Arithmetic Transformers Can Length-Generalize in Both Operand Length and Count

Hanseul Cho, Jaeyoung Cha, Srinadh Bhojanapalli, Chulhee Yun
2/12/2026

No official code found. Best implementation: HanseulJo/position-coupling (14 stars, MIT, last commit 2025-10-26).

Semantic Scholar

Code Implementations(1)

HanseulJo/position-couplingCommunity55%
MIT

Position Coupling: Improving Length Generalization of Arithmetic Transformers Using Task Structure (NeurIPS 2024) + Arithmetic Transformers Can Length-Generalize in Both Operand Length and Count (ICLR 2025)

141Shell, PythonMay 30, 202411 months agoMIT

Similar Papers

AllSim: Simulating and Benchmarking Resource Allocation Policies in Multi-User Systems

96% match2026

Task Ambiguity in Humans and Language Models

96% match2026

Understanding and Minimising Outlier Features in Transformer Training

95% match2026

On Linear Convergence in Smooth Convex-Concave Bilinearly-Coupled Saddle-Point Optimization: Lower Bounds and Optimal Algorithms

95% match2026

Finding Wasserstein Ball Center: Efficient Algorithm and The Applications in Fairness

95% match2026

Do LLM Agents Have Regret? A Case Study in Online Learning and Games

94% match2026

Cite this paper

@article{cho2026arithmetic,
  title  = {Arithmetic Transformers Can Length-Generalize in Both Operand Length and Count},
  author = {Hanseul Cho and Jaeyoung Cha and Srinadh Bhojanapalli and Chulhee Yun},
  year   = {2026},
  url    = {https://api.semanticscholar.org/CorpusID:52930b240e5f057f2e6377cd839c7796c39bec0e},
  journal = {ICLR 2025 2025}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/52930b240e5f057f2e6377cd839c7796c39bec0e.svg)](https://codeofpaper.com/paper/52930b240e5f057f2e6377cd839c7796c39bec0e)

Discussion