New AI safety method fixes low-resource language failures
Training-free technique recovers refusal safety in Yoruba, Hausa without retraining costs...
Godwin Abuh Faruna introduces Latent Space Refusal Anchoring (LSR-Anchoring), a training-free method that extracts refusal directions from English prompts and clamps them onto the residual stream at inference time. Tested on Llama-3-8B, Llama-3.1-70B, Mistral-7B-Instruct, and Qwen2.5-7B, it recovers safety on Mistral and Qwen with benign degradation below 0.08. On Llama-3-8B it overcorrects, causing degraded performance on legitimate prompts, which the SAE-derived variant fixes by reducing KL divergence 3.5–7x. Arabic fails across all architectures, while MMLU accuracy drops stay below 0.35 percentage points.
- LSR-Anchoring recovers AI refusal safety in low-resource African languages without retraining or labeled data
- Tested on Llama-3-8B, Llama-3.1-70B, Mistral-7B-Instruct, and Qwen2.5-7B with near-perfect safety recovery (DPL < 0.08 on Mistral/Qwen)
- Sparse Autoencoder-based steering reduces KL divergence by 3.5-7x while keeping accuracy drops <0.35% on MMLU
Why It Matters
This method could democratize AI safety for 3k+ low-resource languages, reducing harmful compliance risks without costly retraining.