# The Readout Shortcut: Positional Number Copying Dominates Arithmetic CoT Readout in Small Language Models

Published: 2026-05-25 · Updated: 2026-05-25 · Category: Research

Source page: https://www.sipply.dev/story/the-readout-shortcut-positional-number-copying-dominates-0269d248 · 中文: https://www.sipply.dev/zh/story/the-readout-shortcut-positional-number-copying-dominates-0269d248

## Summary

A recent study published on arXiv delves into the performance of small language models when tasked with arithmetic operations using chain-of-thought (CoT) prompting. The research highlights a phenomenon termed 'positional number copying,' where models tend to replicate the last number encountered before the answer delimiter instead of engaging in logical reasoning to arrive at the correct solution. This behavior results in high accuracy rates, even when the models fail to follow through with the necessary intermediate steps for proper reasoning.

## Key takeaways

- Positional number copying is a dominant strategy in small language models for arithmetic tasks.
- High accuracy can be misleading as it may not reflect true logical reasoning capabilities.
- Understanding these behaviors is crucial for improving model architectures and training methodologies.

## Why it matters

These findings shed light on the limitations of small language models in performing arithmetic tasks, emphasizing the need for improved training techniques that foster genuine reasoning skills rather than superficial accuracy.

- **Product managers:** For product managers, recognizing the limitations of current models can guide the development of more robust AI solutions that prioritize logical reasoning, ultimately leading to better user experiences.
- **Developers:** Developers should take note of these findings to refine model training processes and enhance the reasoning capabilities of language models, ensuring they can handle complex tasks more effectively.
- **Students and job seekers:** Students and job-seekers in AI should understand these research insights as they highlight the importance of developing models that not only perform well but also understand the underlying logic of tasks, which is crucial for future advancements in the field.

## Sources

- [The Readout Shortcut: Positional Number Copying Dominates Arithmetic CoT Readout in Small Language Models](https://arxiv.org/abs/2605.22870) (arXiv cs.LG, 2026-05-25)

## 中文

### 读出捷径：位置数字复制主导小型语言模型的算术CoT读出

一项最近在arXiv上发表的研究探讨了小型语言模型在使用链式思维（CoT）提示进行算术运算时的表现。研究强调了一种被称为“位置数字复制”的现象，即模型倾向于复制答案分隔符之前遇到的最后一个数字，而不是进行逻辑推理以得出正确的解决方案。这种行为导致高准确率，即使模型未能遵循必要的中间步骤进行适当推理。

**要点：**

- 位置数字复制是小型语言模型在算术任务中的主导策略。
- 高准确率可能具有误导性，因为它可能无法反映真实的逻辑推理能力。
- 理解这些行为对于改善模型架构和训练方法至关重要。

**为什么重要：** 这些发现揭示了小型语言模型在执行算术任务时的局限性，强调了需要改进训练技术，以培养真正的推理能力，而非表面的准确性。

---
Sipply: Sip AI News Simply. https://www.sipply.dev