Challenge: Existing methods for interpreting the underlying dynamics of Transformers have been criticized for their lack of reliability.
Approach: They propose a token attribution analysis method that incorporates all components in the encoder block and aggregates this across layers.
Outcome: The proposed method significantly outperforms existing methods on saliency scores and correlation with gradient-based salience scores.

Similar Papers

Inceptive Transformers: Enhancing Contextual Representations through Multi-Scale Feature Learning Across Domains and Languages (2025.emnlp-main)

Copied to clipboard

Challenge: Encoder transformer models compress information from all tokens into a single [CLS] token to represent global context.
Approach: They propose a 1-D convolution module that augments token representations with multi-scale local features to improve performance.
Outcome: Experiments on five diverse tasks show that the proposed framework outperforms baseline models by 1% to 14% while maintaining efficiency.
Token-Wise Kernels (TWiKers) for Vicinity-Aware Attention in Transformers (2026.findings-eacl)

Copied to clipboard

Challenge: Token-Wise Kernels (TWiKers) are a novel enhancement to transformers that learn token-specific convolutional kernels applied to the keys or values.
Approach: They propose a transformer enhancement that learns token-specific convolutional kernels applied to the keys or values.
Outcome: The proposed transformers learn token-specific convolutional kernels applied to the keys or values . the results show that content words retain self-focus while function words shift attention toward their neighbors .
DecoderLens: Layerwise Interpretation of Encoder-Decoder Transformers (2024.findings-naacl)

Copied to clipboard

Challenge: Existing interpretability methods have been proposed to interpret the inner workings of Transformer models at different levels of precision and complexity.
Approach: They propose a method to analyze encoder-decoder Transformers by using the decoder module Model Output encoder to cross-attend representations of intermediate encoder activations instead of using the default output.
Outcome: The proposed method maps uninterpretable representations to human-interpreted sequences of words or symbols, shedding new light on the information flow in this popular but understudied class of models.
Measuring the Mixing of Contextual Information in the Transformer (2022.emnlp-main)

Copied to clipboard

Challenge: Experimentally, we show that ALTI provides more faithful explanations and increased robustness than gradient-based methods.
Approach: They propose to measure token-to-token interactions within each layer and then use them to aggregate model predictions.
Outcome: The proposed method provides more faithful explanations and increased robustness than gradient-based methods.
An Attribution Method for Siamese Encoders (2023.emnlp-main)

Copied to clipboard

Challenge: Despite the success of Siamese encoder models, little is known about the aspects of inputs they base their decisions on.
Approach: They propose a local attribution method for Siamese encoders by generalizing the principle of integrated gradients to models with multiple inputs.
Outcome: The proposed method can be reduced to a token–token matrix and account for the model’s full computation graph and is guaranteed to converge to the actual prediction.
FNet: Mixing Tokens with Fourier Transforms (2022.naacl-main)

Copied to clipboard

Challenge: Using simple linear transformations, Transformer encoders can be sped up with limited accuracy costs by replacing the self-attention sublayers with simple linear mixing mechanisms.
Approach: They propose to replace the self-attention sublayer with a linear transformation that "mixes" input tokens.
Outcome: The proposed model outperforms the “efficient Transformers” on the GLUE benchmark at longer input lengths and on smaller models with a light memory footprint.
DecompX: Explaining Transformers Decisions by Propagating Token Decomposition (2023.acl-long)

Copied to clipboard

Challenge: Existing vector-based explanation methods for Transformer-based models are limited in their ability to explain the decisions of multiple layers.
Approach: They propose a vector-based explanation method based on the construction of decomposed token representations and their successive propagation throughout the model without mixing them in between layers.
Outcome: The proposed method outperforms existing vector-based and gradient-based methods on transformer-based models by a wide margin.
Incorporating Residual and Normalization Layers into Analysis of Masked Language Models (2021.emnlp-main)

Copied to clipboard

Challenge: Transformer architecture is composed of multi-head attention, which has been extensively analyzed.
Approach: They extended the scope of the analysis of Transformers from solely the attention patterns to the whole attention block, i.e., multi-head attention, residual connection, and layer normalization.
Outcome: The proposed method incorporates the whole attention block, i.e., multi-head attention, residual connection, and layer normalization into the analysis.
Telling BERT’s Full Story: from Local Attention to Global Aggregation (2021.eacl-main)

Copied to clipboard

Challenge: Recent work discouraging the use of attention distributions for explaining a model’s behaviour suggests that attention distribution can provide insights into local behaviour of attention heads.
Approach: They propose a distinction between local patterns revealed by attention and global patterns that refer back to the input and analyze BERT from both angles.
Outcome: The proposed model can explain local behaviour of attention heads by comparing local and global patterns from both angles.
LAIT: Efficient Multi-Segment Encoding in Transformers with Layer-Adjustable Interaction (2023.acl-long)

Copied to clipboard

Challenge: In many NLP tasks, the input text can be seen as a sequence of related segments.
Approach: They propose a layer-adjustable interactions framework that contextualizes token representations by attending to all other tokens at each layer, leading to quadratic increase in compute effort with the input length.
Outcome: The proposed model reduces 30-50% of attention FLOPs while maintaining high accuracy.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations