arXiv:2607. 17696v1 Announce Type: cross Abstract: We develop an adjoint-sensitivity framework for positional influence in causal residual Transformers and separate unconditional analytic results from conditional boundary-shape conclusions.
arXiv:2609.37717v1 Announce Type: new Abstract: Decoder-only transformers are trained only through a terminal next-token prediction loss, yet this loss constrains every intermediate hidden state thro...
arXiv:2608.30720v1 Announce Type: new Abstract: Representational similarity is foundational to analyses of deep networks, yet distances between point-valued representations are not intrinsically tied...
arXiv:2602. 18849v2 Announce Type: replace-cross Abstract: We develop a sensitivity analysis for transformer attention in a geometry aligned with tokenwise computation.
The paper proposes that two architectural assumptions—(1) attention and MLPs share a key‑value form <phi(S)>U, and (2) components read from an additive residual stream—are sufficient to answer three interpretability questions: component interaction, information routing, and token attribution. By treating these selections as a computational graph, the authors develop Unpack, a backward attribution method that validates interaction scores, recovered routes, and token attribution against established tests across models ranging from 160M to 6.9B parameters. The study also shows that contribution and causal effect can differ, with a recognizable signature in how components change when a task is removed.
arXiv:2605. 25225v2 Announce Type: replace-cross Abstract: Mechanistic interpretability often studies Transformer behavior by intervening on internal activations through activation patching, causal tracing, path patching, and steering directions.
arXiv:2608. 03629v1 Announce Type: new Abstract: A companion paper studies when activation patching and weight-space ablation agree, inside an idealized model where a conditional computation is carried additively through a residual stream.
We develop a finite-width geometric framework describing how learned feature geometries are organized, transported, and selectively aligned in deep neural networks. Incompatibility among weight-genera...
The paper introduces a finite‑width geometric framework that explains how learned feature geometries are organized, transported, and selectively aligned in deep neural networks. It quantifies incompatibilities among weight‑generated covariance, gates, and backward sensitivities using three families of commutators, and provides exact layerwise identities that decompose these commutators into sources such as downstream transport, adjacent‑layer imbalance, and nonlinear gate‑covariance interactions. The study demonstrates that spectral alignment is a layer‑ and scale‑dependent compatibility phenomenon governed by transport, interaction, cancellation, and possible damping, rather than a universal consequence of training.
arXiv:2606. 26538v1 Announce Type: cross Abstract: Deep Transformers are composed of uniformly stacked residual blocks, yet their deepest layers often add little value.
arXiv:2608. 03620v1 Announce Type: cross Abstract: Activation patching and weight-space ablation both claim a component is causally responsible for a behavior, yet they act on different objects: one forward pass versus the parameters behind every forward pass.
arXiv:2608. 11797v1 Announce Type: new Abstract: Model merging by task arithmetic works until it doesn't, and the field diagnoses why with magnitudes: layerwise representation bias, deviations from cross-task linearity, parameter overlap.
arXiv:2606. 29693v1 Announce Type: new Abstract: We ask a simple question about decoder-only transformers: \emph{between which two layers is the probability of a predicted token actually produced?