Research Papers 论文研究 2d ago Updated 1d ago 更新于 1天前 48

Nine Emotion Centroids: A Label-Free Valence Axis That Transfers Across Four Modalities 九种情感质心:一种跨四种模态迁移的无标签效价轴

A single internal direction (V-axis) encoding emotional valence can be extracted from LLMs using only 9 emotion category names and ~450 narrative paragraphs, requiring roughly 1,500 fewer labels than standard supervised methods The same valence direction transfers across four modalities—text, vision, audio, and human EEG—despite never being jointly trained, demonstrating cross-modal alignment of emotional representation Mechanistic ablation confirms the direction is causally responsible for sent 发现现代语言模型内部存在一个"效价轴"(V-axis),可追踪文本的积极/消极程度,仅需9个情绪类别+每类50段故事(约1500标签)即可定位 该效价轴具有跨模态迁移能力,在视觉、音频、人脑EEG编码器中均被发现(从未联合训练),2参数分类器可跨模态迁移 消融实验证实该方向机制上活跃:消融后LLM情感准确率下降5.5-37.2pp,而随机方向仅下降≤0.88pp 方法仅适用于连续属性(情感效价),对分类概念测试接近随机;模型家族特异性明显(Llama/Mistral有效,Qwen/Gemma无效)

65
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • A single internal direction (V-axis) encoding emotional valence can be extracted from LLMs using only 9 emotion category names and ~450 narrative paragraphs, requiring roughly 1,500 fewer labels than standard supervised methods
  • The same valence direction transfers across four modalities—text, vision, audio, and human EEG—despite never being jointly trained, demonstrating cross-modal alignment of emotional representation
  • Mechanistic ablation confirms the direction is causally responsible for sentiment tracking, with accuracy drops of 5.5–37.2 percentage points versus at most 0.88 pp for matched random directions
  • A lightweight 2-parameter classifier trained on text labels transfers to images (AUC 0.961), audio (AUC 0.764), and brain recordings (AUC 0.828) without any target-modality labels
  • The approach is bounded to continuous attributes (categorical concept tests return near-chance) and steering effectiveness is family-specific (works on Llama/Mistral, not Qwen/Gemma)

Why It Matters

This work demonstrates that emotional valence is a robust, cross-modal feature encoded in AI systems, discoverable with dramatically fewer labels than conventional supervised approaches. For AI practitioners, it opens a low-cost pathway to inject emotional awareness into multimodal systems without expensive annotation pipelines. For researchers, it provides evidence that continuous psychological constructs like valence have shared representational geometry across fundamentally different input modalities.

Technical Details

  • Method: Embed nine emotion-anchored story sets in a frozen encoder, compute the top principal component of the nine averaged embeddings to extract the V-axis direction
  • Text evaluation: On SST-2 with Llama-3-8B-Instruct, the label-free V-axis achieves AUC 0.772 versus 0.828 for fully supervised baselines (93% of supervised performance)
  • Cross-modal transfer: Correlation with human valence ratings on 11,811 EmoSet images at r=0.636; AUC 0.906 on ESC-50 audio (p<2.2e-15); AUC 0.720±0.055 on EEG from 123 subjects (p<3.65e-8)
  • Ablation study: Removing the V-axis collapses sentiment accuracy by 5.5–37.2 pp across three LLMs, while matched random directions cause at most 0.88 pp drop (z>12), confirming mechanistic specificity
  • Classifier transfer: A 2-parameter linear classifier trained on text labels transfers to images (AUC 0.961), audio (0.764), and brain recordings (0.828); a generic 16-D subspace performs at chance (0.525), ruling out dimensionality artifacts
  • Limitations: Steering is family-specific (effective on Llama/Mistral, ineffective on Qwen/Gemma); the recipe is bounded to continuous attributes with near-chance results on seven categorical concept tests

Industry Insight

  • The dramatic reduction in labeling requirements (~1,500 fewer labels) makes cross-modal emotional alignment economically viable for resource-constrained teams, potentially replacing expensive human annotation with lightweight narrative-based extraction
  • Family-specific steering suggests that interpretability and controllability techniques cannot be assumed portable across model architectures; practitioners should validate mechanistic interventions on each model family before deployment
  • The cross-modal transfer of a single continuous attribute raises the possibility that other psychological dimensions (arousal, dominance, etc.) may share similar transferable representations, warranting systematic exploration for building emotionally intelligent multimodal systems

TL;DR

  • 发现现代语言模型内部存在一个"效价轴"(V-axis),可追踪文本的积极/消极程度,仅需9个情绪类别+每类50段故事(约1500标签)即可定位
  • 该效价轴具有跨模态迁移能力,在视觉、音频、人脑EEG编码器中均被发现(从未联合训练),2参数分类器可跨模态迁移
  • 消融实验证实该方向机制上活跃:消融后LLM情感准确率下降5.5-37.2pp,而随机方向仅下降≤0.88pp
  • 方法仅适用于连续属性(情感效价),对分类概念测试接近随机;模型家族特异性明显(Llama/Mistral有效,Qwen/Gemma无效)

为什么值得看

本文为"机制可解释性"研究提供了低成本、高迁移性的新范式,证明只需极少标注即可在冻结编码器中提取有意义的语义方向。跨模态一致性(文本→图像→音频→脑电)暗示情感效价可能是多模态表征的通用底层结构,对AI安全、可控生成和神经科学交叉研究具有重要参考价值。

技术解析

  • 方法核心:在冻结编码器中嵌入9个情绪锚定故事集,计算9个平均嵌入的主成分方向(PCA第一主成分),即为V-axis。仅需约1500个标签,比传统监督方法少约1500个。
  • 基准性能:SST-2上AUC 0.772(监督基线0.828,达93%性能);EmoSet图像r=0.636;ESC-50音频AUC 0.906;EEG(123受试者)AUC 0.720±0.055。
  • 跨模态迁移:2参数分类器仅在文本标签上训练,可迁移至图像(AUC 0.961)、音频(0.764)、脑电(0.828);而通用16维子空间仅达随机水平(0.525)。
  • 机制验证:消融V-axis导致三个LLM情感准确率下降5.5-37.2pp,匹配随机方向仅下降≤0.88pp(z>12),证明该方向非偶然。
  • 边界条件:七项分类概念测试接近随机;模型家族特异性显著(Llama/Mistral可引导,Qwen/Gemma无效)。

行业启示

  • 低成本可解释性工具:证明仅需极少标注即可提取模型内部语义方向,为AI可解释性研究提供可扩展范式,降低机制分析成本。
  • 跨模态统一表征线索:V-axis在文本、视觉、音频、脑电中的一致性,暗示情感效价可能是多模态学习的通用底层结构,对多模态模型设计和神经科学交叉研究有启发。
  • 模型家族差异需重视:引导效果在Llama/Mistral有效而在Qwen/Gemma无效,提示可解释性技术存在架构依赖性,实际部署前需针对性验证。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Multimodal 多模态 Embedding Model 嵌入模型 Research 科学研究