Qwen3-32B Temporal Preference Steering via Contrastive Activation Addition

A novel technique has been created by researchers to influence the temporal preferences of large language models, particularly Qwen3-32B, through contrastive activation addition. The findings, published as arXiv preprint 2608.03892, reveal linear representations of temporal horizons within the model's residual stream. By training contrastive linear probes on teacher-forced temporal-choice responses, a distinction between short-term and long-term preferences was discovered. When applying contrastive activation-addition steering to binary temporal-choice tasks, out-of-distribution monetary intertemporal choices, and the TravelPlanner benchmark, notable bidirectional preference shifts were achieved. The key takeaway is that simple linear probes can identify temporal-horizon directions, enabling adjustments in the model's indifference threshold regarding smaller-sooner versus larger-later rewards, which has significant implications for AI alignment and time-related decision-making.