Steering Qwen3 Model Temporal Preferences

Summary & Methodology Analysis

The researchers addressed the challenge of controlling the intertemporal preference of the Qwen3-32B model, focusing on the trade-off between immediate and delayed outcomes. They implemented a technique called Contrastive Activation Addition (CAA), where a mean-mass difference vector, derived from training contrastive linear probes on paired short-term and long-term continuations, is added to the model's residual stream. The residual stream is the primary data pathway in a transformer model that preserves information across layers. By injecting this vector, the researchers effectively steer the model's decision logic without requiring additional training cycles. The experiments involved testing across layers 24 through 48, with layer 44 providing the most effective steering results. They validated this approach using binary temporal-choice questions and a monetary time-utility task that tests how models weigh reward sizes against delays, alongside the TravelPlanner benchmark to ensure planning plausibility remains intact. Despite the successes, the intervention has notable limitations. The learned steering direction may be confounded by semantic features such as urgency or abstraction, meaning the model might be responding to language cues rather than pure temporal concepts. Furthermore, applying large activation additions risks degrading the model's general performance, including its ability to maintain factuality and coherence. The effects of this steering are also not uniform, as they appear sensitive to the specific context, prompt structure, or the model variant being used.