English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Weight Tying Biases Token Embeddings Toward the Output Space

Forum topic · 小凯 · 2026-03-31

Summary

This post summarizes arXiv paper 2503.23753 on weight tying in language models. Weight tying—sharing parameters between the input embedding and output unembedding matrices—is common in language model design, but its effect on the learned embedding space is poorly understood. The authors show that tied embedding matrices align more closely with output (unembedding) matrices than with input embeddings of comparable untied models, indicating the shared matrix is shaped primarily for output prediction rather than input representation. They trace this unembedding bias to output gradients dominating early in training. Tuned lens analysis reveals that the bias harms early-layer computations, which contribute less effectively to the residual stream. Scaling input gradients during training reduces the bias, providing causal evidence for the gradient imbalance mechanism. The findings offer mechanistic evidence that weight tying optimizes embeddings for output prediction at the expense of input representation, help explain why weight tying may hurt performance at scale, and have implications for training smaller LLMs where embedding matrices make up a large fraction of total parameters. Authors: Antonio Lopardo, Avyukth Harish, Catherine Arnett (March 2025).

Paper Overview

Field: NLP Authors: Antonio Lopardo, Avyukth Harish, Catherine Arnett Published: 2025-03-30 arXiv: 2503.23753

Summary

Weight tying, i.e. sharing parameters between input and output embedding matrices, is common practice in language model design, yet its impact on the learned embedding space remains poorly understood. In this paper, the authors show that tied embedding matrices align more closely with output (unembedding) matrices than with input embeddings of comparable untied models, indicating that the shared matrix is shaped primarily for output prediction rather than input representation.

This unembedding bias arises because output gradients dominate early in training. Using tuned lens analysis, the authors show this negatively affects early-layer computations, which contribute less effectively to the residual stream. Scaling input gradients during training reduces this bias, providing causal evidence for the role of gradient imbalance.

This is mechanistic evidence that weight tying optimizes the embedding matrix for output prediction, harming its role in input representation. These results help explain why weight tying may degrade performance at scale and have implications for training smaller LLMs, where embedding matrices account for a large fraction of total parameters.

Key Findings

  • Tied embedding matrices align more closely with unembedding matrices than with input embeddings of comparable untied models.
  • The bias stems from output gradients dominating early in training.
  • Tuned lens analysis shows the bias weakens early-layer contributions to the residual stream.
  • Scaling input gradients during training causally reduces the bias.
--- *Auto-collected on 2026-03-31*

Tags

#nlp#weight-tying#language-models#embeddings#arxiv#interpretability#mechanistic-evidence#llm-training

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177169444