# Contrastive Error Attribution for Finetuned Language Models

arXiv: https://arxiv.org/abs/2212.10722 · HTML: https://mlsanity.com/paper/2212.10722/

## Key facts

- Cited by 12 papers in the map (data through 14 Oct 2026).
- Builds on 26 of its 36 references that are in the map.
- First posted on arXiv 21 Dec 2022.

## Builds on (most cited first, 26 of 26)

- [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://mlsanity.com/paper/1910.13461/) (2019)
- [Understanding Black-box Predictions via Influence Functions](https://mlsanity.com/paper/1703.04730/) (2017)
- [Don't Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarization](https://mlsanity.com/paper/1808.08745/) (2018)
- [On Faithfulness and Factuality in Abstractive Summarization](https://mlsanity.com/paper/2005.00661/) (2020)
- [The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networks](https://mlsanity.com/paper/1802.08232/) (2018)
- [BARTScore: Evaluating Generated Text as Text Generation](https://mlsanity.com/paper/2106.11520/) (2021)
- [Estimating Training Data Influence by Tracing Gradient Descent](https://mlsanity.com/paper/2002.08484/) (2020)
- [FEQA: A Question Answering Evaluation Framework for Faithfulness Assessment in Abstractive Summarization](https://mlsanity.com/paper/2005.03754/) (2020)
- [Understanding Factuality in Abstractive Summarization with FRANK: A Benchmark for Factuality Metrics](https://mlsanity.com/paper/2104.13346/) (2021)
- [Symbolic Knowledge Distillation: from General Language Models to Commonsense Models](https://mlsanity.com/paper/2110.07178/) (2021)
- [The E2E Dataset: New Challenges For End-to-End Generation](https://mlsanity.com/paper/1706.09254/) (2017)
- [Datamodels: Predicting Predictions from Training Data](https://mlsanity.com/paper/2202.00622/) (2022)
- [Representer Point Selection for Explaining Deep Neural Networks](https://mlsanity.com/paper/1811.09720/) (2018)
- [Explaining Black Box Predictions and Unveiling Data Artifacts through Influence Functions](https://mlsanity.com/paper/2005.06676/) (2020)
- [Faithful to the Original: Fact Aware Neural Abstractive Summarization](https://mlsanity.com/paper/1711.04434/) (2017)
- [Scaling Up Influence Functions](https://mlsanity.com/paper/2112.03052/) (2021)
- [Entity-level Factual Consistency of Abstractive Text Summarization](https://mlsanity.com/paper/2102.09130/) (2021)
- [Evaluating the State-of-the-Art of End-to-End Natural Language Generation: The E2E NLG Challenge](https://mlsanity.com/paper/1901.07931/) (2019)
- [Improved Natural Language Generation via Loss Truncation](https://mlsanity.com/paper/2004.14589/) (2020)
- [Data Cleansing for Models Trained with SGD](https://mlsanity.com/paper/1906.08473/) (2019)
- [Semantic Noise Matters for Neural Natural Language Generation](https://mlsanity.com/paper/1911.03905/) (2019)
- [Improving Factual Consistency of Abstractive Summarization via Question Answering](https://mlsanity.com/paper/2105.04623/) (2021)
- [RelatIF: Identifying Explanatory Training Examples via Relative Influence](https://mlsanity.com/paper/2003.11630/) (2020)
- [Explaining and Improving Model Behavior with k Nearest Neighbor Representations](https://mlsanity.com/paper/2010.09030/) (2020)
- [SynthBio: A Case Study in Human-AI Collaborative Curation of Text Datasets](https://mlsanity.com/paper/2111.06467/) (2021)
- [Mind The Facts: Knowledge-Boosted Coherent Abstractive Text Summarization](https://mlsanity.com/paper/2006.15435/) (2020)

## Built on by (most cited first, 12 of 12)

- [Training Data Attribution via Approximate Unrolled Differentiation](https://mlsanity.com/paper/2405.12186/) (2024)
- [Where Did It All Go Wrong? A Hierarchical Look into Multi-Agent Error Attribution](https://mlsanity.com/paper/2510.04886/) (2025)
- [Token-wise Influential Training Data Retrieval for Large Language Models](https://mlsanity.com/paper/2405.11724/) (2024)
- [An Audit on the Perspectives and Challenges of Hallucinations in NLP](https://mlsanity.com/paper/2404.07461/) (2024)
- [Interpretation Meets Safety: A Survey on Interpretation Methods and Tools for Improving LLM Safety](https://mlsanity.com/paper/2506.05451/) (2025)
- [Enhancing Training Data Attribution with Representational Optimization](https://mlsanity.com/paper/2505.18513/) (2025)
- [Enhancing Training Data Attribution for Large Language Models with Fitting Error Consideration](https://mlsanity.com/paper/2410.01285/) (2024)
- [Error Norm Truncation: Robust Training in the Presence of Data Noise for Text Generation Models](https://mlsanity.com/paper/2310.00840/) (2023)
- [Sparse, Efficient and Explainable Data Attribution with DualXDA](https://mlsanity.com/paper/2402.12118/) (2024)
- [A Curious Case of Searching for the Correlation between Training Data and Adversarial Robustness of Transformer Textual Models](https://mlsanity.com/paper/2402.11469/) (2024)
- [Large Language Model Sourcing: A Survey](https://mlsanity.com/paper/2510.10161/) (2025)
- [On the Benefits of Fine-Grained Loss Truncation: A Case Study on Factuality in Summarization](https://mlsanity.com/paper/2403.05788/) (2024)

## Abstract (arXiv)

Recent work has identified noisy and misannotated data as a core cause of hallucinations and unfaithful outputs in Natural Language Generation (NLG) tasks. Consequently, identifying and removing these examples is a key open challenge in creating reliable NLG systems. In this work, we introduce a framework to identify and remove low-quality training instances that lead to undesirable outputs, such as faithfulness errors in text summarization. We show that existing approaches for error tracing, such as gradient-based influence measures, do not perform reliably for detecting faithfulness errors in NLG datasets. We overcome the drawbacks of existing error tracing methods through a new, contrast-based estimate that compares undesired generations to human-corrected outputs. Our proposed method can achieve a mean average precision of 0.93 at detecting known data errors across synthetic tasks with known ground truth, substantially outperforming existing approaches. Using this approach and re-training models on cleaned data leads to a 70% reduction in entity hallucinations on the NYT dataset and a 55% reduction in semantic errors on the E2E dataset.