Ian Rios-Sialer's research causally localizes a subgraph for temporal preference in a distilled LLM, identifying mid-to-upper-layer nodes through gradient-based attribution and activation patching. The study finds that unintervened LLMs discount the future several times less steeply than humans, yet this preference is unstable across contexts. The work presents suggestive evidence that steering vectors can shift temporal preference, demonstrating how mechanistic interpretability can enable control over how LLMs plan and reason.
No score is assigned. Sources and their independence are shown in the citation chain below.