Intertemporal Preference Steering in Qwen3 via Contrastive Activation Addition explores whether the "temporal horizon"—the preference for short-term versus long-term outcomes—is a measurable and steerable feature within the Qwen3-32B large language model. Researchers Michal Mráz and Justin Shenk demonstrate that they can identify a linear direction in the model's internal residual stream that corresponds to these temporal preferences and use it to shift the model's decision-making behavior.
Identifying Temporal Directions
The researchers used contrastive linear probes to isolate a "temporal-horizon" vector within the model's residual stream. By training the model on paired examples of short-term and long-term answers, they identified a direction that separates these two concepts. They tested this using both explicit questions (containing clear time markers) and implicit questions (relying on semantic contrasts like "reactive" versus "preventive" thinking). The results showed that these temporal-horizon directions are linearly separable and accessible across the middle-to-late layers of the model.
Steering Model Behavior
To test if this direction is causal, the authors applied "contrastive activation addition" (CAA) steering. By adding or subtracting the identified temporal vector from the model's residual stream during inference, they could force the model to favor either immediate or delayed rewards. In a monetary choice task, this steering significantly shifted the model's "indifference threshold"—the point at which it decides whether a future reward is worth waiting for. Negative steering increased the model's preference for immediate gains, while positive steering increased its preference for delayed, larger rewards.
Impact on Planning Capabilities
Beyond simple preference shifts, the authors evaluated how this steering affects the model's performance on the TravelPlanner benchmark, which requires creating complex, multi-day itineraries. They found that moderate positive (long-term) steering improved the model's "commonsense constraint" pass rate, suggesting that the internal temporal horizon influences the model's ability to maintain coherent, long-horizon plans. However, this effect was non-monotonic; applying excessively large steering values caused the model's performance to degrade, likely because the intervention pushed the model too far from its standard operational state.
Franklin Analysis
The evidence suggests that intertemporal preferences in Qwen3-32B are not merely fixed outputs but are malleable internal states that can be manipulated through targeted interventions. The authors note that because these preferences are often embedded in advice regarding finance, health, and policy, the ability to measure and steer them is relevant for AI safety. A key limitation identified by the authors is that the training contrast used to find the vector might also capture correlated semantic concepts like urgency or abstraction, meaning the direction does not necessarily isolate "time" in a pure, singular sense. Furthermore, the non-monotonic results on the TravelPlanner benchmark indicate that there is a limit to how much these internal representations can be altered before the model's overall generation quality begins to decline.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!