Temporal Preference Concepts and their Functions in a Large Language Model

In three linesMechanistic interpretability study on Qwen3-4B-Instruct-2507: causal localization of temporal preference circuits via gradient attribution and activation patching. LLMs encode time horizon in residual stream and discount future less steeply than humans, but preference is context-unstable. Steering vectors show potential for explicit control.

Summary generated by Claude — human-verified