Distillation Robustifies Unlearning

Current LLM unlearning methods are not robust: they can be reverted easily with a few steps of finetuning. This is true even for the idealized unlearning method of training to imitate an oracle model that was never exposed to unwanted information, suggesting that output-based finetuning is insufficient to achieve robust unlearning. In a similar vein, we find that training a randomly initialized student to imitate an unlearned model transfers desired behaviors while leaving undesired capabilities behind. In other words, distillation robustifies unlearning. Building on this insight, we propose Unlearn-Noise-Distill-on-Outputs (UNDO), a scalable method that distills an unlearned model into a partially noised copy of itself. UNDO introduces a tunable tradeoff between compute cost and robustness, establishing a new Pareto frontier on synthetic language and arithmetic tasks. At its strongest setting, UNDO matches the robustness of a model retrained from scratch with perfect data filtering while using only 60-80% of the compute and requiring only 0.01% of the pretraining data to be labeled. We also show that UNDO robustifies unlearning on the more realistic Weapons of Mass Destruction Proxy (WMDP) benchmark. Since distillation is widely used in practice, incorporating an unlearning step beforehand offers a convenient path to robust capability removal.

An Approach to Technical AGI Safety and Security

Artificial General Intelligence (AGI) promises transformative benefits but also presents significant risks. We develop an approach to address the risk of harms consequential enough to significantly harm humanity. We identify four areas of risk: misuse, misalignment, mistakes, and structural risks. Of these, we focus on technical approaches to misuse and misalignment. For misuse, our strategy aims to prevent threat actors from accessing dangerous capabilities, by proactively identifying dangerous capabilities, and implementing robust security, access restrictions, monitoring, and model safety mitigations. To address misalignment, we outline two lines of defense. First, model-level mitigations such as amplified oversight and robust training can help to build an aligned model. Second, system-level security measures such as monitoring and access control can mitigate harm even if the model is misaligned. Techniques from interpretability, uncertainty estimation, and safer design patterns can enhance the effectiveness of these mitigations. Finally, we briefly outline how these ingredients could be combined to produce safety cases for AGI systems.

Gradient Routing: Masking Gradients to Localize Computation in Neural Networks

Neural networks are trained primarily based on their inputs and outputs, without regard for their internal mechanisms. These neglected mechanisms determine properties that are critical for safety, like (i) transparency; (ii) the absence of sensitive information or harmful capabilities; and (iii) reliable generalization of goals beyond the training distribution. To address this shortcoming, we introduce gradient routing, a training method that isolates capabilities to specific subregions of a neural network. Gradient routing applies data-dependent, weighted masks to gradients during backpropagation. These masks are supplied by the user in order to configure which parameters are updated by which data points. We show that gradient routing can be used to (1) learn representations which are partitioned in an interpretable way; (2) enable robust unlearning via ablation of a pre-specified network subregion; and (3) achieve scalable oversight of a reinforcement learner by localizing modules responsible for different behaviors. Throughout, we find that gradient routing localizes capabilities even when applied to a limited, ad-hoc subset of the data. We conclude that the approach holds promise for challenging, real-world applications where quality data are scarce.

Steering Llama 2 via Contrastive Activation Addition

We introduce Contrastive Activation Addition (CAA), an innovative method for steering language models by modifying activations during their forward passes. CAA computes ``steering vectors'' by averaging the difference in residual stream activations between pairs of positive and negative examples of a particular behavior such as factual versus hallucinatory responses. During inference, these steering vectors are added at all token positions after the user's prompt with either a positive or negative coefficient, allowing precise control over the degree of the targeted behavior. We evaluate CAA's effectiveness on Llama 2 Chat using both multiple-choice behavioral question datasets and open-ended generation tasks. We demonstrate that CAA significantly alters model behavior, outperforms traditional methods like finetuning and few-shot prompting, and minimally reduces capabilities. Moreover, by employing various activation space interpretation methods, we gain deeper insights into CAA's mechanisms. CAA both accurately steers model outputs and also sheds light on how high-level concepts are represented in Large Language Models (LLMs).

Understanding and Controlling a Maze-Solving Policy Network

To understand the goals and goal representations of AI systems, we carefully study a pretrained reinforcement learning policy that solves mazes by navigating to a range of target squares. We find this network pursues multiple context-dependent goals, and we further identify circuits within the network that correspond to one of these goals. In particular, we identified eleven channels that track the location of the goal. By modifying these channels, either with hand-designed interventions or by combining forward passes, we can partially control the policy. We show that this network contains redundant, distributed, and retargetable goal representations, shedding light on the nature of goal-direction in trained policy networks.

Activation Addition: Steering Language Models Without Optimization

Reliably controlling the behavior of large language models is a pressing open problem. Existing methods include supervised finetuning, reinforcement learning from human feedback, prompt engineering, and guided decoding. We instead investigate activation engineering: modifying activations at inference time to predictably alter model behavior. In particular, we bias the forward pass with an added 'steering vector' implicitly specified through natural language. Unlike past work which learned these steering vectors, our Activation Addition (ActAdd) method computes them by taking the activation differences that result from pairs of prompts. We demonstrate ActAdd on GPT-2 on OpenWebText and ConceptNet. Our inference-time approach yields control over high-level properties of output and preserves off-target model performance. It involves far less compute and implementation effort than finetuning, allows users to provide natural language specifications, and its overhead scales naturally with model size.

Parametrically Retargetable Decision-Makers Tend To Seek Power

If capable AI agents are generally incentivized to seek power in service of the objectives we specify for them, then these systems will pose enormous risks, in addition to enormous benefits. In fully observable environments, most reward functions have an optimal policy which seeks power by keeping options open and staying alive. However, the real world is neither fully observable, nor will agents be perfectly optimal. We consider a range of models of AI decision-making, from optimal, to random, to choices informed by learning and interacting with an environment. We discover that many decision-making functions are retargetable, and that retargetability is sufficient to cause power-seeking tendencies. Our functional criterion is simple and broad. We show that a range of qualitatively dissimilar decision-making procedures incentivize agents to seek power. We demonstrate the flexibility of our results by reasoning about learned policy incentives in Montezuma's Revenge. These results suggest a safety risk: Eventually, highly retargetable training procedures may train real-world agents which seek power over humans.

Formalizing the Problem of Side Effect Regularization

AI objectives are often hard to specify properly. Some approaches tackle this problem by regularizing the AI's side effects: Agents must weigh off "how much of a mess they make" with an imperfectly specified proxy objective. We propose a formal criterion for side effect regularization via the assistance game framework. In these games, the agent solves a partially observable Markov decision process (POMDP) representing its uncertainty about the objective function it should optimize. We consider the setting where the true objective is revealed to the agent at a later time step. We show that this POMDP is solved by trading off the proxy reward with the agent's ability to achieve a range of future tasks. We empirically demonstrate the reasonableness of our problem formalization via ground-truth evaluation in two gridworld environments.

On Avoiding Power-Seeking by Artificial Intelligence

We do not know how to align a very intelligent AI agent's behavior with human interests. I investigate whether -- absent a full solution to this AI alignment problem -- we can build smart AI agents which have limited impact on the world, and which do not autonomously seek power. In this thesis, I introduce the attainable utility preservation (AUP) method. I demonstrate that AUP produces conservative, option-preserving behavior within toy gridworlds and within complex environments based off of Conway's Game of Life. I formalize the problem of side effect avoidance, which provides a way to quantify the side effects an agent had on the world. I also give a formal definition of power-seeking in the context of AI agents and show that optimal policies tend to seek power. In particular, most reward functions have optimal policies which avoid deactivation. This is a problem if we want to deactivate or correct an intelligent agent after we have deployed it. My theorems suggest that since most agent goals conflict with ours, the agent would very probably resist correction. I extend these theorems to show that power-seeking incentives occur not just for optimal decision-makers, but under a wide range of decision-making procedures.

Avoiding Side Effects in Complex Environments

Reward function specification can be difficult, even in simple environments. Realistic environments contain millions of states. Rewarding the agent for making a widget may be easy, but penalizing the multitude of possible negative side effects is hard. In toy environments, Attainable Utility Preservation (AUP) avoids side effects by penalizing shifts in the ability to achieve randomly generated goals. We scale this approach to large, randomly generated environments based on Conway's Game of Life. By preserving optimal value for a single randomly generated reward function, AUP incurs modest overhead, completes the specified task, and avoids side effects.