Confidence Regulation Neurons in Language Models
This work addresses the problem of understanding uncertainty mechanisms in LLMs for researchers, but it is incremental as it builds on known components and introduces new ones without broad SOTA impact.
The study investigated how large language models regulate uncertainty in next-token predictions by analyzing entropy neurons and newly discovered token frequency neurons, showing that entropy neurons influence confidence by scaling logits via the residual stream and are present in models up to 7 billion parameters, while token frequency neurons adjust logits based on token frequency to shift output distributions.
Despite their widespread use, the mechanisms by which large language models (LLMs) represent and regulate uncertainty in next-token predictions remain largely unexplored. This study investigates two critical components believed to influence this uncertainty: the recently discovered entropy neurons and a new set of components that we term token frequency neurons. Entropy neurons are characterized by an unusually high weight norm and influence the final layer normalization (LayerNorm) scale to effectively scale down the logits. Our work shows that entropy neurons operate by writing onto an unembedding null space, allowing them to impact the residual stream norm with minimal direct effect on the logits themselves. We observe the presence of entropy neurons across a range of models, up to 7 billion parameters. On the other hand, token frequency neurons, which we discover and describe here for the first time, boost or suppress each token's logit proportionally to its log frequency, thereby shifting the output distribution towards or away from the unigram distribution. Finally, we present a detailed case study where entropy neurons actively manage confidence in the setting of induction, i.e. detecting and continuing repeated subsequences.