---
title: "Intertemporal Preference Steering in Qwen3 via Contrastive Activation Addition"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.03892v1"
permalink: "https://feed.astralor.com/items/itm_70fe4c6acd87ebc68fe445b2"
published: "2026-08-04T16:25:38.000Z"
score: 77
channel: "ai"
tags: ["AI Safety","Qwen3","Contrastive Activation Addition","Linear Probing","Intertemporal choice"]
---

# Intertemporal Preference Steering in Qwen3 via Contrastive Activation Addition

## AI commentary

> What's interesting here is that it turns 'temporal preference'—a psychological concept—into a manipulable vector inside the model. For alignment researchers, this offers a lighter-weight alternative to RLHF for preference adjustment; for product folks, it could someday enable dynamic adjustment of an AI advisor's horizon based on user traits, like making a finance assistant more patient or more decisive. I particularly like the surprising finding that moderate temporal steering improves planning capability, suggesting that some preference directions are coupled with capability—worth exploring further.

## At a glance

This study explores linear representations of temporal horizon in the large language model Qwen3-32B, identifying short-term vs long-term directions in the residual stream using contrastive linear probes. Contrastive activation addition (CAA) steering is evaluated on a held-out binary temporal-choice task, an out-of-distribution monetary intertemporal-choice task, and the TravelPlanner benchmark. Results show large bidirectional shifts in the model's indifference threshold between smaller-sooner and larger-later rewards, and moderate steering improves planning-related capability metrics. The findings indicate that model intertemporal preferences are measurable and steerable, with implications for AI systems giving advice on delayed costs and benefits, as well as safety concerns around long-horizon planning.

## Key points

- Contrastive linear probes identify temporal-horizon directions in Qwen3-32B's residual stream, enabling bidirectional preference manipulation.
- Steering strongly shifts the indifference threshold between smaller-sooner and larger-later rewards on an out-of-distribution monetary intertemporal-choice task.
- Moderate temporal steering improves planning-related capability metrics on the TravelPlanner benchmark.
- The work shows that LLM intertemporal preferences are linearly representable and steerable, offering new tools for alignment and safety research.

## Full text

We study linear representations of temporal horizon in the large language model Qwen3-32B and use them to change the model's time-related preferences, recommendations, and capabilities. We train contrastive linear probes on teacher-forced temporal-choice answers to find a short-term versus long-term direction in the model's residual stream, and evaluate contrastive activation-addition steering on a held-out binary temporal-choice task, an out-of-distribution monetary intertemporal-choice task, and a TravelPlanner capability benchmark. The central result is that temporal-horizon directions can be identified with simple contrastive linear probes and then used for steering to induce large, bidirectional preference changes. On an out-of-distribution monetary choice task that varies reward size and delay, steering strongly shifts the model's indifference threshold between smaller-sooner and larger-later rewards in both directions. We further show improvements on a planning-related capability metric under moderate temporal steering. These results suggest that model intertemporal preferences are measurable and steerable, which is relevant for AI systems that give advice involving delayed costs and benefits, and for safety questions about long-horizon planning.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_70fe4c6acd87ebc68fe445b2