Not all language model features are linear

J Engels, EJ Michaud, I Liao, W Gurnee… - arxiv preprint arxiv …, 2024 - arxiv.org
Recent work has proposed that language models perform computation by manipulating one-
dimensional representations of concepts (" features") in activation space. In contrast, we …

Dichotomy of early and late phase implicit biases can provably induce grokking

K Lyu, J **, Z Li, SS Du, JD Lee, W Hu - arxiv preprint arxiv:2311.18817, 2023 - arxiv.org
Recent work by Power et al.(2022) highlighted a surprising" grokking" phenomenon in
learning arithmetic tasks: a neural net first" memorizes" the training set, resulting in perfect …

Fourier circuits in neural networks: Unlocking the potential of large language models in mathematical reasoning and modular arithmetic

J Gu, C Li, Y Liang, Z Shi, Z Song… - arxiv preprint arxiv …, 2024 - openreview.net
In the evolving landscape of machine learning, a pivotal challenge lies in deciphering the
internal representations harnessed by neural networks and Transformers. Building on recent …

Emergence in non-neural models: grokking modular arithmetic via average gradient outer product

N Mallinar, D Beaglehole, L Zhu… - arxiv preprint arxiv …, 2024 - arxiv.org
Neural networks trained to solve modular arithmetic tasks exhibit grokking, a phenomenon
where the test accuracy starts improving long after the model achieves 100% training …

Why do you grok? a theoretical analysis of grokking modular addition

MA Mohamadi, Z Li, L Wu, DJ Sutherland - arxiv preprint arxiv …, 2024 - arxiv.org
We present a theoretical explanation of the``grokking''phenomenon, where a model
generalizes long after overfitting, for the originally-studied problem of modular addition. First …

Do Mice Grok? Glimpses of Hidden Progress During Overtraining in Sensory Cortex

T Kumar, B Bordelon, C Pehlevan, VN Murthy… - arxiv preprint arxiv …, 2024 - arxiv.org
Does learning of task-relevant representations stop when behavior stops changing?
Motivated by recent theoretical advances in machine learning and the intuitive observation …

Progressive distillation induces an implicit curriculum

A Panigrahi, B Liu, S Malladi, A Risteski… - arxiv preprint arxiv …, 2024 - arxiv.org
Knowledge distillation leverages a teacher model to improve the training of a student model.
A persistent challenge is that a better teacher does not always yield a better student, to …

Pre-trained Large Language Models Use Fourier Features to Compute Addition

T Zhou, D Fu, V Sharan, R Jia - arxiv preprint arxiv:2406.03445, 2024 - arxiv.org
Pre-trained large language models (LLMs) exhibit impressive mathematical reasoning
capabilities, yet how they compute basic arithmetic, such as addition, remains unclear. This …

Composing Global Optimizers to Reasoning Tasks via Algebraic Objects in Neural Nets

Y Tian - arxiv preprint arxiv:2410.01779, 2024 - arxiv.org
We prove rich algebraic structures of the solution space for 2-layer neural networks with
quadratic activation and $ L_2 $ loss, trained on reasoning tasks in Abelian group (eg …

Unifying and Verifying Mechanistic Interpretations: A Case Study with Group Operations

W Wu, L Jaburi, J Drori, J Gross - arxiv preprint arxiv:2410.07476, 2024 - arxiv.org
A recent line of work in mechanistic interpretability has focused on reverse-engineering the
computation performed by neural networks trained on the binary operation of finite groups …