Paper · arXiv 2307.15771
The Hydra Effect: Emergent Self-repair in Language Model Computations
Thomas McGrath, Matthew Rahtz, Janos Kramar, Vladimir Mikulik, Shane Legg
Analysis reveals adaptive computation and downregulation mechanisms in language models, showing loose coupling among layers even without dropout.
Abstract
We investigate the internal structure of language model computations using causal analysis and demonstrate two motifs: (1) a form of adaptive computation where ablations of one attention layer of a language model cause another layer to compensate (which we term the Hydra effect) and (2) a counterbalancing function of late MLP layers that act to downregulate the maximum-likelihood token. Our ablation studies demonstrate that language model layers are typically relatively loosely coupled (ablations to one layer only affect a small number of downstream layers). Surprisingly, these effects occur even in language models trained without any form of dropout. We analyse these effects in the context of factual recall and consider their implications for circuit-level attribution in language models.