Modular addition

The mathematically natural computation for cyclic concepts (e.g., addition mod 12 for months), which the paper shows Llama does NOT directly implement

Neighborhood — ranked by edge-count

Concepts (1)

concept

Base-10 addition
contradicts
The generic addition mechanism that Llama-3.1-8B actually uses to compute sums before mapping back to cyclic concept space

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Activation Additionmethod0.801
Intervention method that adds a learned direction vector to residual stream activations to steer model behavior
Modularityconcept0.784
Property of developmental systems where functions are encapsulated in modules with simple triggers, enhancing evolvability.
Functional Modularityconcept0.740
The principle that specific computations are localized to specific model components, evidenced by Style Modulation Heads
Activation Addition (ActAdd)framework0.738
Steering method deriving vectors from contrastive prompt pairs and adding to first-token activations.
Generic addition mechanismconcept0.729
The core finding: Llama-3.1-8B reuses one addition mechanism across all cyclic tasks rather than learning task-specific modular arithmetic
introducesconcept0.711
Does Llama compute modular addition or base-10 addition for cyclic tasks?question0.711
The specific computational question the paper resolves empirically
integralmethod0.708
Primitive signal transformer computing integration of input signal over time; enables velocity-to-position conversion in paddleball.