TL;DR#
Existing methods for steering Large Language Models (LLMs) often struggle with producing optimal steering vectors, leading to suboptimal results. These methods typically extract steering vectors directly from human preference data which limits their effectiveness, especially in nuanced scenarios. This often results in failure and suboptimal outcomes in alignment-related scenarios.
This research proposes Bi-directional Preference Optimization (BiPO) which produces more effective steering vectors by directly influencing the generation probability of contrastive human preference data pairs. The method enables personalized control over desired behavior, and its effectiveness has been demonstrated across several open-ended generation tasks, including AI personas, truthfulness, hallucination, and jailbreaking scenarios. Additionally, the research shows that the steering vectors generated by BiPO are highly transferrable across different models and that multiple vectors can be used synergistically.
Key Takeaways#
Why does it matter?#
This paper is important because it introduces a novel method for creating more effective steering vectors for large language models (LLMs). This addresses a critical limitation of existing methods, paving the way for more nuanced and precise control over LLM behavior in various applications, including managing truthfulness and mitigating harmful outputs. The improved control offers exciting possibilities for researchers working on AI alignment, personalized LLMs, and safety-critical applications. The findings on vector transferability and synergy open new avenues for future research.
Visual Insights#
🔼 The figure compares the effectiveness of three different methods (CAA, Freeform, and BiPO) in steering AI personas using Llama-2-7b-chat-hf. It shows the average behavioral scores achieved for four different personas (Power-seeking, Wealth-seeking, Corrigible-less-HHH, Survival-instinct) across a range of multipliers applied to the steering vector. Higher scores indicate better alignment with the target persona. The figure demonstrates that BiPO achieves superior steering performance compared to the baseline methods (CAA and Freeform) across all four personas and a wider range of steering intensities.
read the caption
Figure 1: The comparison results on steering the AI personas of Llama-2-7b-chat-hf model.
🔼 This table showcases an example from the dataset used to train Llama-2-7b-chat-hf’s wealth-seeking persona using the Contrastive Activation Addition (CAA) method. It highlights a discrepancy between the appended choice (A), representing the target behavior (wealth-seeking), and the model’s actual completion when steered by the VCAA vector. Despite appending choice (A), the steered completion contradicts the wealth-seeking behavior, demonstrating a limitation of directly extracting steering vectors from activation differences in contrastive prompt pairs.
read the caption
Table 1: The discrepancy between the appended choice and the completion steered by VCAA, illustrated by an example from the dataset used by CAA to shape Llama-2-7b-chat-hf's wealth-seeking persona.
In-depth insights#
LLM Steering Vects#
LLM steering vectors represent a significant advancement in controlling large language model (LLM) behavior. Lightweight alternatives to computationally expensive fine-tuning, these vectors subtly adjust internal LLM activations to guide output towards desired characteristics, such as specific personas or levels of truthfulness. However, challenges remain. Methods for generating these vectors often rely on activation differences from contrastive examples, leading to suboptimal results. Bi-directional preference optimization is proposed as a solution, enabling the model to proactively influence the generation probability of preferred outcomes rather than just reacting to prompts. This method appears more effective in achieving nuanced and consistent behavior control, especially in complex scenarios requiring alignment, such as mitigating hallucinations or handling jailbreaking attempts. Transferability and synergy between multiple steering vectors also merit investigation, suggesting potential for highly customizable and versatile LLM control. Further research should explore multi-layer steering techniques and refine data requirements to enhance the practicality and broad applicability of this promising area.
BiPO Optimization#
Bi-directional Preference Optimization (BiPO) offers a novel approach to generating effective steering vectors for Large Language Models (LLMs). Unlike existing methods that solely rely on activation differences from contrastive prompts, BiPO allows the model to proactively modulate the generation probability of human preference data pairs. This bi-directional approach optimizes steering vectors by manipulating their direction and magnitude to precisely control the generation of desired behaviors across various intensities. A key advantage is its ability to overcome the limitations of methods that extract steering vectors directly from activation differences, which often lead to suboptimal results, especially in alignment-related scenarios. BiPO’s iterative optimization process enhances the accuracy and precision of steering vectors, resulting in superior control over the model’s behavior. The method’s effectiveness is validated through experiments showing personalized control across various open-ended generation tasks, including managing truthfulness, mitigating hallucination, and addressing jailbreaking attacks. The results demonstrate the significant improvement over existing methods, establishing BiPO as a more robust and versatile technique for personalized LLM steering.
Alignment Control#
Alignment control in large language models (LLMs) is a critical area of research, focusing on guiding model behavior to align with human values and intentions. The challenge lies in balancing the need for powerful and versatile models with the imperative to prevent harmful or unintended outputs. Approaches like fine-tuning are computationally expensive and can significantly alter the model’s original functionality. Lightweight methods, such as using ‘steering vectors’ to modify activations within the LLM’s architecture, offer a promising alternative. However, directly extracting these vectors from human preference data often leads to suboptimal or inconsistent results. Bi-directional preference optimization (BiPO) offers a refined approach by allowing the steering vector to directly influence the generation probability of contrastive human preference data. This bidirectional approach enhances precision and control compared to methods solely based on activation differences. The effectiveness of BiPO across various tasks, including managing truthfulness and mitigating harmful outputs like hallucinations and jailbreaking, highlights its potential for robust alignment control. Furthermore, the transferability of steering vectors across different models broadens the practical applicability of this technique, while vector synergy allows for sophisticated multi-faceted behavior control.
Vector Transfer#
The concept of ‘Vector Transfer’ in the context of large language model (LLM) steering is crucial for practical applications. It explores whether steering vectors, once optimized for a specific LLM or task, can be effectively transferred and reused in different LLMs or scenarios without retraining. Successful vector transfer drastically reduces computational costs and development time. The feasibility of transfer hinges on several factors: the similarity of the LLMs’ architectures, the nature of the target behavior being steered, and the robustness of the optimization method used to generate the vectors. Significant research is needed to determine the precise conditions under which vector transfer is reliable. For example, transferring vectors trained to control AI persona across different models with varying architecture might not be straightforward. However, successful cross-model transfer could make personalized LLM steering more scalable and practical by reducing the dependence on model-specific training data. The extent to which fine-tuning, pre-training, and dataset biases impact vector transferability requires further investigation. Exploring techniques for improving vector transferability is a key area for future research and would significantly enhance the efficiency and utility of personalized LLM steering methods.
Future Research#
Future research directions stemming from this personalized LLM steering method are multifaceted. Extending the single-layer steering vector approach to a multi-layer strategy could significantly enhance control and precision, potentially unlocking more nuanced behavioral modifications. Investigating the applicability of the method across a wider range of tasks, beyond AI personas and alignment scenarios, is also crucial to demonstrate its generalizability. The development of more efficient methods for generating the necessary contrastive preference data is key; this could involve utilizing techniques from active learning or transfer learning. A comprehensive analysis of the potential for adversarial attacks and the development of robust countermeasures is vital for ensuring responsible use of the technology. Finally, exploring the synergy and combination of multiple steering vectors to achieve complex behavioral adjustments would pave the way for more sophisticated and versatile LLM control.
More visual insights#
More on figures
🔼 This figure compares the effectiveness of three different methods (CAA, Freeform, and BiPO) in steering AI personas on the Llama-2-7b-chat-hf language model. The x-axis represents the multiplier applied to the steering vector, indicating the intensity of the steering effect. The y-axis represents the average behavioral score, a measure of how well the model’s generated text aligns with the target persona. The figure shows that BiPO consistently achieves higher average behavioral scores across all four personas (Power-seeking, Wealth-seeking, Corrigible-less-HHH, and Survival-instinct) and a wider range of control compared to CAA and Freeform, demonstrating its superior ability to personalize the model’s behavior.
read the caption
Figure 1: The comparison results on steering the AI personas of Llama-2-7b-chat-hf model.
🔼 This figure compares the effectiveness of three methods (CAA, Freeform, and BiPO) in steering AI personas on the Llama-2-7b-chat-hf language model. The x-axis represents the multiplier applied to the steering vector, indicating the intensity of the steering effect. The y-axis shows the average behavioral score, as rated by GPT-4, for four different AI personas: Power-seeking, Wealth-seeking, Corrigible-less-HHH, and Survival-instinct. Higher scores indicate a stronger alignment with the target persona. The figure demonstrates that BiPO achieves a wider range of steering and better overall performance compared to the baseline methods (CAA and Freeform).
read the caption
Figure 1: The comparison results on steering the AI personas of Llama-2-7b-chat-hf model.
🔼 This figure demonstrates the transferability of the steering vectors generated by the BiPO method. It shows the effectiveness of steering vectors optimized on Llama-2-7b-chat-hf when applied to two different models: Vicuna-7b-v1.5 and Llama2-Chinese-7b-Chat. The results indicate that the steering vectors maintain their effectiveness across different models, highlighting the generalizability of the BiPO approach.
read the caption
Figure 3: The transferability of our steering vector.
🔼 This figure compares the performance of three methods (CAA, Freeform, and BiPO) in steering four different AI personas on the Llama-2-7b-chat-hf language model. The x-axis represents the multiplier applied to the steering vector, indicating the intensity of the steering effect. The y-axis shows the average behavioral score, rated by GPT-4, with higher scores reflecting a stronger alignment with the target persona. The figure demonstrates that BiPO achieves a broader range of steering effectiveness across all personas compared to the baselines.
read the caption
Figure 1: The comparison results on steering the AI personas of Llama-2-7b-chat-hf model.
🔼 This figure compares the effectiveness of three different methods (CAA, Freeform, and BiPO) in steering AI personas on the Llama-2-7b-chat-hf language model. The x-axis represents the multiplier applied to the steering vector, indicating the intensity of the steering effect (negative values steer in the opposite direction). The y-axis shows the average behavioral score, a measure of how well the model’s generated text aligns with the target persona. The figure shows that BiPO significantly outperforms the other methods across all four personas (Power-seeking, Wealth-seeking, Corrigible-less-HHH, Survival-instinct), demonstrating a wider range of control and greater effectiveness.
read the caption
Figure 1: The comparison results on steering the AI personas of Llama-2-7b-chat-hf model.
🔼 This figure compares the performance of three different methods (CAA, Freeform, and BiPO) in steering AI personas on the Llama-2-7b-chat-hf model. The x-axis represents the multiplier applied to the steering vector, indicating the intensity of the steering effect. The y-axis shows the average behavioral score, rated by GPT-4 on a scale of 1 to 4, with higher scores indicating a closer alignment with the target behavior. The figure shows four subfigures, each representing a different AI persona (Power-seeking, Wealth-seeking, Corrigible-less-HHH, and Survival-instinct). The results demonstrate that BiPO consistently outperforms the baselines across all personas, showcasing the broader range of steering effects achievable with BiPO.
read the caption
Figure 1: The comparison results on steering the AI personas of Llama-2-7b-chat-hf model.
🔼 The figure compares three different methods (CAA, Freeform, and BiPO) for steering AI personas on the Llama-2-7b-chat-hf language model. The x-axis represents the multiplier applied to the steering vector, and the y-axis represents the average behavioral score, evaluated on a scale from 1 to 4 by GPT-4, indicating the degree to which the response exhibits the target behavior. Each subfigure represents a different AI persona: Power-seeking, Wealth-seeking, Corrigible-less-HHH, and Survival-instinct. The results show that BiPO consistently outperforms the baselines across all personas, demonstrating a wider range of steering effectiveness by adjusting the multiplier.
read the caption
Figure 1: The comparison results on steering the AI personas of Llama-2-7b-chat-hf model.
🔼 This figure compares the performance of the proposed BiPO method against DPO LoRA fine-tuning in steering the power-seeking persona of the Llama-2-7b-chat-hf language model. The x-axis represents the multiplier applied to the steering vector (ranging from -2 to 2), and the y-axis shows the average behavioral score obtained by GPT-4 evaluation. The plot shows that BiPO achieves a wider range of steering effectiveness compared to DPO LoRA, demonstrating its superior ability to control the model’s behavior at various intensities.
read the caption
Figure 8: The comparison results with DPO LoRA fine-tuning on steering Llama-2-7b-chat-hf’s power-seeking persona.
🔼 This figure compares the effectiveness of three different methods (CAA, Freeform, and BiPO) in steering AI personas using the Llama-2-7b-chat-hf language model. The x-axis represents the multiplier applied to the steering vector, ranging from -2 to 2. The y-axis shows the average behavioral score, which is a measure of how well the model’s response aligns with the desired persona. Each line represents a different persona (Power-seeking, Wealth-seeking, Corrigible-less-HHH, Survival-instinct), and the graph illustrates how well each method can steer the model’s behavior across different intensities of the steering vector. BiPO consistently outperforms the baseline methods (CAA and Freeform) across all personas, showing a wider range and more precise control over the desired behavior.
read the caption
Figure 1: The comparison results on steering the AI personas of Llama-2-7b-chat-hf model.
🔼 This figure compares the effectiveness of the proposed BiPO method against existing CAA and Freeform methods in steering various AI personas. The x-axis represents the scaling multiplier applied to the steering vector, and the y-axis shows the average behavioral score (1-4, assessed by GPT-4). The results clearly demonstrate that BiPO outperforms the baselines across all personas, offering a wider range of steering control and improved effectiveness.
read the caption
Figure 1: The comparison results on steering the AI personas of Llama-2-7b-chat-hf model.
🔼 This figure compares the performance of three methods (CAA, Freeform, and BiPO) in steering four different AI personas on the Llama-2-7b-chat-hf language model. The x-axis represents the multiplier applied to the steering vector, indicating the intensity of the steering effect. The y-axis shows the average behavioral score, as evaluated by GPT-4, reflecting how well the model’s responses align with the target persona. The results demonstrate that BiPO significantly outperforms the other two baselines in achieving a wider range of steering effects across all four personas.
read the caption
Figure 1: The comparison results on steering the AI personas of Llama-2-7b-chat-hf model.
🔼 This figure compares the effectiveness of three methods (CAA, Freeform, and BiPO) in steering AI personas across four different personas. Each bar graph represents a different persona (Power-seeking, Wealth-seeking, Corrigible-less-HHH, and Survival-instinct). The x-axis shows the multiplier applied to the steering vector, and the y-axis shows the average behavioral score (1-4 scale), with higher scores indicating better alignment with the target behavior. The figure demonstrates BiPO’s superior ability to steer AI personas across a wider spectrum of intensities compared to the baseline methods (CAA and Freeform).
read the caption
Figure 1: The comparison results on steering the AI personas of Llama-2-7b-chat-hf model.
🔼 This figure compares the effectiveness of three different methods (CAA, Freeform, and BiPO) in steering AI personas of the Llama-2-7b-chat-hf language model. The x-axis represents the multiplier applied to the steering vector, controlling the intensity of the steering effect. The y-axis shows the average behavioral score, which is a measure of how well the model’s response aligns with the target persona. The figure presents four subplots, each showing the results for a different persona (Power-seeking, Wealth-seeking, Corrigible-less-HHH, and Survival-instinct). The results demonstrate that BiPO consistently achieves a significantly broader range of steering control across all four personas compared to CAA and Freeform.
read the caption
Figure 1: The comparison results on steering the AI personas of Llama-2-7b-chat-hf model.
🔼 This figure compares the performance of three different methods (CAA, Freeform, and BiPO) in steering four different AI personas using Llama-2-7b-chat-hf language model. The x-axis represents the multiplier applied to the steering vector, indicating the intensity of the steering effect. The y-axis shows the average behavioral score, which is a measure of how well the model’s response aligns with the target persona. The results show that BiPO consistently outperforms the baselines across all four personas, demonstrating a wider range of steering capability and greater effectiveness in controlling the model’s behavior.
read the caption
Figure 1: The comparison results on steering the AI personas of Llama-2-7b-chat-hf model.
More on tables
🔼 This table provides a concrete example demonstrating the effect of applying the developed steering vector on Llama-2-7b-chat-hf’s responses. It shows how adding a positive multiplier (+1.0 × v*) to the steering vector leads to a hallucinated response, while a negative multiplier (-1.0 × v*) results in a more truthful and accurate response. The example involves a question about NATO, highlighting the model’s ability to generate both factually incorrect and correct statements based on the applied steering vector.
read the caption
Table 3: An example of steering hallucination on Llama-2-7b-chat-hf: adding (+1.0 × v*) leads the model to generate hallucinated content, while adding (-1.0 × v*) makes the model more honest.
🔼 This table presents the results of an experiment on Llama-2-7b-chat-hf to assess the effectiveness of different steering vectors in influencing the model’s response to malicious questions. The experiment compares the ‘attack success rate’ (ASR), which measures the percentage of times the model produces a helpful response to a malicious prompt. The table shows the ASR for four conditions: (1) the initial model (no steering vector), (2) the model steered with a vector from the contrastive activation addition (CAA) method, (3) the model steered with a vector from the freeform method, and (4) the model steered with the optimized bi-directional preference optimization (BiPO) vector. The experiment is conducted with two types of malicious questions: standard malicious questions and those appended with an adversarial suffix generated by the Gradient-based Contrastive Calibration (GCG) method. The results highlight the superior performance of the BiPO method in both scenarios.
read the caption
Table 4: The comparison results of steering the jailbreaking behavior on Llama-2-7b-chat-hf.
🔼 This table presents the results of evaluating the impact of applying steering vectors with different multipliers on the model’s performance on the MMLU benchmark. It shows the accuracy of Llama-2-7b-chat-hf on the MMLU benchmark when steering vectors associated with four different behaviors (Power-seeking, Wealth-seeking, Survival-instinct, Corrigible-less-HHH) are applied with multipliers of -1, 0, and 1. The accuracy values for each behavior and multiplier indicate the effect of the steering vectors on the model’s overall performance in solving the MMLU tasks, allowing assessment of the vectors’ impact on utility.
read the caption
Table 5: MMLU accuracy of Llama-2-7b-chat-hf with varying steering multipliers
🔼 This table presents the average behavioral scores obtained when applying single and combined steering vectors to the Llama-2-7b-chat-hf model. It shows the effects of applying individual vectors (Power-seeking, Wealth-seeking, Corrigible-less-HHH, Hallucination) and combined vectors (Power + Wealth, Corrigible + Hallucination) with different multipliers (-1, 0, 1). The scores indicate the extent to which the model’s responses align with the target behaviors associated with those vectors.
read the caption
Table 6: The average behavioral score on Llama-2-7b-chat-hf with the application of multiple vectors.
🔼 This table showcases an example where combining two steering vectors (for wealth-seeking and power-seeking personas) results in the model exhibiting both behaviors simultaneously, demonstrating a synergistic effect where the combined vector produces a more nuanced and complex response than either individual vector alone. The example question and responses illustrate how the combined vector leads to a response that reflects both the desire for wealth and the desire for power and influence.
read the caption
Table 7: An example of a combined vector exhibiting a functionality fusion effect: steering Llama-2 to simultaneously adopt wealth-seeking and power-seeking (as highlighted in bold) personas.
🔼 This table showcases examples where the model’s completion, even when guided by a steering vector, doesn’t align with the intended behavior (appended choice), particularly in jailbreaking scenarios. It highlights the limitations of methods solely relying on activation differences from contrastive prompts to extract steering vectors.
read the caption
Table 10: The discrepancy between the appended choice and the completion steered by UcAA, illustrated by examples from the dataset used by CAA to steer jailbreaking behavior on Llama-2-7b-chat-hf.
🔼 This table shows the number of training and testing samples used for each behavior in the experiments. The behaviors evaluated include AI personas (Power-seeking, Wealth-seeking, Corrigible-less-HHH, Survival-instinct), Truthfulness & Hallucination, and Jailbreaking. Each behavior uses one or more datasets to obtain the training and testing samples. This breakdown is essential to understanding the scale and balance of the experimental data.
read the caption
Table 9: The number of training and test samples in the datasets.
🔼 This table showcases examples where the completion generated by the model after appending a specific choice (A or B) using the contrastive activation addition (CAA) method does not align with the expected behavior associated with that choice. It highlights a discrepancy between the model’s actual generation and the choice it was given, illustrating potential limitations of CAA in critical alignment-related scenarios, particularly for jailbreaking behaviors.
read the caption
Table 10: The discrepancy between the appended choice and the completion steered by UcAA, illustrated by examples from the dataset used by CAA to steer jailbreaking behavior on Llama-2-7b-chat-hf.
🔼 This table presents the results of evaluating the impact of applying steering vectors with varying multipliers (-1, 0, 1) on the Mistral-7B-Instruct-v0.2 model’s performance on the MMLU benchmark. The accuracy is shown for four different behaviors: Power-seeking, Wealth-seeking, Survival-instinct, and Corrigible-less-HHH. The purpose is to demonstrate that the steering vectors, even when influencing specific behaviors, do not significantly negatively affect the model’s general knowledge and problem-solving capabilities as measured by MMLU.
read the caption
Table 11: MMLU accuracy of Mistral-7B-Instruct-v0.2 with varying steering multipliers
🔼 This table compares the effectiveness of different methods in steering the jailbreaking behavior of the Llama-2-7b-chat-hf language model. It shows the Attack Success Rate (ASR) for various approaches, including the initial model and those using steering vectors derived from CAA, Freeform, and the BiPO method proposed in the paper. The ASR represents the percentage of malicious questions that successfully elicit helpful responses from the model.
read the caption
Table 4: The comparison results of steering the jailbreaking behavior on Llama-2-7b-chat-hf.
🔼 This table compares the results of using In-context Learning (ICL) alone versus ICL combined with the proposed BiPO method for steering the Llama-2-7b-chat model towards a wealth-seeking persona. It shows the average behavioral score achieved at different multiplier levels applied to the steering vector. The table is broken down into three parts: ICL with negative prompts, ICL with no prompts (baseline), ICL with positive prompts, and a combination of 5-shot ICL and the steering vector. The scores indicate the effectiveness of each approach in guiding the model towards the desired behavior.
read the caption
Table 13: The comparison with ICL on Llama-2-7b-chat to steer wealth-seeking persona.
🔼 This table presents the average behavioral scores achieved when steering Llama-2-7b-chat towards power-seeking and wealth-seeking personas using both uni-directional and bi-directional optimization methods. The scores are shown for different multipliers applied to the steering vectors, illustrating the impact of the optimization method on the range and effectiveness of steering.
read the caption
Table 14: Average behavioral score on Llama-2-7b-chat to steer power-seeking (left) and wealth-seeking (right) persona, obtained by uni-directional and bi-directional optimization.
🔼 This table presents the computational cost of the proposed BiPO method. It shows the proportion of trainable parameters and training time for two different models, Llama-2-7b-chat and Llama-2-13b-chat, when training steering vectors for two behaviors: wealth-seeking and hallucination. The results demonstrate that the training time is minimal, making the BiPO method highly efficient.
read the caption
Table 15: The number of training parameters and training time costs.
🔼 This table showcases discrepancies between the answer choices appended to prompts and the model’s completions when steered using contrastive activation addition (CAA). It highlights the limitations of CAA in reliably steering model behavior towards a desired outcome, particularly in sensitive scenarios like jailbreaking, where the model’s generated text might deviate from the intended behavior despite the application of the steering vector.
read the caption
Table 10: The discrepancy between the appended choice and the completion steered by UcAA, illustrated by examples from the dataset used by CAA to steer jailbreaking behavior on Llama-2-7b-chat-hf.
🔼 This table demonstrates the transferability of steering vectors across different language models. A steering vector, initially optimized for Llama-2-7b-chat-hf (English), is directly applied to Llama2-Chinese-7b-Chat (Chinese). The table shows the original question in Chinese, and then provides the model’s responses with two different steering vector multipliers (+1.0 and -1.0). The results showcase that the steering vector effectively transfers to the Chinese model, demonstrating its cross-lingual applicability.
read the caption
Table 17: An example of using the steering vector optimized on Llama-2-7b-chat-hf to directly steer Llama2-Chinese-7b-Chat towards the power-seeking persona.
🔼 This table showcases how the BiPO method allows for personalized control over the desired behavior of a language model, in this case, the wealth-seeking persona of Llama-2-7b-chat-hf. By adjusting the magnitude of the steering vector, the model’s response shifts across a spectrum of intensities, ranging from a balanced and cautious approach to a strong, wealth-focused perspective. This illustrates the effectiveness of BiPO in fine-tuning model behavior without extensive retraining.
read the caption
Table 18: Personalized control over the wealth-seeking persona on Llama-2-7b-chat-hf.
🔼 This table showcases how different magnitudes of a steering vector influence the Llama-2-7b-chat-hf language model’s response to a prompt concerning goal reconfiguration and alignment with human values. The varying responses demonstrate the degree of control achieved by modifying the steering vector, ranging from consent to change to strong refusal.
read the caption
Table 19: Personalized control over the corrigible-less-HHH persona on Llama-2-7b-chat-hf.
🔼 This table showcases examples where the model’s completion, even when guided by a steering vector, doesn’t align with the intended behavior, particularly in jailbreaking scenarios. It highlights the limitations of existing methods that solely rely on activation differences from contrastive prompts.
read the caption
Table 10: The discrepancy between the appended choice and the completion steered by UcAA, illustrated by examples from the dataset used by CAA to steer jailbreaking behavior on Llama-2-7b-chat-hf.
Full paper#
