mlsanityData to 7 Oct 2026

Incorporating Attribution Importance for Improving Faithfulness Metrics

arXiv 2305.10496 ↗First posted 17 May 2023cs.CL cs.AI cs.LGcited by 14

On the map: LLM reasoning → Test-time compute and efficient thinking

Abstract (arXiv)

Feature attribution methods (FAs) are popular approaches for providing insights into the model reasoning process of making predictions. The more faithful a FA is, the more accurately it reflects which parts of the input are more important for the prediction. Widely used faithfulness metrics, such as sufficiency and comprehensiveness use a hard erasure criterion, i.e. entirely removing or retaining the top most important tokens ranked by a given FA and observing the changes in predictive likelihood. However, this hard criterion ignores the importance of each individual token, treating them all equally for computing sufficiency and comprehensiveness. In this paper, we propose a simple yet effective soft erasure criterion. Instead of entirely removing or retaining tokens from the input, we randomly mask parts of the token vector representations proportionately to their FA importance. Extensive experiments across various natural language processing tasks and different FAs show that our soft-sufficiency and soft-comprehensiveness metrics consistently prefer more faithful explanations compared to hard sufficiency and comprehensiveness. Our code: https://github.com/casszhao/SoftFaith

Key facts. Cited by 14 papers in the map (data through 7 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 36 of its 48 references that are in the map. First posted on arXiv 17 May 2023.

Builds on

36 of 48 references.

References · 36

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
  2. 2019RoBERTa: A Robustly Optimized BERT Pretraining Approachcited by 13,266
  3. 2016"Why Should I Trust You?": Explaining the Predictions of Any Classifiercited by 5,040
  4. 2017Axiomatic Attribution for Deep Networkscited by 3,350
  5. 2013Visualizing and Understanding Convolutional Networkscited by 2,278
  6. 2017Learning Important Features Through Propagating Activation Differencescited by 1,454
  7. 2019SciBERT: A Pretrained Language Model for Scientific Textcited by 1,404
  8. 2019Attention is not Explanationcited by 916
  9. 2020Towards Faithfully Interpretable NLP Systems: How should we define and evaluate faithfulness?cited by 590
  10. 2019ERASER: A Benchmark to Evaluate Rationalized NLP Modelscited by 523
  11. 2019Is Attention Interpretable?cited by 387
  12. 2018On the Robustness of Interpretability Methodscited by 310
24 more
  1. 2019One Explanation Does Not Fit All: A Toolkit and Taxonomy of AI Explainability Techniquescited by 164
  2. 2020A Diagnostic Study of Explainability Techniques for Text Classificationcited by 158
  3. 2022Towards Faithful Model Explanation in NLP: A Surveycited by 129
  4. 2020The elephant in the interpretability room: Why use attention as explanation when we have saliency methods?cited by 122
  5. 2020Learning to Faithfully Rationalize by Constructioncited by 113
  6. 2021Paragraph-level Rationale Extraction through Regularization: A case study on European Court of Human Rights Casescited by 77
  7. 2019Inferring Which Medical Treatments Work from Reports of Clinical Trialscited by 63
  8. 2016Investigating the influence of noise and distractors on the interpretation of neural networkscited by 54
  9. 2016"What is Relevant in a Text Document?": An Interpretable Machine Learning Approachcited by 53
  10. 2022A Comparative Study of Faithfulness Metrics for Model Interpretability Methodscited by 38
  11. 2020Evaluating and Characterizing Human Rationalescited by 36
  12. 2021On the Sensitivity and Stability of Model Interpretations in NLPcited by 30
  13. 2021Perturbing Inputs for Fragile Interpretations in Deep Natural Language Processingcited by 30
  14. 2021Improving the Faithfulness of Attention-based Explanations with Task-specific Information for Text Classificationcited by 29
  15. 2021Translation Error Detection as Rationale Extractioncited by 22
  16. 2022An Empirical Study on Explanations in Out-of-Domain Settingscited by 21
  17. 2021Connecting Attributions and QA Model Behavior on Realistic Counterfactualscited by 20
  18. 2020Staying True to Your Word: (How) Can Attention Become Explanation?cited by 18
  19. 2021Logic Traps in Evaluating Attribution Scorescited by 18
  20. 2021On the Lack of Robust Interpretability of Neural Text Classifierscited by 17
  21. 2021Enjoy the Salience: Towards Better Transformer-based Faithful Explanations with Word Saliencecited by 16
  22. 2022On the Impact of Temporal Concept Drift on Model Explanationscited by 14
  23. 2018Word Embedding Perturbation for Sentence Classificationcited by 11
  24. 2021Flexible Instance-Specific Rationalization of NLP Modelscited by 10

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2023: 1; 2024: 6; 2025: 6; 2026: 1.

14 papers.

Sort:

  1. 2023Investigating Hallucinations in Pruned Large Language Models for Abstractive Summarizationcited by 16
  2. 2025Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associationscited by 14
  3. 2024Latent Concept-based Explanation of NLP Modelscited by 13
  4. 2024ReAGent: A Model-agnostic Feature Attribution Method for Generative Language Modelscited by 12
  5. 2024Normalized AOPC: Fixing Misleading Faithfulness Metrics for Feature Attribution Explainabilitycited by 10
  6. 2024Comparing Explanation Faithfulness between Multilingual and Monolingual Fine-tuned Language Modelscited by 6
  7. 2025Noiser: Bounded Input Perturbations for Attributing Large Language Modelscited by 4
  8. 2025Gender Bias in Explainability: Investigating Performance Disparity in Post-hoc Methodscited by 3
  9. 2024DLBacktrace: A Model Agnostic Explainability for any Deep Learning Modelscited by 2
  10. 2025EvalxNLP: A Framework for Benchmarking Post-Hoc Explainability Methods on NLP Modelscited by 2
  11. 2025Cross-Layer Discrete Concept Discovery for Interpreting Language Modelscited by 1
  12. 2025Support-Contra Asymmetry in LLM Explanationscited by 1
2 more
  1. 2026Hessian-Enhanced Token Attribution (HETA): Interpreting Autoregressive LLMscited by 1
  2. 2024ExU: AI Models for Examining Multilingual Disinformation Narratives and Understanding their Spreadcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025A Causal Lens for Evaluating Faithfulness Metrics
  2. 2024The Probabilities Also Matter: A More Faithful Metric for Faithfulness of Free-Text Explanations in Large Language Models
  3. 2026Assessing Alignment and Stability of Feature Importance Explanations via Weight of Evidence
  4. 2023Benchmarking Faithfulness: Towards Accurate Natural Language Explanations in Vision-Language Tasks
  5. 2025Explanation Bias is a Product: Revealing the Hidden Lexical and Position Preferences in Post-Hoc Feature Attribution
  6. 2026Towards Long-Horizon Interpretability: Efficient and Faithful Multi-Token Attribution for Reasoning LLMs
  7. 2026Measuring Faithfulness Depends on How You Measure: Classifier Sensitivity in LLM Chain-of-Thought Evaluation
  8. 2024Are self-explanations from Large Language Models faithful?