detected by a rotation criterion. A concept typically occupies several segments; we call that set its atlas. This rotation is substantial and consistent across architectures and concept types - a property of the representation, not an estimation artifact. The extracted direction is causal: ablating it suppresses far more separation than ablating a random direction, though no single site carries the effect alone. Probing after rotation completes is more precise than probing during it; a depth-matched control shows this advantage comes from probing at greater depth, not any privilege specific to the handoff boundary. A small number of structured exceptions are documented rather than left unaccounted for. A concept's atlas - its full set of GEMs, not any single one of them - is therefore the unit this method delivers for downstream use. GEM provides the ablation-targeting methodology used by the CAZ Validation paper in this program.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品;中文卡片由大模型生成,请以原文为准。