Small Initialization Matters for Large Language Models

• URL: http://arxiv.org/abs/2606.17945v2
• Date: Wed, 24 Jun 2026 02:25:05 GMT
• Title: Small Initialization Matters for Large Language Models
• Abstract summary: We show that parameter initialization is a gene-like determinant of training and, in particular, of model capacity.<n>Reducing the scale consistently improves pretraining, with the largest gains on reasoning-demanding tasks.<n>We identify two widely used empirical settings that restrain the advantage of small initialization, and show how relaxing them restores favorable scaling.
• Score: 28.726627416564842
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models provide a tractable system for asking how intelligence itself emerges, rather than only how LLMs can be engineered. Although progress is usually attributed to scale, data and architecture, we show that parameter initialization is a gene-like determinant of training and, in particular, of model capacity. Reducing the initialization scale consistently improves pretraining, with the largest gains on reasoning-demanding tasks. We identify two widely used empirical settings that restrain the advantage of small initialization, and show how relaxing them restores favorable scaling. We further uncover a critical initialization that balances the reasoning and training. Mechanistically, small initialization drives a distinct developmental trajectory: parameters first condense into low-complexity structures and later expand into richer representations, giving concrete form to the idea that compression is intelligence. Token-level analyses show that the gains concentrate on non-trivial, context-constrained predictions rather than all tokens uniformly. These results motivate a simple $γ$-initialization rule: expose initialization rage as an explicit knob and use small initialization by default, an almost cost-free intervention that improves pretraining and strengthens reasoning across model scales.

Related papers

• Conditioned Initialization for Attention [32.435722001652174]
  Transformers are a dominant architecture in modern machine learning, powering applications across vision, language, and beyond.<n>At the core of their success lies the attention layer, where the query, key, and value matrices determine how token dependencies are captured.<n>While considerable work has focused on scaling and optimizing Transformers, comparatively little attention has been paid to how the weights of the queries, keys and values are.
  arXiv  Detail & Related papers  (2026-09-07T06:17:58Z)
• Logic Before Language: Pre-pretraining on Formal Derivations Fosters Skill Acquisition and Compressibility [41.15091250761469]
  Pre-pretraining language models (LMs) on symbolic data can accelerate and improve natural language acquisition.<n>We show that logic pre-pretraining substantially accelerates skill acquisition in LMs.<n>Mechanistically, formal derivations induce persistent structural reorganization.
  arXiv  Detail & Related papers  (2026-08-04T17:02:34Z)
• A Theory of How Pretraining Shapes Inductive Bias in Fine-Tuning [51.505728136705564]
  We develop an analytical theory of the pretraining-fine-tuning pipeline in diagonal linear networks.<n>We find that different initialization choices place the network into four distinct fine-tuning regimes.<n>A smaller scale in earlier layers enables the network to both reuse and refine its features, leading to superior generalization.
  arXiv  Detail & Related papers  (2026-02-23T17:19:33Z)
• Making Foundation Models Probabilistic via Singular Value Ensembles [56.4174499669573]
  Foundation models have become a dominant paradigm in machine learning, achieving remarkable performance across diverse tasks through large-scale pretraining.<n>Standard approach to quantifying uncertainty, training an ensemble of independent models, incurs prohibitive computational costs that scale linearly with ensemble size.<n>We propose Singular Value Ensemble (SVE), a parameter-efficient implicit ensemble method that builds on a simple, but powerful core assumption.<n>We show that SVE uncertainty quantification achieves comparable to explicit deep ensembles while increasing the parameter count of the base model by less than 1%.
  arXiv  Detail & Related papers  (2026-01-29T18:07:18Z)
• Bigger Isn't Always Memorizing: Early Stopping Overparameterized Diffusion Models [56.032091696552094]
  Generalization in natural data domains is progressively achieved during training before the onset of memorization.<n>Generalization vs. memorization is then best understood as a competition between time scales.<n>We show that this phenomenology is recovered in diffusion models learning a simple probabilistic context-free grammar with random rules.
  arXiv  Detail & Related papers  (2025-05-22T17:40:08Z)
• Do Larger Language Models Generalize Better? A Scaling Law for Implicit Reasoning at Pretraining Time [73.22651918134808]
  This work shows counterintuitive effects of model size scaling and provides new insights into the relationship between scaling and reasoning in language models (LMs)<n>We pretrain LMs from scratch on a synthetic implicit multihop reasoning environment designed to replicate the structure and distribution of real-world large-scale knowledge graphs.<n>We then assess the LMs' ability to complete the missing edges in the graph, which requires multi-hop reasoning that can be viewed as a simplification of implicit reasoning during real-world pretraining.
  arXiv  Detail & Related papers  (2025-04-04T17:57:22Z)
• The First Few Tokens Are All You Need: An Efficient and Effective Unsupervised Prefix Fine-Tuning Method for Reasoning Models [69.798277882245]
  We introduce Unsupervised Prefix Fine-Tuning (UPFT) to enhance large language models' reasoning efficiency.<n>UPFT removes the need for labeled data or exhaustive sampling.<n> Experiments show that UPFT matches the performance of supervised methods.
  arXiv  Detail & Related papers  (2025-03-04T18:56:03Z)
• An Analysis for Reasoning Bias of Language Models with Small Initialization [8.380004565348619]
  Large Language Models (LLMs) have revolutionized Natural Language Processing by demonstrating exceptional performance across diverse tasks.<n>This study investigates the impact of the parameter initialization scale on the training behavior and task preferences of LLMs.
  arXiv  Detail & Related papers  (2025-02-05T15:23:26Z)
• On the Implicit Bias of Initialization Shape: Beyond Infinitesimal Mirror Descent [55.96478231566129]
  We show that relative scales play an important role in determining the learned model. We develop a technique for deriving the inductive bias of gradient-flow.
  arXiv  Detail & Related papers  (2021-02-19T07:10:48Z)
• Provable Benefits of Overparameterization in Model Compression: From Double Descent to Pruning Neural Networks [38.153825455980645]
  Recent empirical evidence indicates that the practice of overization not only benefits training large models, but also assists - perhaps counterintuitively - building lightweight models. This paper sheds light on these empirical findings by theoretically characterizing the high-dimensional toolsets of model pruning. We analytically identify regimes in which, even if the location of the most informative features is known, we are better off fitting a large model and then pruning.
  arXiv  Detail & Related papers  (2020-12-16T05:13:30Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.