ASIDE accepted to #ICLR2026! ๐ง๐ท๐
We architecturally separate instructions and data in LLMs by rotating data token embeddings 90ยฐ during the forward pass: one extra matmul, virtually no overhead.
Models & code open-sourced โฌ๏ธ
Flying to #ICLR2026 ๐ง๐ท to present our paper, ASIDE: a parameter-free 90ยฐ rotation of data embeddings gives LLMs built-in instruction-data separation, cutting prompt injection rates without explicit safety training.
๐Thu Apr 23, 10:30, Pavilion 4, #3910
โฌ๏ธ Paper, Code, Models
๐ข ๐๐ฎ๐น๐น ๐ณ๐ผ๐ฟ ๐ฃ๐ผ๐๐๐ฒ๐ฟ๐: ๐๐๐ ๐ฆ๐ฎ๐ณ๐ฒ๐๐ ๐ฎ๐ป๐ฑ ๐ฆ๐ฒ๐ฐ๐๐ฟ๐ถ๐๐ ๐ช๐ผ๐ฟ๐ธ๐๐ต๐ผ๐ฝ @ ๐๐๐๐๐ฆ ๐จ๐ป๐๐ผ๐ป๐ณ๐ฒ๐ฟ๐ฒ๐ป๐ฐ๐ฒ
๐ December 2, 2025
๐ Copenhagen
An opportunity to discuss your work with colleagues working on similar problems in LLM safety and security
Results: much higher instruction-data separation, stronger prompt injection robustness, no utility loss. Also, near-perfect linear separability on instructions vs data at every layer of the model.