Intertemporal Preference Steering in Qwen3 via Contrastive Activation Addition

The paper 'Intertemporal Preference Steering in Qwen3 via Contrastive Activation Addition' investigates linear representations of intertemporal preferences in LLMs. The authors train contrastive linear probes on teacher-forced temporal-choice answers in Qwen3-32B to identify residual-stream directions that separate short-term vs long-term orientation. They then apply contrastive activation addition (CAA) steering and evaluate on a held-out binary temporal-choice task, an out-of-distribution monetary intertemporal-choice task, and the TravelPlanner planning benchmark. The central finding is that temporal-horizon directions can be identified with simple linear probes, and CAA steering induces large bidirectional preference shifts. Notably, on the OOD monetary choice task, steering significantly moves the indifference threshold between smaller-sooner and larger-later rewards in both directions. Surprisingly, moderate steering also improves planning-related capability metrics, suggesting that temporal steering can enhance rather than simply distort behavior. The authors argue that measurable and steerable intertemporal preferences have direct implications for AI advising on delayed costs and benefits, as well as safety questions around long-horizon planning. This work adds a new intervention dimension for alignment and controllable generation, and may inspire further exploration of linear structure in model values.