Researchers propose MD-SigLIP, a margin-regularized framework to align brain embeddings with text embeddings for improved brain-language decoding

Researchers propose MD-SigLIP, a margin-regularized structured semantic alignment framework that directly aligns brain embeddings with text embeddings in a shared semantic space to enable retrieval-based decoding. The method addresses ambiguity in brain-language decoding about whether decoded content reflects genuine neural representations or reconstructions by the language model.

The framework builds upon duplicate-aware sigmoid contrastive learning and introduces a listwise margin-regularized term that enforces structured ranking constraints between positive semantic clusters and negative samples. This formulation aims to explicitly model the correspondence between neural representations and language semantics.

By modeling multi-positive semantic structure and margin-based ordering simultaneously, the method seeks to capture the manifold organization of language embeddings reflected in neural signals. The authors report that experiments demonstrate state-of-the-art retrieval performance under both full-vocabulary and subset evaluation settings.

The work is presented as a pre-print on arXiv under the category Computation and Language (cs.CL) and Machine Learning (cs.LG), with authors Jiaqi Wang, Huawen Hu, and Shu Zhang. The submission history indicates the paper was posted on August 17, 2026.