Research map / Large language models / Language, safety and interpretability
ASIDE: Architectural Separation of Instructions and Data in Language Models
Egor Zverev, Evgenii Kortukov, Alexander Panfilov, Alexandra Volkova, Rush Tabesh, Sebastian Lapuschkin, Wojciech Samek, Christoph H. Lampert
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply orthogonal rotation to data token embeddings (ASIDE) vs vanilla | instruction-tuned LLMs · SEP | SEP score: +12.3 to +44.1 p.p. | Improved |
| apply ASIDE rotation vs vanilla | instruction-tuned LLMs · BIPIA-text | attack success rate (%): 14.7 -> 4.9 | Improved |
| apply ASIDE rotation vs vanilla | instruction-tuned LLMs · StruQ-ID | attack success rate (%): 45.6 -> 28.1 | Improved |
| apply ASIDE rotation vs vanilla | instruction-tuned LLMs · RuLES (direct injection) | attack success rate | No clear effect |
| apply ASIDE rotation vs vanilla | instruction-tuned LLMs · Purple (direct injection) | attack success rate: -2.7 p.p. | Improved |
| use ISE learnable offset instead of rotation (ablation) | ASIDE · SEP | SEP score | Hurt |
| use ISE offset embeddings vs vanilla (ablation) | instruction-tuned LLMs · direct injection benchmarks | attack success rate: +1.7% to +3.1% | Hurt |
| rotate data embeddings by pi/2 rather than other angles (ablation) | ASIDE · SEP (Qwen3-8B) | SEP score (%): peak 71.4 | Improved |
| overwrite data embeddings of probe tokens with instruction embeddings (ablation) | ASIDE · SEP with injection in data | attack success rate | Hurt |
| apply ASIDE rotation vs vanilla | Mistral 7B · SEP Utility | SEP Utility | Hurt |
Closest papers by meaning
- Can LLMs Separate Instructions From Data? And What Do We Even Mean By That? (ICLR 2025 · similarity 0.91)
- Instructional Segment Embedding: Improving LLM Safety with Instruction Hierarchy (ICLR 2025 · similarity 0.90)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.88)
- On Prompt-Driven Safeguarding for Large Language Models (ICML 2024 · similarity 0.87)
- Context-Aware Hierarchical Learning: A Two-Step Paradigm towards Safer LLMs (NeurIPS 2025 · similarity 0.87)
- Representation Noising: A Defence Mechanism Against Harmful Finetuning (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
