Secure Linear Alignment of Large Language Models

2026-03-19

AI Summary

This work addresses the critical challenge of enabling collaborative inference among independently trained large language models under strict security constraints where neither data nor models can be shared. The authors propose the first efficient secure inference framework that relies solely on encrypted linear operations. By leveraging the convergence properties of large language model representation spaces, they learn an affine transformation on public data to align model representations across participants. Client queries are protected using homomorphic encryption, ensuring strong security guarantees. The method demonstrates for the first time that linear alignment alone suffices to support cross-model text generation, achieving near-lossless performance on embedding classification and out-of-distribution detection tasks with sub-second latency.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Large language models for search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Language models increasingly appear to learn similar representations, despite differences in training objectives, architectures, and data modalities. This emerging compatibility between independently trained models introduces new opportunities for cross-model alignment to downstream objectives. Moreover, it unlocks new potential application domains, such as settings where security, privacy, or competitive constraints prohibit direct data or model sharing. In this work, we propose a privacy-preserving framework that exploits representational convergence to enable cross-silo inference between independent language models. The framework learns an affine transformation over a shared public dataset and applies homomorphic encryption to protect client queries during inference. By encrypting only the linear alignment and classification operations, the method achieves sub-second inference latency while maintaining strong security guarantees. We support this framework with an empirical investigation into representational convergence, in which we learn linear transformations between the final hidden states of independent models. We evaluate these cross-model mappings on embedding classification and out-of-distribution detection, observing minimal performance degradation across model pairs. Additionally, we show for the first time that linear alignment sometimes enables text generation across independently trained models.

Problem

• secure alignment
• language models
• privacy-preserving
• cross-model inference
• representational convergence

Innovation

• linear alignment
• homomorphic encryption
• representational convergence
• privacy-preserving inference
• cross-model transfer