GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions.
Unveiling T ransformers with LEGO : a synthetic reasoning task
5 Pith papers cite this work. Polarity classification is still indexing.
5
Pith papers citing it
representative citing papers
Massive activations are constant large values in LLMs that function as indispensable bias terms and concentrate attention probabilities on specific tokens.
Transformers reuse the same modular addition MLP for direct and variable-assigned inputs, with learning progressing through three phases that enable compositional generalization to unseen combinations.
BERT learns shortcut solutions that impair generalization and forward transfer in continual LEGO, while ALBERT learns loop-like solutions for better performance, yet both fail at cross-experience composition, with ALBERT rescued by mixed-data training.
Static depth-staggered Fibonacci sparse attention improves perplexity over fixed/learned variants and extrapolates to 4x context while dense attention fails.
citing papers explorer
-
Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small
GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions.
-
Massive Activations in Large Language Models
Massive activations are constant large values in LLMs that function as indispensable bias terms and concentrate attention probabilities on specific tokens.
-
Assign and Add: A Mechanistic Study of Compositional Arithmetic
Transformers reuse the same modular addition MLP for direct and variable-assigned inputs, with learning progressing through three phases that enable compositional generalization to unseen combinations.
-
Shortcut Solutions Learned by Transformers Impair Continual Compositional Reasoning
BERT learns shortcut solutions that impair generalization and forward transfer in continual LEGO, while ALBERT learns loop-like solutions for better performance, yet both fail at cross-experience composition, with ALBERT rescued by mixed-data training.
-
Depth-Staggered Fibonacci Spacing for Sparse Attention: Static Schedules Beat Learned Dilation and Extrapolate Where Dense Attention Fails
Static depth-staggered Fibonacci sparse attention improves perplexity over fixed/learned variants and extrapolates to 4x context while dense attention fails.