arXiv:2607. 22367v1 Announce Type: new Abstract: Feature-attribution methods assign scores relating input variables to a model's output, but do not by themselves characterize how explicitly defined interaction operators compose across its intermediate layers.
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...
The paper introduces an influence score that measures how much each attention head contributes to classification decisions in Transformer models, specifically for prompt injection detection. The score blends directional effects on logits with structural impact within the residual stream, allowing analysis at head, layer, and network scales. When applied to a DeBERTa model, the framework uncovers different decision patterns for correct versus incorrect predictions, offering a balanced approach between detailed circuit analysis and global output methods.
arXiv:2609.39188v1 Announce Type: new Abstract: Developmental psychology holds that certain priors are given to infants prior to experience rather than induced from data, and that the influence of su...
arXiv:2606. 03885v1 Announce Type: new Abstract: Feature attribution methods explain predictions by assigning importance scores to input features.
arXiv:2601. 17952v2 Announce Type: replace-cross Abstract: Interpretability remains a key challenge for deploying language models (LM) in clinical settings such as progression diagnosis of Alzheimer disease, where early and trustworthy predictions are essential.
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:2609.36448v1 Announce Type: new Abstract: Transformers have demonstrated remarkable in-context learning (ICL) capabilities, enabling them to perform new tasks without additional fine-tuning. Ho...
arXiv:2608. 12447v1 Announce Type: new Abstract: Trained transformer models develop privileged bases: coordinate axes whose statistics differ from the rest of the residual stream.
The projection of queries and keys are central to the attention mechanism in Transformer architectures. While they are mathematically symmetric, they play different roles in attention mechanisms.
arXiv:2606. 29256v1 Announce Type: cross Abstract: In recent years, models based on the Transformer architecture have seen widespread applications and have become one of the core tools in the field of deep learning.
The paper introduces a coupled query‑key transformation that jointly evolves queries and keys via an invertible coupling before the standard dot‑product scoring in attention mechanisms. Implemented as a lightweight alternating affine map, the coupling is added on top of existing attention methods and preserves the original softmax and architecture. Experiments on WikiText‑103 show that coupling improves performance when combined with Differential Attention, query‑key normalization, and Multi‑Token Attention, especially at larger model scales, while its standalone benefit diminishes with size.
arXiv:2608. 09417v1 Announce Type: new Abstract: Deep decoder-only Transformers often replace the original Post-Norm architecture with Pre-Norm variants because Post-Norm training is highly sensitive to warmup and learning rate under conventional initialization schemes.