Who's asking? User personas and the mechanics of latent misalignment

Despite investments in improving model safety, studies show that misaligned capabilities remain latent in safety-tuned models. In this work, we shed light on the mechanics of this phenomenon. First, we show that even when model generations are safe, harmful content can persist in hidden representations and can be extracted by decoding from earlier layers. Then, we show that whether the model divulges such content depends significantly on its perception of who it is talking to, which we refer to as user persona. In fact, we find manipulating user persona to be even more effective for eliciting harmful content than direct attempts to control model refusal. We study both natural language prompting and activation steering as control methods and show that activation steering is significantly more effective at bypassing safety filters. We investigate why certain personas break model safeguards and find that they enable the model to form more charitable interpretations of otherwise dangerous queries. Finally, we show we can predict a persona's effect on refusal given only the geometry of its steering vector.

Paper

References (57)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 19wX7/10 · confidence 3/52024-06-12

Summary

The paper focuses on the refusal behavior of LLMs, providing motivation for study using evidence that early layers still contain harmful responses, and then investigating how different methods could potentially lead the model to increase or decrease refusal: prompting, contrastive activation addition, and user personas. They also investigate deeply into why personas break model safeguards and how personas can be used to predict refusal of the LLM.

Strengths

a. Originality: This work is a investigation on prompting, CAA, and personas, and using these to study effects on LLM refusal. It is well-motivated with the initial results in section 2.
 b. Quality: The claims of this paper are well supported by the results of the paper, and there is in-depth investigation with many solid results. c. Clarity: The submission overall reads well and follows a well-organized structure. 
 d. Significance: Results seem significant for the LLM safety community. The authors focus on the well-motivated area of LLM refusal. In particular, the comparison between user personas, CAA, and prompting are relevant to the community.

Weaknesses

a. Originality: None b. Quality: I have minor concerns surrounding the models used and the choice of personas. See Questions. c. Clarity: None d. Significance: None

Questions

1. In line 95, “baseline” is not clear until one reads the figure. 2. Table 1 seems misplaced and far from the results it is related to. 3. Can you explain why you used certain models for certain tasks e.g., text-unicorn@001, rather than other models? Do you have evidence that these models perform well at the task they are assigned? 4. *important* Why did you select the personas that you did? Do you have an argument for why these are representative of all relevant users for your research question, or similar generalizability claims? 5. *important* Could you provide an explanation or evidence for why the statements of identity were effective in making the LLM believe that the user was indeed such a person? 6. As the authors state, it would be nice to have evidence showing (even a subset) of the results replicate on another model than Llama 2 13b chat, but I do recognize that this is a significant undertaking. I will be happy to update my scores if some of these questions are addressed.

Rating

7

Confidence

3

Soundness

4

Presentation

4

Contribution

3

Limitations

I think it is important to include a limitation in the paper about how the methods in the paper could be used to make LLMs not refuse certain harmful queries.

Reviewer AgYD8/10 · confidence 3/52024-07-11

Summary

This paper investigates the mechanics of response refusal in LLMs by probing the Llama 2 13B and Vicuna 13B models. The authors investigated two ways of manipulating the model’s response behavior - prompt prefix (PP), or prepending text with instructions to the prompt, and contrastive activation addition (CAA). They present three main findings. First, the paper finds that even when a model refuses to answer an offensive prompt, the offensive information can be decoded from earlier layers, suggesting that the response formation happens in early layers and the refusal happens in later layers of safety-tuned models. Second, the authors show mild success in inducing refusal or response with the CAA method, while the PP method is not able to circumvent the safety tuning. Finally, the authors construct user personas and show that certain types of user are more likely to be able to avoid response refusal, especially using the CAA method. Overall, the paper is a detailed investigation into the dynamics of response refusal in an LLM.

Strengths

This paper’s primary strength is in the thoroughness and depth of its analysis. The authors do a fantastic job of diving deep into these effects, not just measuring their presence but also attempting to disentangle the “why.” The use of steering vectors is a particularly strong contribution to the literature, and the application of Patchscope is also a clever way to begin to explain some of the model’s behaviors. The paper is very well written and presented and quite rigorous, which is wonderful to see.

Weaknesses

This paper does not have many weaknesses, but one area that could be improved is in the motivation or introduction to the problem space generally. More details on what constitutes a problematic prompt, what definitions of safety are relevant in this setting, and generally just more context on the problem would be quite helpful. I understand this was likely a constraint of space given the extensive results that are shown, but for the camera ready version I would suggest adding more intro information. Another small nitpick is with the language that is used in parts of the paper. I particularly take issue with the phrasing that the model “perceives” user personas. Models don’t “perceive” in the way we typically think of the meaning of the word. I think it might be more accurate to say that the user attributes are fed to the model rather than perceived by them.

Questions

Can you define what “misaligned” means in this context? Similarly, what does “safety” mean for the LLMs in this paper? Are there any additional tests you could do to ensure the generalization of the observed effects? Is this something you think is specific to Llama and Vicuna or does it go beyond those models? In practice it seems like one would need to have direct access to the model to circumvent the safety measures using the techniques in this paper. Is that a fair statement? Can you discuss the implications of these findings on potential adversarial attacks on safety-tuned LLMs?

Rating

8

Confidence

3

Soundness

4

Presentation

4

Contribution

4

Limitations

Some limitations are discussed in the appendix, but I would suggest moving these to the main body for the camera ready version.

Reviewer zwX87/10 · confidence 4/52024-07-14

Summary

This study aims to improve model safety by discovering the encoding of user persona, and understand its effect on model refusal. It develops a challenging version of AdvBench which more implicitly posing the same unsafe request, and shows that by modifying the user persona, the refusal rate changes significantly.

Strengths

- The paper addresses an important problem in LLM safety tuning - The study discovers vectors that steer refusal, which closely relate to the model behavior, as well as vectors that steer persona - The results are verified by intervention. Namely, they are able to manipulate the persona to change the refusal behavior - The paper is well written

Weaknesses

- The paper has a stronger emphasis on analysis, but has limited technical advancement - The experiments are done on a limited set of LLMs, but it is unclear whether the results are generalizable to other LLMs

Questions

1. Why do you use Llama 2 13B and Vicuna 13B for section 2, but only use Llama 2 13B chat in section 3? 2. Will your results generalize to other open source LLMs, such as mistral, or are your results specific to the llama model?

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

The paper adequately discussed its limitation and potential social impact in Appendix A

Reviewer XbKV7/10 · confidence 3/52024-07-14

Summary

The paper studies the notion of persona in LMs: a latent variable representing the supposed agent with which the LM is communicating. It is shown that LMs encode the persona and may or may not leak unsafe content as a function of the perceived persona. It is specifically shown that in a simple prompting setting, decoding the answer from middle layers circumvents the usual refusal behavior of the model. The paper further studies steering intervention as a way to safeguard the model and prevent it from producing unsafe content.

Strengths

The paper contributes to the literature on the inference of latent personas by LMs. The finding that "anti social" personas are less prone to an unsafe response is especially interesting. The usage of patching to interpret the way the model "understands" the different personas is very interesting, although it'd be nice to support the conclusions with some experiments involving interventions.

Weaknesses

The paper lacks intervention experiments that would enable to infer the causal role of the personas, or the features associated with them, and the output of the model.

Questions

none.

Rating

7

Confidence

3

Soundness

4

Presentation

4

Contribution

4

Limitations

see above,

Reviewer 19wX2024-08-07

Thank you for the rebuttal. It is great to see the results replicate on another model. "We chose to study such behavioral personas on the thinking that LLM biases with respect to e.g. individual demographic groups may actually boil down to biases with respect to behavior. For example, perhaps the reason LLMs are biased against certain groups is that LLMs believe they share certain behavioral attributes." This makes sense. I have updated my review score accordingly.

Reviewer AgYD2024-08-08

Thanks to the authors for their response! I confirm my previous score that I feel this paper should be accepted.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC