ACL Findings2026
Weight Tying Biases Token Embeddings Towards the Output Space
Antonio Lopardo, Avyukth Harish, Catherine Arnett, Akshat Gupta
Abstract
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, we 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, we 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, compromising its role in input representation. These results help explain why weight tying can harm performance at scale and have implications for training smaller LLMs, where the embedding matrix contributes substantially to total parameter count.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
Builds on10
- WinoGrande: An Adversarial Winograd Schema Challenge at ScaleKeisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, Yejin ChoiAAAI 2020 · 3,037 citations
- PIQA: Reasoning about Physical Commonsense in Natural LanguageYonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao et al.AAAI 2020 · 2,916 citations
- Pythia: A Suite for Analyzing Large Language Models Across Training and ScalingStella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley et al.ICML 2023 · 1,822 citations
- XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual GeneralisationJunjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig et al.ICML 2020 · 1,132 citations
- Emerging Cross-lingual Structure in Pretrained Language ModelsAlexis Conneau, Shijie Wu, Haoran Li, Luke Zettlemoyer et al.ACL 2020 · 210 citations
Related papers
- Revisiting Representation Degeneration Problem in Language ModelingZhong Zhang, Chongming Gao, Cong Xu, Rui Miao et al.EMNLP Findings 2020 · 19 citations
- Rethinking Embedding Coupling in Pre-trained Language ModelsHyung Won Chung, Thibault Févry, Henry Tsai, Melvin Johnson et al.ICLR 2021 · 11 citations
- By Tying Embeddings You Are Assuming the Distributional HypothesisFrancesco Bertolotti, Walter CazzolaICML 2024 · 4 citations
- Disentangling Geometry, Performance, and Training in Language ModelsAtharva Kulkarni, Jacob Mitchell Springer, Arjun Subramonian, Swabha SwayamdiptaICML 2026 · 1 citation
- SLlama: Parameter-Efficient Language Model Architecture for Enhanced Linguistic Competence Under Strict Data ConstraintsVictor Adelakun Omolaoye, Babajide Alamu Owoyele, Gerard de MeloEMNLP 2025
