Nine Emotion Centroids: A Label-Free Valence Axis That Transfers Across Four Modalities
Nine Emotion Centroids: A Label-Free Valence Axis That Transfers Across Four Modalities
九个情感质心：一种跨越四种模态的无标签效价轴
Abstract: Inside a modern language model sits a single internal direction that tracks how positive or negative a sentence feels. We show how to find this valence axis (V-axis) from just 9 emotion category names plus 50 short narrative paragraphs per emotion — about 1,500 fewer labels than the usual supervised approach — and that the same direction appears in vision, audio, and human-brain encoders never jointly trained.
摘要： 在现代语言模型内部，存在一个单一的内部方向，用于追踪句子的情感正负极性。我们展示了如何仅通过 9 个情感类别名称以及每个情感 50 段短叙述性段落来找到这条效价轴（V-axis）——这比通常的监督学习方法减少了约 1,500 个标签——并且发现相同的方向也存在于从未进行过联合训练的视觉、音频和人脑编码器中。
The recipe: embed nine emotion-anchored story sets in a frozen encoder, take the top principal direction of the nine averaged embeddings. Projecting new inputs onto it captures 93% of supervised performance on SST-2 (Llama-3-8B-Instruct, AUC 0.772 vs. 0.828), correlates with human valence ratings on 11,811 EmoSet images at r=0.636, reaches AUC 0.906 on ESC-50 audio (p<2.2e-15), and AUC 0.720+/-0.055 on EEG from 123 subjects (p<3.65e-8).
具体方法是：将九组以情感为锚点的故事集嵌入到一个冻结的编码器中，并提取这九个平均嵌入向量的主成分方向。将新输入投影到该方向上，在 SST-2 任务上可达到监督学习性能的 93%（Llama-3-8B-Instruct 模型，AUC 0.772 对比 0.828）；在 11,811 张 EmoSet 图像上与人类效价评分的相关性达到 r=0.636；在 ESC-50 音频数据集上达到 AUC 0.906 (p<2.2e-15)；在 123 名受试者的脑电图（EEG）数据上达到 AUC 0.720+/-0.055 (p<3.65e-8)。
The direction is mechanistically active: ablating it collapses sentiment accuracy by 5.5-37.2 pp across three LLMs vs. at most 0.88 pp for matched random directions (z>12). A 2-parameter classifier trained on text labels transfers to images (AUC 0.961), audio (0.764), and brain recordings (0.828) without target-modality labels; a generic 16-D subspace stays at chance (0.525). The recipe is bounded to continuous attributes — seven tests on categorical concepts return near-chance — and steering is family-specific (Llama/Mistral yes, Qwen/Gemma no).
该方向在机制上是活跃的：消融该方向会导致三个大语言模型的情感准确率下降 5.5-37.2 个百分点，而匹配的随机方向最多仅导致 0.88 个百分点的下降 (z>12)。一个在文本标签上训练的 2 参数分类器可以在无需目标模态标签的情况下，迁移到图像 (AUC 0.961)、音频 (0.764) 和脑电记录 (0.828) 中；而通用的 16 维子空间则保持在随机水平 (0.525)。该方法仅限于连续属性——在分类概念上的七项测试结果接近随机水平——且引导效果具有模型家族特异性（Llama/Mistral 有效，Qwen/Gemma 无效）。