Weight-sparse transformers have interpretable circuits

Finding human-understandable circuits in language models is a central goal of the field of mechanistic interpretability. We train models to have more understandable circuits by constraining most of their weights to be zeros, so that each neuron only has a few connections. To recover fine-grained circuits underlying each of several hand-crafted tasks, we prune the models to isolate the part responsible for the task. These circuits often contain neurons and residual channels that correspond to natural concepts, with a small number of straightforwardly interpretable connections between them. We study how these models scale and find that making weights sparser trades off capability for interpretability, and scaling model size improves the capability-interpretability frontier. However, scaling sparse models beyond tens of millions of nonzero parameters while preserving interpretability remains a challenge. In addition to training weight-sparse models de novo, we show preliminary results suggesting our method can also be adapted to explain existing dense models. Our work produces circuits that achieve an unprecedented level of human understandability and validates them with considerable rigor.

Persona Features Control Emergent Misalignment

Understanding how language models generalize behaviors from their training to a broader deployment distribution is an important problem in AI safety. Betley et al. discovered that fine-tuning GPT-4o on intentionally insecure code causes "emergent misalignment," where models give stereotypically malicious responses to unrelated prompts. We extend this work, demonstrating emergent misalignment across diverse conditions, including reinforcement learning on reasoning models, fine-tuning on various synthetic datasets, and in models without safety training. To investigate the mechanisms behind this generalized misalignment, we apply a "model diffing" approach using sparse autoencoders to compare internal model representations before and after fine-tuning. This approach reveals several "misaligned persona" features in activation space, including a toxic persona feature which most strongly controls emergent misalignment and can be used to predict whether a model will exhibit such behavior. Additionally, we investigate mitigation strategies, discovering that fine-tuning an emergently misaligned model on just a few hundred benign samples efficiently restores alignment.

Investigating task-specific prompts and sparse autoencoders for activation monitoring

Language models can behave in unexpected and unsafe ways, and so it is valuable to monitor their outputs. Internal activations of language models encode additional information that could be useful for this. The baseline approach for activation monitoring is some variation of linear probing on a particular layer: starting from a labeled dataset, train a logistic regression classifier on that layer's activations. Recent work has proposed several approaches which may improve on naive linear probing, by leveraging additional computation. One class of techniques, which we call "prompted probing," leverages test time computation to improve monitoring by (1) prompting the model with a description of the monitoring task, and (2) applying a learned linear probe to resulting activations. Another class of techniques uses computation at train time: training sparse autoencoders offline to identify an interpretable basis for the activations, and e.g. max-pooling activations across tokens using that basis before applying a linear probe. However, one can also prompt the model with a description of the monitoring task and use its output directly. We develop and test novel refinements of these methods and compare them against each other. We find asking the model zero-shot is a reasonable baseline when inference-time compute is not limited; however, activation probing methods can substantially outperform this baseline given sufficient training data. Specifically, we recommend prompted probing when inference-time compute is available, due to its superior data efficiency and good generalization performance. Alternatively, if inference-time compute is limited, we find SAE-based probing methods outperform raw activation probing.