Problem
Research questions and friction points this paper is trying to address.
Cross-lingual alignment
Decoder-only models
Mixture-of-Experts (MoE)
Contrastive learning
Multilingual tokenization
Cross-Lingual Alignment for Decoder-Only Models using MoE Routers
🤖 AI Summary
This study addresses the challenge of cross-lingual representation alignment in decoder-only large language models, which arises from tokenization discrepancies across languages. To tackle this issue, this work proposes a novel contrastive learning paradigm that leverages Mixture-of-Experts (MoE) router outputs as alignment anchors. Departing from conventional auxiliary losses applied to hidden states, the method constructs robust sequence-level alignment objectives through pooling and optimizes them via controlled continual pre-training. Experimental results demonstrate that the proposed approach effectively aligns lower-layer hidden representations, substantially enhancing the multilingual performance of open-source MoE models. Overall, this research offers a promising new direction for achieving cross-lingual alignment in large language models.
📝 Abstract
Cross-lingual contrastive learning has been a core component of multilingual encoder training, but the ability to explicitly align representations is not possible in decoder-only LLMs because of varying multilingual tokenization. However, a growing amount of research suggests that even in LLMs, higher cross-lingual representational alignment leads to improved cross-lingual transfer. In this paper, we propose a novel approach to reimagine cross-lingual contrastive learning given the architectural constraints of modern LLMs. Rather than applying an auxiliary alignment loss on hidden states, we propose using the outputs of the mixture-of-experts (MoE) routers as the target for alignment. Router outputs lend themselves better to pooling over many tokens, enabling more reliable cross-lingual comparisons at the sequence-level. Controlled continual pre-training experiments on four open-source MoEs show that incorporating this routing loss also aligns the underlying hidden representations across languages. Most importantly, this loss improves multilingual performance on our diverse evaluation suite, demonstrating the potential of cross-lingual MoE router alignment.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Cross-Lingual Alignment
Decoder-Only Models
Mixture-of-Experts (MoE)
Contrastive Learning
Router Outputs
🔎 Similar Papers
No similar papers found.
