Base-10 addition

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

Neighborhood — ranked by edge-count

concept

Base-10 arithmetic
related_to
High pretraining exposure numeral system used in E2.
Generic addition mechanism
implements
The core finding: Llama-3.1-8B reuses one addition mechanism across all cyclic tasks rather than learning task-specific modular arithmetic
Two-stage cyclic computation
uses
Llama's discovered algorithm: first compute integer sum via base-10 addition, then map the sum back to cyclic concept space
Modular addition
contradicts
The mathematically natural computation for cyclic concepts (e.g., addition mod 12 for months), which the paper shows Llama does NOT directly implement

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.

Base-9 arithmeticconcept0.833
Low pretraining exposure numeral system used in E2.
Activation Additionmethod0.809
Intervention method that adds a learned direction vector to residual stream activations to steer model behavior
Base-8 arithmeticconcept0.803
Moderate pretraining exposure numeral system used in E2.
Base Capabilityconcept0.761
A model's task-solving capability without harness evolution, used as baseline for comparing evolution gains
Base-10 Addition Mechanism in Llamaconcept0.755
Activation Addition (ActAdd)framework0.748
Steering method deriving vectors from contrastive prompt pairs and adding to first-token activations.
k50 for base-10 two-digit addition: 0.28 ± 0.05 shotsfinding0.737
Shot midpoint from logistic fit over 10 runs.
Reflection Enhancement via Activation Additionmethod0.734
Adding steering vector in forward direction to push model activations toward stronger reflective behavior.