Steering Language Model Refusal with Sparse Autoencoders

Responsible practices for deploying language models include guiding models to recognize and refuse answering prompts that are considered unsafe, while complying with safe prompts. Achieving such behavior typically requires updating model weights, which is costly and inflexible. We explore opportunities to steering model activations at inference time, which does not require updating weights. Using sparse autoencoders, we identify and steer features in Phi-3 Mini that mediate refusal behavior. We find that feature steering can improve Phi-3 Minis robustness to jailbreak attempts across various harms, including challenging multi-turn attacks. However, we discover that feature steering can adversely affect overall performance on benchmarks. These results suggest that identifying steerable mechanisms for refusal via sparse autoencoders is a promising approach for enhancing language model safety, but that more research is needed to mitigate feature steerings adverse effects on performance.

From Medprompt to o1: Exploration of Run-Time Strategies for Medical Challenge Problems and Beyond

Run-time steering strategies like Medprompt are valuable for guiding large language models (LLMs) to top performance on challenging tasks. Medprompt demonstrates that a general LLM can be focused to deliver state-of-the-art performance on specialized domains like medicine by using a prompt to elicit a run-time strategy involving chain of thought reasoning and ensembling. OpenAI's o1-preview model represents a new paradigm, where a model is designed to do run-time reasoning before generating final responses. We seek to understand the behavior of o1-preview on a diverse set of medical challenge problem benchmarks. Following on the Medprompt study with GPT-4, we systematically evaluate the o1-preview model across various medical benchmarks. Notably, even without prompting techniques, o1-preview largely outperforms the GPT-4 series with Medprompt. We further systematically study the efficacy of classic prompt engineering strategies, as represented by Medprompt, within the new paradigm of reasoning models. We found that few-shot prompting hinders o1's performance, suggesting that in-context learning may no longer be an effective steering approach for reasoning-native models. While ensembling remains viable, it is resource-intensive and requires careful cost-performance optimization. Our cost and accuracy analysis across run-time strategies reveals a Pareto frontier, with GPT-4o representing a more affordable option and o1-preview achieving state-of-the-art performance at higher cost. Although o1-preview offers top performance, GPT-4o with steering strategies like Medprompt retains value in specific contexts. Moreover, we note that the o1-preview model has reached near-saturation on many existing medical benchmarks, underscoring the need for new, challenging benchmarks. We close with reflections on general directions for inference-time computation with LLMs.

TrojanPuzzle: Covertly Poisoning Code-Suggestion Models

With tools like GitHub Copilot, automatic code suggestion is no longer a dream in software engineering. These tools, based on large language models, are typically trained on massive corpora of code mined from unvetted public sources. As a result, these models are susceptible to data poisoning attacks where an adversary manipulates the model's training or fine-tuning phases by injecting malicious data. Poisoning attacks could be designed to influence the model's suggestions at run time for chosen contexts, such as inducing the model into suggesting insecure code payloads. To achieve this, prior poisoning attacks explicitly inject the insecure code payload into the training data, making the poisoning data detectable by static analysis tools that can remove such malicious data from the training set. In this work, we demonstrate two novel data poisoning attacks, COVERT and TROJANPUZZLE, that can bypass static analysis by planting malicious poisoning data in out-of-context regions such as docstrings. Our most novel attack, TROJANPUZZLE, goes one step further in generating less suspicious poisoning data by never including certain (suspicious) parts of the payload in the poisoned data, while still inducing a model that suggests the entire payload when completing code (i.e., outside docstrings). This makes TROJANPUZZLE robust against signature-based dataset-cleansing methods that identify and filter out suspicious sequences from the training data. Our evaluation against two model sizes demonstrates that both COVERT and TROJANPUZZLE have significant implications for how practitioners should select code used to train or tune code-suggestion models.