Tag
This research paper explores emotion-sensitive neurons in multimodal foundation models, revealing shared affective mechanisms between speech and facial emotion recognition through causal interventions and cross-modal analysis.
This paper investigates when large language models develop domain-specific parametric shells (causally necessary neuron populations), finding that modular training data at the token level (e.g., languages, code) produces functional shells, while academic subject domains do not, despite being linearly decodable.
A first neuron-level interpretability study of how large audio-language models encode multilingual emotion, introducing Consistency-Regularized Fusion to identify Multilingual Emotion Neurons across 12 languages and showing cross-lingual transfer benefits.
This paper investigates whether arithmetic heuristic neurons in LLMs are form-invariant across symbolic arithmetic, natural language word problems, and Python code. Using activation patching, they find a shared circuit of neurons that is necessary and sufficient for arithmetic computation, and that cross-format failures arise from activation states rather than distinct circuits.
This paper introduces a technique to disentangle a single convolutional neuron in Inceptionv1 by analyzing Hadamard products, revealing clean monosemantic clusters (cars, cats, dogs) and also low-valued clusters (letters, faces) with distributed weights as evidence of gradient descent behavior.
This paper introduces YFPO, a neuron-guided preference optimization framework that uses internal activation signals to improve mathematical reasoning in large language models.
This research demonstrates that safety alignment in large language models can be bypassed by targeting single neurons responsible for refusal, revealing that safety mechanisms are not robustly distributed but mediated by individual neurons.
This paper introduces IRiS, a training-free framework for situational personality steering in LLMs that moves beyond static persona modeling by identifying and leveraging situation-dependent persona neurons. The approach demonstrates that LLM behavior varies contextually and proposes neuron-based identification, retrieval, and weighted steering methods validated on PersonalityBench and a new SPBench benchmark.
MNAFT (Modality Neuron-Aware Fine-Tuning) is a novel approach that selectively updates language-specific and language-agnostic neurons in multimodal large language models to improve image translation while preserving pre-trained knowledge. The method outperforms state-of-the-art image translation techniques including cascaded models and standard fine-tuning approaches.