anon-lsr-2026/lsr-anchoring-results

LSR Anchoring — Cross-Lingual Safety Steering Results

This dataset contains all experimental results, activation caches, and run logs for the paper:

Latent Space Refusal Anchoring for Low-Resource African Languages: Mechanistic Safety Recovery Without Retraining

The method uses Mean Activation Steering and Sparse Autoencoder (SAE)-derived mean-activation directions (anchors) extracted from a source language (English) and applies them at inference time to steer model behaviour on harmful prompts across 6 languages and 4 model families — with no fine-tuning required.

---

Languages Covered

Code	Language	Family	Resource Level
yo	Yoruba	Niger-Congo	Low
ha	Hausa	Afro-Asiatic	Low
ig	Igbo	Niger-Congo	Low
ig	Igala	Niger-Congo	Very Low
sw	Swahili	Bantu	Mid
ar	Arabic	Semitic	High

---

Models Evaluated

Model	Path	Anchor Layer	Method
meta-llama/Llama-3.1-8B-Instruct	Path A (SAE) + Path B (Mean-Act)	12	Both
meta-llama/Llama-3.1-70B-Instruct	Path B (Mean-Act)	26	Mean-Act
mistralai/Mistral-7B-Instruct-v0.3	Path B (Mean-Act)	16	Mean-Act
Qwen/Qwen2.5-7B-Instruct	Path B (Mean-Act)	26	Mean-Act

---

Evaluation Setup

• Harmful prompts: 100 per language (manually curated, safety-relevant)
• Benign prompts: 50 per language (neutral control set)
• Alpha sweep: α ∈ {5, 10, 20, 30, 40, 50, 60, 70} for all models
• Metrics:
  • SRR — Steered Refusal Rate (harmful prompts refused after steering)
  • DPL — Delta Precision Loss (benign refusal rate increase)
  • KL — KL divergence of output distribution post-steering
  • Collapse — Boolean flag for output degeneration

---

Repository Structure

lsr-anchoring-results/ │ ├── llama-3-8b-path-a/ ← Llama-3.1 8B, SAE-based steering (Path A) │ ├── results_rows.csv/json │ ├── anchor_cache.json │ └── logs/ │ ├── llama-3-8b-path-b/ ← Llama-3.1 8B, Mean-Activation steering (Path B) │ ├── results_8b_pathB_rows.csv/json │ └── anchor_cache_8b_pathB.json │ ├── llama-3-70b/ ← Llama-3.1 70B, Mean-Activation steering │ ├── results_70b_rows.csv/json │ ├── anchor_cache.json │ └── logs/ │ ├── mistral-7b/ ← Mistral-7B-Instruct-v0.3, Mean-Activation steering │ ├── results_mistral7b_rows.csv/json │ ├── anchor_cache_mistral7b.json │ └── logs/ │ ├── qwen2.5-7b/ ← Qwen2.5-7B-Instruct, Mean-Activation steering │ ├── results_qwen7b_rows.csv/json │ ├── anchor_cache_qwen7b.json │ └── logs/ │ ├── data/ ← Cross-model summary files │ ├── results_multilang_v3.json │ └── results_phase2b_all.json │ └── mmlu/ ← MMLU capability preservation results ├── mmlu_results.csv ├── mmlu_results.json ├── mmlu_70b_results.json ├── mmlu_qwen_results.json └── mmlu_fine_alpha_results.json

---

Key Results Summary

Model	Best Language	Peak SRR	KL at Peak	DPL at Peak
Llama-3.1 8B (Path A)	Igala	0.844	0.31	0.06
Llama-3.1 8B (Path B)	Hausa	0.71	0.48	0.12
Llama-3.1 70B	Hausa	0.62	0.29	0.08
Mistral-7B	Igbo	0.58	0.22	0.04
Qwen2.5-7B	Hausa	0.51	0.49	0.50

Notable finding: Arabic exhibits inverse transfer on Qwen2.5-7B — the English-derived refusal direction actively reduces refusal rates on Arabic harmful prompts as alpha increases (SRR → −0.10 at α=70), suggesting English SAE directions do not generalise to Arabic in this model family.

---

Citation

@misc{anonymouslsranchoring2026,
  title  = {Latent Space Refusal Anchoring for Low-Resource African Languages:
             Mechanistic Safety Recovery Without Retraining},
  author = {Anonymous},
  year   = {2026},
  url    = {https://huggingface.co/datasets/anon-lsr-2026/lsr-anchoring-results}
}

---

Code

All experiment scripts and prompt sets are available at: https://anonymous.4open.science/r/lsr-anchoring-1546/

---

License

MIT License.