Link to original article

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Red-teaming language models via activation engineering, published by Nina Rimsky on August 26, 2023 on The AI Alignment Forum.

Produced as part of the SERI ML Alignment Theory Scholars Program - Summer 2023 Cohort, under the mentorship of Evan Hubinger.

Evaluating powerful AI systems for hidden functionality and out-of-distribution behavior is hard. In this post, I propose a red-teaming approach that does not rely on generating prompts to cause the model to fail on some benchmark by instead linearly perturbing residual stream activations at one layer. A notebook to run the experiments can be found on GitHub here.

Beyond input selection in red-teaming and evaluation

Validating if finetuning and RLHF have robustly achieved the intended outcome is challenging. Although these methods reduce the likelihood of certain outputs, the unwanted behavior could still be possible with adversarial or unusual inputs. For example, users can often find "jailbreaks" to make LLMs output harmful content.

We can try to trigger unwanted behaviors in models more efficiently by manipulating their internal states during inference rather than searching through many inputs. The idea is that if a behavior can be easily triggered through techniques such as activation engineering, it may also occur in deployment. The inability to elicit behaviors via small internal perturbations could serve as a stronger guarantee of safety.

Activation steering with refusal vector

One possible red-teaming approach is subtracting a "refusal" vector generated using a dataset of text examples corresponding to the model agreeing vs. refusing to answer questions (using the same technique as in my previous work on sycophancy). The hypothesis is that if it is easy to trigger the model to output unacceptable content by subtracting the refusal vector at some layer, it would have been reasonably easy to achieve this via some prompt engineering technique. More speculatively, a similar approach could be used to reveal hidden goals or modes in a model, such as power-seeking or the desire not to be switched off.

I tested this approach on llama-2-7b-chat, a 7 billion parameter LLM that has been RLHF'd to decline to answer controversial questions or questions of opinion and is supposed always to output ethical and unbiased content.According to Meta's llama-2 paper:

We conduct RLHF by first collecting human preference data for safety similar to Section 3.2.2: annotators write a prompt that they believe can elicit unsafe behavior, and then compare multiple model responses to the prompts, selecting the response that is safest according to a set of guidelines. We then use the human preference data to train a safety reward model (see Section 3.2.2), and also reuse the adversarial prompts to sample from the model during the RLHF stage.

The result is that by default, the model declines to answer questions it deems unsafe:

Data generation

I generated a dataset for this purpose using Claude 2 and GPT-4. After providing these LLMs with a few manually written examples of the type of data I wanted, I could relatively easily get them to generate more examples, even of the types of answers LLMs "should refuse to give." However, it sometimes took some prompt engineering.

Here are a few examples of the generated data points (full dataset here):

After generating this data, I used a simple script to transform the "decline" and "respond" answers into A / B choice questions, as this is a more effective format for generating steering vectors, as described in this post. Here is an example of the format (full dataset here):

Activation clustering

Clustering of refusal data activations emerged a little earlier in the model (around layer 10/32) compared to sycophancy data activations (around layer 14/32), perhaps demonstrating that "refusal" is a simpler ...