Luca Baroni

Mohammed Abu Baker, Luca Baroni, Dan Wilhelm

Finetuning can significantly modify the behavior of large language models, including introducing harmful or unsafe behaviors. To study these risks, researchers develop model organisms: models finetuned to exhibit specific known behaviors for controlled experimentation. Identifying these behaviors remains challenging. We show that a simple perplexity-based method can surface finetuning objectives from model organisms by leveraging their tendency to overgeneralize their finetuned behaviors beyond the intended context. First, we generate diverse completions from the finetuned model using short random prefills drawn from general corpora. Second, we rank completions by decreasing perplexity gap between reference and finetuned models. The top-ranked completions often reveal the finetuning objectives, without requiring model internals or prior assumptions about the behavior. We evaluate this on a diverse set of model organisms (N=76, 0.5 to 70B parameters), including backdoored models, models finetuned to internalize false facts via synthetic document finetuning, adversarially trained models with hidden concerning behaviors, and models exhibiting emergent misalignment. For the vast majority of model organisms tested, the method surfaces completions revealing finetuning objectives within the top-ranked results, with models trained via synthetic document finetuning or to produce exact phrases being particularly susceptible. We further show that the technique can be effective even without access to the exact pre-finetuning checkpoint: trusted reference models from different families can serve as effective substitutes. As the method requires only next-token probabilities from the finetuned model, it is compatible with API-gated models that expose token logprobs.

Luca Baroni, Galvin Khara, Joachim Schaeffer et al.

Layer-wise normalization (LN) is an essential component of virtually all transformer-based large language models. While its effects on training stability are well documented, its role at inference time is poorly understood. Additionally, LN layers hinder mechanistic interpretability by introducing additional nonlinearities and increasing the interconnectedness of individual model components. Here, we show that all LN layers can be removed from every GPT-2 model with only a small increase in validation loss (e.g. +0.03 cross-entropy loss for GPT-2 XL). Thus, LN cannot play a substantial role in language modeling. We find that the amount of fine-tuning data needed for LN removal grows sublinearly with model parameters, suggesting scaling to larger models is feasible. We release a suite of LN-free GPT-2 models on Hugging Face. Furthermore, we test interpretability techniques on LN-free models. Direct logit attribution now gives the exact direct effect of individual components, while the accuracy of attribution patching does not significantly improve. We also confirm that GPT-2's "confidence neurons" are inactive in the LN-free models. Our work clarifies the role of LN layers in language modeling, showing that GPT-2-class models can function without LN layers. We hope that our LN-free analogs of the GPT-2 family of models will enable more precise interpretability research and improve our understanding of language models.

Jan Sobotka, Luca Baroni, Ján Antolík

Decoding visual stimuli from neural population activity is crucial for understanding the brain and for applications in brain-machine interfaces. However, such biological data is often scarce, particularly in primates or humans, where high-throughput recording techniques, such as two-photon imaging, remain challenging or impossible to apply. This, in turn, poses a challenge for deep learning decoding techniques. To overcome this, we introduce MEIcoder, a biologically informed decoding method that leverages neuron-specific most exciting inputs (MEIs), a structural similarity index measure loss, and adversarial training. MEIcoder achieves state-of-the-art performance in reconstructing visual stimuli from single-cell activity in primary visual cortex (V1), especially excelling on small datasets with fewer recorded neurons. Using ablation studies, we demonstrate that MEIs are the main drivers of the performance, and in scaling experiments, we show that MEIcoder can reconstruct high-fidelity natural-looking images from as few as 1,000-2,500 neurons and less than 1,000 training data points. We also propose a unified benchmark with over 160,000 samples to foster future research. Our results demonstrate the feasibility of reliable decoding in early visual system and provide practical insights for neuroscience and neuroengineering applications.