A Shared Subcircuit Lets LLMs Count Down Across Tasks

2026-07-13

AI Summary

This study addresses the lack of explicit positional tracking mechanisms in large language models for tasks requiring precise control over output length. Through controlled experiments, the authors discover and validate, for the first time in Llama-3.1-70B-Instruct, a “countdown subcircuit” that estimates the number of remaining generation steps. This subcircuit exhibits remarkable generality, being shared across diverse tasks and even across different models, and maintains a stable representational geometry. It is consistently activated in both explicitly and implicitly length-dependent tasks. The findings uncover a fundamental mechanism underlying behavioral generalization in large language models and offer a novel perspective on the functional modularity within their internal representations.

Abstract

Writing a sentence of exactly twelve words; ending a DNA sequence at the right codon; formatting an ASCII table. These are all tasks that language models can do that requires tracking how many tokens remain before a target. In this work, we identify in Llama-3.1-70B-Instruct a general mechanism for performing these tasks: a "countdown subcircuit" that compares the current position to a goal length and estimates the time remaining until then. We first isolate a countdown subcircuit in a controlled setting, in which the model is tasked with writing a fixed-length sentence ending in a specified word. We then investigate the geometry of the representations used by the subcircuit, and find that the subcircuit uses an identical motif previously identified in a frontier LLM on a separate task, thus suggesting that this motif is shared across models. Finally, we use unsupervised probing on a natural language dataset to find a variety of other tasks where this subcircuit is used, including tasks where the goal length is inferred from context rather than explicitly stated. Our work suggests that reverse-engineering subcircuits allows us to understand how behaviors generalize from a single example to many different tasks and even models.

Problem

• countdown
• subcircuit
• large language models
• token tracking
• generalization

Innovation

• countdown subcircuit
• mechanistic interpretability
• representation geometry
• cross-task generalization
• unsupervised probing