Analyzing the Generalization and Reliability of Steering Vectors

• URL: http://arxiv.org/abs/2407.12404v2
• Date: Mon, 22 Jul 2024 08:28:31 GMT
• Title: Analyzing the Generalization and Reliability of Steering Vectors
• Authors: Daniel Tan, David Chanin, Aengus Lynch, Dimitrios Kanoulas, Brooks Paige, Adria Garriga-Alonso, Robert Kirk,
• Abstract summary: We show that steering vectors have substantial limitations both in- and out-of-distribution. In-distribution, steerability is highly variable across different inputs. Out-of-distribution, while steering vectors often generalise well, for several concepts they are brittle to reasonable changes in the prompt.
• Score: 8.253773195379166
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Steering vectors (SVs) are a new approach to efficiently adjust language model behaviour at inference time by intervening on intermediate model activations. They have shown promise in terms of improving both capabilities and model alignment. However, the reliability and generalisation properties of this approach are unknown. In this work, we rigorously investigate these properties, and show that steering vectors have substantial limitations both in- and out-of-distribution. In-distribution, steerability is highly variable across different inputs. Depending on the concept, spurious biases can substantially contribute to how effective steering is for each input, presenting a challenge for the widespread use of steering vectors. Out-of-distribution, while steering vectors often generalise well, for several concepts they are brittle to reasonable changes in the prompt, resulting in them failing to generalise well. Overall, our findings show that while steering can work well in the right circumstances, there remain many technical difficulties of applying steering vectors to guide models' behaviour at scale.

Related papers

• Improving Steering Vectors by Targeting Sparse Autoencoder Features [2.4188584949331053]
  We use SAEs to measure the effects of steering vectors, giving us a method that can be used to understand the causal effect of any steering vector intervention. We develop an improved steering method, SAE-Targeted Steering (SAE-TS), which finds steering vectors to target specific SAE features while minimizing unintended side effects.
  arXiv  Detail & Related papers  (2024-11-04T15:46:20Z)
• Steering Without Side Effects: Improving Post-Deployment Control of Language Models [61.99293520621248]
  Language models (LMs) have been shown to behave unexpectedly post-deployment. We present KL-then-steer (KTS), a technique that decreases the side effects of steering while retaining its benefits. Our best method prevents 44% of jailbreak attacks compared to the original Llama-2-chat-7B model.
  arXiv  Detail & Related papers  (2024-06-21T01:37:39Z)
• Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization [34.05163996072159]
  "steering vectors" are extracted from the activations of human preference data. This work proposes an innovative approach that could produce more effective steering vectors through bi-directional preference optimization. Our method is designed to allow steering vectors to directly influence the generation probability of contrastive human preference data pairs.
  arXiv  Detail & Related papers  (2024-05-28T05:10:40Z)
• Towards Generalizable and Interpretable Motion Prediction: A Deep Variational Bayes Approach [54.429396802848224]
  This paper proposes an interpretable generative model for motion prediction with robust generalizability to out-of-distribution cases. For interpretability, the model achieves the target-driven motion prediction by estimating the spatial distribution of long-term destinations. Experiments on motion prediction datasets validate that the fitted model can be interpretable and generalizable.
  arXiv  Detail & Related papers  (2024-03-10T04:16:04Z)
• Extending Activation Steering to Broad Skills and Multiple Behaviours [5.40770929004319]
  We investigate the efficacy of activation steering for broad skills and multiple behaviours. We find that steering broader skills is competitive to steering narrower skills. We steer models to become more or less myopic and wealth-seeking.
  arXiv  Detail & Related papers  (2024-03-09T02:30:04Z)
• InferAligner: Inference-Time Alignment for Harmlessness through Cross-Model Guidance [56.184255657175335]
  We develop textbfInferAligner, a novel inference-time alignment method that utilizes cross-model guidance for harmlessness alignment. Experimental results show that our method can be very effectively applied to domain-specific models in finance, medicine, and mathematics. It significantly diminishes the Attack Success Rate (ASR) of both harmful instructions and jailbreak attacks, while maintaining almost unchanged performance in downstream tasks.
  arXiv  Detail & Related papers  (2024-01-20T10:41:03Z)
• Steering Llama 2 via Contrastive Activation Addition [41.54815073311959]
  Contrastive Activation Addition (CAA) is a method for steering language models by modifying their activations during forward passes. CAA accurately steers model outputs and sheds light on how high-level concepts are represented in Large Language Models (LLMs)
  arXiv  Detail & Related papers  (2023-12-09T04:40:46Z)
• Unsupervised Domain Adaptation for Self-Driving from Past Traversal Features [69.47588461101925]
  We propose a method to adapt 3D object detectors to new driving environments. Our approach enhances LiDAR-based detection models using spatial quantized historical features. Experiments on real-world datasets demonstrate significant improvements.
  arXiv  Detail & Related papers  (2023-09-21T15:00:31Z)
• On Learning the Tail Quantiles of Driving Behavior Distributions via Quantile Regression and Flows [13.540998552232006]
  We consider the problem of learning models that accurately capture the diversity and tail quantiles of human driver behavior probability distributions. We adapt two flexible quantile learning frameworks for this setting that avoid strong distributional assumptions. We evaluate our approach in a one-step acceleration prediction task, and in multi-step driver simulation rollouts.
  arXiv  Detail & Related papers  (2023-05-22T15:09:04Z)
• Control-Aware Prediction Objectives for Autonomous Driving [78.19515972466063]
  We present control-aware prediction objectives (CAPOs) to evaluate the downstream effect of predictions on control without requiring the planner be differentiable. We propose two types of importance weights that weight the predictive likelihood: one using an attention model between agents, and another based on control variation when exchanging predicted trajectories for ground truth trajectories.
  arXiv  Detail & Related papers  (2022-04-28T07:37:21Z)
• Causally-motivated Shortcut Removal Using Auxiliary Labels [63.686580185674195]
  Key challenge to learning such risk-invariant predictors is shortcut learning. We propose a flexible, causally-motivated approach to address this challenge. We show both theoretically and empirically that this causally-motivated regularization scheme yields robust predictors.
  arXiv  Detail & Related papers  (2021-05-13T16:58:45Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.