Papers with private token sampling
Hyperion: Private Token Sampling with Homomorphic Encryption (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) enable an extraordinary range of applications, yet their adoption is limited by a fundamental concern: users must send sensitive data to remote model providers. |
| Approach: | They propose an efficient homomorphic encryption algorithm for inverse transform sampling that allows private token sampling under HE. |
| Outcome: | The proposed method samples tokens in 0.14 seconds on GPU, achieving a 100 latency improvement over prior work. |