Refusal in LLMs is an Affine Function
- URL: http://arxiv.org/abs/2411.09003v2
- Date: Tue, 19 Nov 2024 04:53:47 GMT
- Title: Refusal in LLMs is an Affine Function
- Authors: Thomas Marshall, Adam Scherlis, Nora Belrose,
- Abstract summary: We propose affine concept editing (ACE) as an approach for steering language models' behavior.
ACE combines affine subspace projection and activation addition to reliably control the model's refusal responses.
Our experiments demonstrate that ACE consistently achieves more precise control over model behavior than existing methods.
- Score: 1.722461331472526
- License:
- Abstract: We propose affine concept editing (ACE) as an approach for steering language models' behavior by intervening directly in activations. We begin with an affine decomposition of model activation vectors and show that prior methods for steering model behavior correspond to subsets of terms of this decomposition. We then provide a derivation of ACE and use it to control refusal behavior on ten different models, including Llama 3 70B. ACE combines affine subspace projection and activation addition to reliably control the model's refusal responses across prompt types. We evaluate the results using LLM-based scoring on a collection of harmful and harmless prompts. Our experiments demonstrate that ACE consistently achieves more precise control over model behavior than existing methods and generalizes to models where directional ablation via affine subspace projection alone produces incoherent outputs. Code for reproducing our results is available at https://github.com/EleutherAI/steering-llama3 .
Related papers
- MASALA: Model-Agnostic Surrogate Explanations by Locality Adaptation [3.587367153279351]
Existing local Explainable AI (XAI) methods select a region of the input space in the vicinity of a given input instance, for which they approximate the behaviour of a model using a simpler and more interpretable surrogate model.
We propose a novel method, MASALA, for generating explanations, which automatically determines the appropriate local region of impactful model behaviour for each individual instance being explained.
arXiv Detail & Related papers (2024-08-19T15:26:45Z) - 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 Discovery of Interpretable Directions in h-space of
Pre-trained Diffusion Models [63.1637853118899]
We propose the first unsupervised and learning-based method to identify interpretable directions in h-space of pre-trained diffusion models.
We employ a shift control module that works on h-space of pre-trained diffusion models to manipulate a sample into a shifted version of itself.
By jointly optimizing them, the model will spontaneously discover disentangled and interpretable directions.
arXiv Detail & Related papers (2023-10-15T18:44:30Z) - Causal Disentangled Variational Auto-Encoder for Preference
Understanding in Recommendation [50.93536377097659]
This paper introduces the Causal Disentangled Variational Auto-Encoder (CaD-VAE), a novel approach for learning causal disentangled representations from interaction data in recommender systems.
The approach utilizes structural causal models to generate causal representations that describe the causal relationship between latent factors.
arXiv Detail & Related papers (2023-04-17T00:10:56Z) - Diffusion Action Segmentation [63.061058214427085]
We propose a novel framework via denoising diffusion models, which shares the same inherent spirit of such iterative refinement.
In this framework, action predictions are iteratively generated from random noise with input video features as conditions.
arXiv Detail & Related papers (2023-03-31T10:53:24Z) - Predictable MDP Abstraction for Unsupervised Model-Based RL [93.91375268580806]
We propose predictable MDP abstraction (PMA)
Instead of training a predictive model on the original MDP, we train a model on a transformed MDP with a learned action space.
We theoretically analyze PMA and empirically demonstrate that PMA leads to significant improvements over prior unsupervised model-based RL approaches.
arXiv Detail & Related papers (2023-02-08T07:37:51Z) - Generative Slate Recommendation with Reinforcement Learning [49.75985313698214]
reinforcement learning algorithms can be used to optimize user engagement in recommender systems.
However, RL approaches are intractable in the slate recommendation scenario.
In that setting, an action corresponds to a slate that may contain any combination of items.
In this work we propose to encode slates in a continuous, low-dimensional latent space learned by a variational auto-encoder.
We are able to (i) relax assumptions required by previous work, and (ii) improve the quality of the action selection by modeling full slates.
arXiv Detail & Related papers (2023-01-20T15:28:09Z) - Lifted Model Checking for Relational MDPs [12.574454799055026]
pCTL-REBEL is a lifted model checking approach for verifying pCTL properties on relational MDPs.
We show that the pCTL model checking approach is decidable for relational MDPs even for possibly infinite domains.
arXiv Detail & Related papers (2021-06-22T13:12:36Z) - Control as Hybrid Inference [62.997667081978825]
We present an implementation of CHI which naturally mediates the balance between iterative and amortised inference.
We verify the scalability of our algorithm on a continuous control benchmark, demonstrating that it outperforms strong model-free and model-based baselines.
arXiv Detail & Related papers (2020-07-11T19:44:09Z) - Data Driven Control with Learned Dynamics: Model-Based versus Model-Free
Approach [0.0]
We compare two types of data-driven control methods, representing model-based and model-free approaches.
One is a recently proposed method - Deep Koopman Representation for Control (DKRC), which utilizes a deep neural network to map an unknown nonlinear dynamical system to a high-dimensional linear system.
The other is a classic model-free control method based on an actor-critic architecture - Deep Deterministic Policy Gradient (DDPG), which has been proved to be effective in various dynamical systems.
arXiv Detail & Related papers (2020-06-16T22:18:21Z)
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.