Why Can't Transformers Learn Multiplication? Reverse-Engineering Reveals Long-Range Dependency Pitfalls

• URL: http://arxiv.org/abs/2510.00184v1
• Date: Tue, 30 Sep 2025 19:03:26 GMT
• Title: Why Can't Transformers Learn Multiplication? Reverse-Engineering Reveals Long-Range Dependency Pitfalls
• Authors: Xiaoyan Bai, Itamar Pres, Yuntian Deng, Chenhao Tan, Stuart Shieber, Fernanda Viégas, Martin Wattenberg, Andrew Lee,
• Abstract summary: Language models are increasingly capable, yet still fail at a seemingly simple task of multi-digit multiplication.<n>We study why, by reverse-engineering a model that successfully learns multiplication via emphimplicit chain-of-thought'
• Score: 54.57326125204404
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Language models are increasingly capable, yet still fail at a seemingly simple task of multi-digit multiplication. In this work, we study why, by reverse-engineering a model that successfully learns multiplication via \emph{implicit chain-of-thought}, and report three findings: (1) Evidence of long-range structure: Logit attributions and linear probes indicate that the model encodes the necessary long-range dependencies for multi-digit multiplication. (2) Mechanism: the model encodes long-range dependencies using attention to construct a directed acyclic graph to ``cache'' and ``retrieve'' pairwise partial products. (3) Geometry: the model implements partial products in attention heads by forming Minkowski sums between pairs of digits, and digits are represented using a Fourier basis, both of which are intuitive and efficient representations that the standard fine-tuning model lacks. With these insights, we revisit the learning dynamics of standard fine-tuning and find that the model converges to a local optimum that lacks the required long-range dependencies. We further validate this understanding by introducing an auxiliary loss that predicts the ``running sum'' via a linear regression probe, which provides an inductive bias that enables the model to successfully learn multi-digit multiplication. In summary, by reverse-engineering the mechanisms of an implicit chain-of-thought model we uncover a pitfall for learning long-range dependencies in Transformers and provide an example of how the correct inductive bias can address this issue.

Related papers

• Implicit Models: Expressive Power Scales with Test-Time Compute [17.808479563949074]
  Implicit models, an emerging model class, compute outputs by iterating a single parameter block to a fixed point.<n>We study this gap through a nonparametric analysis of expressive power.<n>We prove that for a broad class of implicit models, this process lets the model's expressive power scale with test-time compute.
  arXiv  Detail & Related papers  (2025-10-04T02:49:22Z)
• Beyond Memorization: Extending Reasoning Depth with Recurrence, Memory and Test-Time Compute Scaling [60.63703438729223]
  We show how different architectures and training methods affect model multi-step reasoning capabilities.<n>We confirm that increasing model depth plays a crucial role for sequential computations.
  arXiv  Detail & Related papers  (2025-08-22T18:57:08Z)
• Learning Linear Attention in Polynomial Time [115.68795790532289]
  We provide the first results on learnability of single-layer Transformers with linear attention. We show that linear attention may be viewed as a linear predictor in a suitably defined RKHS. We show how to efficiently identify training datasets for which every empirical riskr is equivalent to the linear Transformer.
  arXiv  Detail & Related papers  (2024-10-14T02:41:01Z)
• Beyond DAGs: A Latent Partial Causal Model for Multimodal Learning [80.44084021062105]
  We propose a novel latent partial causal model for multimodal data, featuring two latent coupled variables, connected by an undirected edge, to represent the transfer of knowledge across modalities.<n>Under specific statistical assumptions, we establish an identifiability result, demonstrating that representations learned by multimodal contrastive learning correspond to the latent coupled variables up to a trivial transformation.<n>Experiments on a pre-trained CLIP model embodies disentangled representations, enabling few-shot learning and improving domain generalization across diverse real-world datasets.
  arXiv  Detail & Related papers  (2024-02-09T07:18:06Z)
• Quantized Fourier and Polynomial Features for more Expressive Tensor Network Models [9.18287948559108]
  We exploit the tensor structure present in the features by constraining the model weights to be an underparametrized tensor network. We show that, for the same number of model parameters, the resulting quantized models have a higher bound on the VC-dimension as opposed to their non-quantized counterparts.
  arXiv  Detail & Related papers  (2023-09-11T13:18:19Z)
• Exposing Attention Glitches with Flip-Flop Language Modeling [55.0688535574859]
  This work identifies and analyzes the phenomenon of attention glitches in large language models. We introduce flip-flop language modeling (FFLM), a family of synthetic benchmarks designed to probe the extrapolative behavior of neural language models. We find that Transformer FFLMs suffer from a long tail of sporadic reasoning errors, some of which we can eliminate using various regularization techniques.
  arXiv  Detail & Related papers  (2023-06-01T17:44:35Z)
• A Unified View of Long-Sequence Models towards Modeling Million-Scale Dependencies [0.0]
  We compare existing solutions to long-sequence modeling in terms of their pure mathematical formulation. We then demonstrate that long context length does yield better performance, albeit application-dependent. Inspired by emerging sparse models of huge capacity, we propose a machine learning system for handling million-scale dependencies.
  arXiv  Detail & Related papers  (2023-02-13T09:47:31Z)
• The Effects of Invertibility on the Representational Complexity of Encoders in Variational Autoencoders [16.27499951949733]
  We show that if the generative map is "strongly invertible" (in a sense we suitably formalize), the inferential model need not be much more complex. Importantly, we do not require the generative model to be layerwise invertible. We provide theoretical support for the empirical wisdom that learning deep generative models is harder when data lies on a low-dimensional manifold.
  arXiv  Detail & Related papers  (2021-07-09T19:53:29Z)
• Goal-directed Generation of Discrete Structures with Conditional Generative Models [85.51463588099556]
  We introduce a novel approach to directly optimize a reinforcement learning objective, maximizing an expected reward. We test our methodology on two tasks: generating molecules with user-defined properties and identifying short python expressions which evaluate to a given target value.
  arXiv  Detail & Related papers  (2020-10-05T20:03:13Z)
• A Solution for Large Scale Nonlinear Regression with High Rank and Degree at Constant Memory Complexity via Latent Tensor Reconstruction [0.0]
  This paper proposes a novel method for learning highly nonlinear, multivariate functions from examples. Our method takes advantage of the property that continuous functions can be approximated by bys, which in turn are representable by tensors. For learning the models, we present an efficient-based algorithm that can be implemented in linear time.
  arXiv  Detail & Related papers  (2020-05-04T14:49:14Z)

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.