claim

active

claim:language-models-prefer-reusing-generic-arithmetic-mechanisms-over-learning-task-specific-modular-computations-even-when-task-specific-geometry-exists

Language models prefer reusing generic arithmetic mechanisms over learning task-specific modular computations even when task-specific geometry exists

Broader interpretive claim about LM learning bias inferred from the findings

Source paper

extracted_from

Arithmetic in the Wild: Llama uses Base-10 Addition to Reason About Cyclic Concepts

(2026) · Sheridan Feucht · Tal Haklay · Usha Bhalla · Daniel Wurgaft +8

Neighborhood — ranked by edge-count

Findings (1)

finding

Llama-3.1-8B reuses a single generic addition mechanism across all cyclic tasks independently of concept-specific geometry
supports
Key mechanistic finding showing task-agnostic reuse of arithmetic circuitry

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.

It's tricky, because for a typical language model the entity is sort of tricameral: the base simulator, the simulated simulator, and the simulated awareness.quote0.800
Antra's earlier definitive statement of the tricameral model.
An interplay between causal abstraction and feature geometry deepens mechanistic understanding of language modelsclaim0.798
Methodological claim about the scientific value of combining causal abstraction with representational geometry analysis
There are fewer representations competent for N tasks than M<N tasks, so training more general models should yield fewer possible solutionshypothesis0.792
Selective pressure toward convergence via task generality
Software implementations for all of the models/behaviours presented are common for n = 2, and can be made very efficient for α_i that map many objects onto a much smaller set of object families.claim0.789
Claim about current practical feasibility and efficiency of 2-way associative implementations.
Language models are some of the most remarkable computer programs in existence.quote0.789
Opening sentence setting the stage for the importance of interpretability.
Representation and computation can diverge; cyclic geometry is representational invariant while operations use generic substrate.claim0.788
Language models implement algorithms humans have tried and failed to write by hand for decadesclaim0.785
Opening interpretive claim about the remarkable nature of language models.
Analogous alignment between representation manifold and behavior manifold is found across months, letters, ages, and synthetic in-context learning tasks in language models.finding0.782
Generalization finding from the full paper extending beyond days-of-week to other structured concepts.