Research map / Large language models / Language, safety and interpretability

An Analysis for Reasoning Bias of Language Models with Small Initialization

ICML 2025 · Large language models › Language, safety and interpretability · approach: Reasoning mechanisms and failure modes · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that small parameter initialization in Transformers/LLMs biases learning toward reasoning over memorization, and explains this via embedding-space label-distribution dynamics and attention averaging.

Problem

Understanding why small initialization scales in Transformers and LLMs make models favor reasoning tasks over memorization tasks, and what training dynamics cause this task preference.

Approach

Small initialization makes token embeddings of reasoning anchors differentiate early because their label distributions are structured, while memory-anchor labels are uniform random and yield indistinguishable embeddings; a theory based on gradient flow of embeddings and average-like first attention explains the bias.

Key results

• On the synthetic task with gamma=0.8, losses on reasoning train and test sets fall much faster than memory loss, while gamma=0.3 mainly memorizes the training data.
• On GPT-2 trained on PrOntoQA and TinyStories, the average embedding cosine similarity is 0.123 for PrOntoQA tokens versus 0.531 for TinyStories tokens.
• The reasoning-bias metric Delta L between TinyStories and PrOntoQA losses increases as the initialization rate gamma increases, indicating a growing reasoning bias with smaller initialization.

Measured effects reported in the paper

Surprising findings

• Memory tasks whose labels share structure with anchors (label range depending on anchors) are learned faster than uniform-random memory tasks, showing label distribution alone drives learning speed.

Limitations

• Theory relies on infinite-width/limit approximations and small-initialization assumptions (n, gamma to infinity).
• Analysis of real language tasks is limited to GPT-2 on two datasets with small-scale training (10,000 sequences).

Experimental setup

Datasets: PrOntoQA TinyStories synthetic composition anchor-function dataset

Benchmarks: synthetic composition task (reasoning vs memory anchors) PrOntoQA TinyStories

Metrics: training loss prediction accuracy cosine similarity of embeddings relative loss gap Delta L

Baselines: GPT-2 with larger initialization (gamma=0.3) gamma=0.5 default initialization

Builds on: anchor functions neural condensation LeCun initialization He initialization

Closest papers by meaning

• On the Inductive Bias of Stacking Towards Improving Reasoning (NeurIPS 2024 · similarity 0.90)
• Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.90)
• Calibrating Reasoning in Language Models with Internal Consistency (NeurIPS 2024 · similarity 0.89)
• What Do Learning Dynamics Reveal About Generalization in LLM Mathematical Reasoning? (ICML 2025 · similarity 0.89)
• Generalization v.s. Memorization: Tracing Language Models’ Capabilities Back to Pretraining Data (ICLR 2025 · similarity 0.89)
• A Simple Model of Inference Scaling Laws (ICML 2025 · similarity 0.89)