Elias Kempf

Elias Kempf, Simon Schrodi, Max Argus et al.

The remarkable generalization performance of contrastive vision-language models like CLIP is often attributed to the diversity of their training distributions. However, key questions remain unanswered: Can CLIP generalize to an entirely unseen domain when trained on a diverse mixture of domains (domain generalization)? Can it generalize to unseen classes within partially seen domains (compositional generalization)? What factors affect such generalization? To answer these questions, we trained CLIP models on systematically constructed training distributions with controlled domain diversity and object class exposure. Our experiments show that domain diversity is essential for both domain and compositional generalization, yet compositional generalization can be surprisingly weaker than domain generalization when the training distribution contains a suboptimal subset of the test domain. Through data-centric and mechanistic analyses, we find that successful generalization requires the learning of sufficiently shared representations in intermediate layers and circuits.

Elias Kempf, Simon Schrodi, Bartosz Cywiński et al.

Standard LLM evaluations only test capabilities or dispositions that evaluators designed them for, missing unexpected differences such as behavioral shifts between model revisions or emergent misaligned tendencies. Model diffing addresses this limitation by automatically surfacing systematic behavioral differences. Recent approaches include LLM-based methods that generate natural language descriptions and sparse autoencoder (SAE)-based methods that identify interpretable features. However, no systematic comparison of these approaches exists nor are there established evaluation criteria. We address this gap by proposing evaluation metrics for key desiderata (generalization, interestingness, and abstraction level) and use these to compare existing methods. Our results show that an improved LLM-based baseline performs comparably to the SAE-based method while typically surfacing more abstract behavioral differences.

Simon Schrodi, Elias Kempf, Fazl Barez et al.

Language models can transfer hidden biases during distillation. For example, a teacher that "likes owls" can make its student "like owls" too, even when the training data consists only of lists of numbers. This surprising phenomenon is called subliminal learning. Subliminal learning can be expected under soft distillation, where the student is trained on the teacher's full next-token distribution. But the fact that this also occurs under hard distillation-where the student only sees sampled tokens-raises a deeper question: when and how does subliminal learning actually occur? We answer this question through controlled experiments and mechanistic analysis. Our results show that subliminal learning does not need (global) token entanglement or logit leakage. Instead, it comes down to a small set of divergence tokens-rare cases where teachers with different biases would predict different tokens. Masking out these tokens mostly removes the hidden bias transfer. Mechanistically, divergence tokens reveal that early layers are critical. Surprisingly, finetuning even a single such early layer is sufficient for subliminal learning. Finally, we find that subliminal learning is fragile. Even small changes, like paraphrasing prompts, are usually sufficient to suppress it.