# Same Function, Different Mechanism: Implementation-Local Geometry Beyond Accuracy

DevFeed: [Same Function, Different Mechanism: Implementation-Local Geometry Beyond Accuracy](<https://devfeed.tech/articles/same-function-different-mechanism-implementation-local-geometry-beyond-accuracy-114527.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.01328>)

Author: Miroslav L\\v{z}i\\v{c}a\\v{r} (Mereology Labs)

Published: 2026-11-03T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [AI Models](<https://devfeed.tech/topics/ai-models.md>)

Tags: [addition](<https://devfeed.tech/tags/addition.md>), [charts](<https://devfeed.tech/tags/charts.md>), [compiler](<https://devfeed.tech/tags/compiler.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [dimension](<https://devfeed.tech/tags/dimension.md>), [function](<https://devfeed.tech/tags/function.md>), [implementation](<https://devfeed.tech/tags/implementation.md>), [learning](<https://devfeed.tech/tags/learning.md>), [local](<https://devfeed.tech/tags/local.md>), [mechanism](<https://devfeed.tech/tags/mechanism.md>), [models](<https://devfeed.tech/tags/models.md>), [modular](<https://devfeed.tech/tags/modular.md>), [rank](<https://devfeed.tech/tags/rank.md>), [sharing](<https://devfeed.tech/tags/sharing.md>), [transformers](<https://devfeed.tech/tags/transformers.md>)

## AI overview

Models with identical predictions can have different implementation-local geometry: a lookup table permits independent case changes, while a shared rule couples them. The paper measures this using specified local logit-perturbation families around a fixed predictive distribution. For exact finite-group sharing, it gives active dimension as input-output orbit count minus input orbit count; softened modular addition yields ranks of p²(p−1) for lookup and p−1 for shared cyclic charts. The abstract also connects rank to local KL curvature via the Fisher pullback, discusses approximate-sharing spectral tails and higher-order learning coefficients, and reports measurements on trained modular-addition models and controlled compiled transformers, with negative controls where discrimination fails. The framework evaluates specified mechanism hypotheses; it neither uniquely infers mechanisms from predictions nor predicts training dynamics.

## Source excerpt

arXiv:2611.01328v1 Announce Type: new Abstract: Two models can make exactly the same predictions yet respond differently when their implementations are perturbed. A lookup table can change one case independently; a shared rule couples changes across cases. We compare these mechanisms through an implementation chart: a specified local family of logit perturbations around a fixed predictive distribution. Our main result counts the independent prediction changes allowed by exact finite-group sharing: the active dimension is the number of input-output orbits minus the number of input orbits. For softened modular addition, lookup and shared cyclic charts therefore have ranks $p^2(p-1)$ and $p-1$ despite identical predictions. The classical Fisher pullback connects these dimensions to local KL curvature; mechanism-preserving reparameterizations leave the rank unchanged. We also bound the spectral tail under approximate sharing and separate quadratic rank from higher-order learning coefficients. Trained modular-addition models verify that chart rank is measurable, while controlled compiled transformers distinguish compiler-supported intermediate charts from matched alternatives. Negative controls show where this discrimination fails. The framework evaluates specified mechanism hypotheses; it does not infer a unique mechanism from predictions or predict training dynamics.