Stefan Heimersheim

Concept Influence attributes model behaviors to semantic directions (like linear probes or sparse autoencoder features) rather than individual test examples, improving identification of the training data that disproportionately drive unintended behaviors. Simple first-order approximations match or outperform standard influence functions while achieving over 20× computational speedups, though they degrade under significant distribution shifts.

Training against white-box deception detectors in a realistic coding environment reveals two obfuscation strategies models can develop: modifying internal representations to evade the detector, or producing deceptive text with justifications that bypass it. However, sufficiently strong KL regularization combined with a detector penalty can suppress both, validating deception detectors as viable training signals against reward hacking.

We show that the apparent performance gains of the Compressed Computation toy model arise from unintended label mixing via a noisy residual stream, not from computation in superposition.

We show that all LayerNorm layers can be removed from GPT-2 models via fine-tuning with minimal performance loss, making inference-time LayerNorm unnecessary.

Misalignment cases might be rare but critical. We study activation probes under extreme class imbalance, and find that leveraging abundant negative examples yields better positive-sample efficiency, larger models probe more efficiently, and careful LLM upsampling can amplify signal from rare positives.

We systematize the mechanistic interpretability process into 3 iterative steps, then proceed to automate one of them: circuit discovery. Two of the algorithms presented automatically discover interpretability results previously established by human inspection.