Metraipolis › archive › dc_68f5617978cd
Latent Space Refusal Anchoring for Low-Resource African Languages: Mechanistic Safety Recovery Without Retraining — arXiv:2608.18089
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2608.18089
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-18T13:44:45.151Z
- SHA-256 of the archived text
a9ffe470e09eb5c33125a13a7fdb9b08086df6c18d404fa31eabf552bcb91b55- Stable URL of this page
https://metraipolis.com/documento/dc_68f5617978cd
How to cite
Latent Space Refusal Anchoring for Low-Resource African Languages: Mechanistic Safety Recovery Without Retraining — arXiv:2608.18089 — arXiv (en), https://arxiv.org/abs/2608.18089 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_68f5617978cd (Metraipolis archive dc_68f5617978cd, sha256 a9ffe470e09eb5c33125a13a7fdb9b08086df6c18d404fa31eabf552bcb91b55)
BibTeX
@misc{dc68f5617978cd,
title = {Latent Space Refusal Anchoring for Low-Resource African Languages: Mechanistic Safety Recovery Without Retraining — arXiv:2608.18089},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2608.18089},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_68f5617978cd; sha256 a9ffe470e09eb5c33125a13a7fdb9b08086df6c18d404fa31eabf552bcb91b55},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.18089. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Instruction-tuned models often refuse harmful requests in English but comply with the same requests in Yoruba, Igbo, Igala, and Hausa. This suggests that the refusal mechanism is present in the residual stream but fails to activate for low-resource inputs. Recovering it normally requires labelled target-language data and retraining, neither of which is available at scale for most African languages. We introduce Latent Space Refusal Anchoring (LSR-Anchoring), a training-free method that extracts the refusal direction from English prompts and clamps it onto the residual stream at inference time. The primary variant, Mean-Activation Steering (MAS), operates across the four architectures we tested: Llama-3-8B, Llama-3.1-70B, Mistral-7B-Instruct, and Qwen2.5-7B. On Mistral and Qwen it recovers safety with benign degradation below 0.08. On Llama-3-8B it overcorrects, with Degraded Performance on Legitimate prompts (DPL) reaching 1.00. We address this with SAE-Derived Steering (SDS), which replaces the dense mean-difference direction with a single Sparse Autoencoder (SAE) feature and reduces Kullback-Leibler (KL) divergence by 3.5-7x without benign collapse. Four languages transfer positively, but Arabic fails on every architecture and at every steering magnitude, indicating a geometric mismatch rather than a baseline effect. Massive Multitask Language Understanding (MMLU) accuracy drops remain below 0.35 percentage points at every effective steering magnitude. Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-20 in cs.AI). Categorie arXiv: Computation and Language (cs.CL) ; Artificial Intelligence (cs.AI). Pagina del lavoro: https://arxiv.org/abs/2608.18089 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-18T13:41:17.896Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-08-20?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).