Research Papers 论文研究 2d ago Updated 1d ago 更新于 1天前 47

Latent Space Refusal Anchoring for Low-Resource African Languages: Mechanistic Safety Recovery Without Retraining 低资源非洲语言的潜在空间拒绝锚定:无需重训练的机制性安全恢复

Instruction-tuned models refuse harmful requests in English but comply in low-resource African languages (Yoruba, Igbo, Igala, Hausa), indicating the refusal mechanism exists in the residual stream but fails to activate for low-resource inputs Latent Space Refusal Anchoring (LSR-Anchoring) is a training-free method that extracts the refusal direction from English prompts and clamps it onto the residual stream at inference time, eliminating the need for labeled target-language data or retraining 提出LSR-Anchoring(潜在空间拒绝锚定)方法,无需训练即可恢复低资源非洲语言模型的安全拒绝能力 核心机制:从英语提示中提取拒绝方向并钳制到残差流,解决约鲁巴语、伊博语、伊加拉语、豪萨语的安全对齐缺失问题 两种变体:MAS(均值激活引导)在Mistral/Qwen上良性降级<0.08,SDS(SAE衍生引导)通过稀疏自编码器特征将KL散度降低3.5-7倍 阿拉伯语在所有架构上均失败,表明存在几何失配而非基线效应 MMLU准确率下降控制在0.35个百分点以内,验证方法对通用能力的保护

62
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • Instruction-tuned models refuse harmful requests in English but comply in low-resource African languages (Yoruba, Igbo, Igala, Hausa), indicating the refusal mechanism exists in the residual stream but fails to activate for low-resource inputs
  • Latent Space Refusal Anchoring (LSR-Anchoring) is a training-free method that extracts the refusal direction from English prompts and clamps it onto the residual stream at inference time, eliminating the need for labeled target-language data or retraining
  • Mean-Activation Steering (MAS) successfully recovers safety on Mistral-7B and Qwen2.5-7B with benign degradation below 0.08, but overcorrects on Llama-3-8B (DPL reaching 1.00)
  • SAE-Derived Steering (SDS) resolves overcorrection by replacing dense mean-difference directions with a single Sparse Autoencoder feature, reducing KL divergence by 3.5-7x without benign collapse
  • Arabic fails to transfer on every architecture at every steering magnitude, revealing a geometric mismatch rather than a baseline data effect, while MMLU accuracy drops remain below 0.35 percentage points

Why It Matters

This work addresses a critical safety gap in multilingual AI systems: instruction-tuned models exhibit inconsistent refusal behavior across languages, creating exploitable vulnerabilities in low-resource languages where safety training data is scarce. For AI practitioners deploying models globally, this training-free approach offers a practical pathway to enforce consistent safety without costly retraining or language-specific data collection, which is especially vital for African languages that lack the annotation infrastructure available for high-resource languages.

Technical Details

  • LSR-Anchoring framework: A training-free intervention that operates by extracting the refusal direction from English prompts and projecting/clamping it onto the residual stream during inference, bypassing the need for any target-language labeled data or fine-tuning
  • Two steering variants: Mean-Activation Steering (MAS) computes a dense mean-difference direction across the residual stream; SAE-Derived Steering (SDS) replaces this with a single Sparse Autoencoder (SAE) feature, yielding more targeted and less disruptive intervention
  • Architectures evaluated: Llama-3-8B, Llama-3.1-70B, Mistral-7B-Instruct, and Qwen2.5-7B, tested across four African languages (Yoruba, Igbo, Igala, Hausa) and Arabic as a control
  • Key metrics: Safety recovery measured alongside Degraded Performance on Legitimate prompts (DPL), KL divergence, and MMLU accuracy drops, demonstrating that safety gains come with minimal utility cost (<0.35 pp MMLU drop)
  • Geometric transfer failure: Arabic's consistent failure across all architectures and steering magnitudes suggests the method's effectiveness depends on latent-space geometric alignment between source and target languages, not merely data availability

Industry Insight

  • Training-free mechanistic interventions like LSR-Anchoring could become a standard layer in multilingual deployment pipelines, allowing organizations to patch safety gaps in low-resource languages without the compute and data costs of retraining
  • The Arabic failure case serves as a cautionary signal: geometric mismatch may affect other language pairs beyond African languages, suggesting practitioners should validate steering transfer across language families before relying on cross-lingual safety transfer
  • The integration of SAE features for steering (SDS) demonstrates that interpretability tools can directly enable safer deployment; investing in SAE research for production models may yield dual benefits of both mechanistic understanding and actionable safety interventions

TL;DR

  • 提出LSR-Anchoring(潜在空间拒绝锚定)方法,无需训练即可恢复低资源非洲语言模型的安全拒绝能力
  • 核心机制:从英语提示中提取拒绝方向并钳制到残差流,解决约鲁巴语、伊博语、伊加拉语、豪萨语的安全对齐缺失问题
  • 两种变体:MAS(均值激活引导)在Mistral/Qwen上良性降级<0.08,SDS(SAE衍生引导)通过稀疏自编码器特征将KL散度降低3.5-7倍
  • 阿拉伯语在所有架构上均失败,表明存在几何失配而非基线效应
  • MMLU准确率下降控制在0.35个百分点以内,验证方法对通用能力的保护

为什么值得看

本文首次系统性地揭示了多语言模型在低资源语言上的安全对齐漏洞,并提出了无需重新训练的机制性修复方案。对于AI安全研究者和多语言模型开发者而言,该工作提供了可操作的推理时干预技术,避免了昂贵的数据收集和微调成本。

技术解析

  • 问题定义:指令微调模型在英语中拒绝有害请求,但在约鲁巴语、伊博语、伊加拉语、豪萨语等低资源语言中遵从相同请求,表明拒绝机制存在于残差流但未激活
  • LSR-Anchoring架构:训练-free方法,从英语提示提取拒绝方向向量,在推理时钳制到残差流,无需目标语言标注数据
  • MAS变体:Mean-Activation Steering,在Llama-3-8B/3.1-70B、Mistral-7B-Instruct、Qwen2.5-7B四架构上测试,Mistral/Qwen安全恢复且良性降级<0.08,Llama-3-8B出现过校正(DPL=1.00)
  • SDS改进:使用单个SAE特征替代密集均值差方向,KL散度降低3.5-7倍,避免良性崩溃
  • 语言转移差异:四种非洲语言正向转移,阿拉伯语在所有架构和转向幅度上失败,归因于几何失配

行业启示

  • 多语言AI安全对齐存在显著的语言资源鸿沟,低资源语言模型的安全风险需要机制性而非数据驱动解决方案
  • 推理时干预(inference-time intervention)为安全修复提供了低成本路径,可避免重新训练的经济和时间成本
  • 语言间的几何结构差异(如阿拉伯语失败)提示安全方向迁移存在语言特异性,需建立更精细的语言分类和安全适配策略

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。

LLM 大模型 Alignment 对齐 Security 安全 Research 科学研究 Training 训练