CURE: Controlled Unlearning for Robust Embeddings — Mitigating Conceptual Shortcuts in Pre-Trained Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Pre-trained language models are susceptible to spurious, concept-driven correlations that impair robustness and fairness. |
| Approach: | They propose a framework that disentangles and suppresses conceptual shortcuts while preserving essential content information. |
| Outcome: | The proposed framework improves on IMDB and Yelp datasets with minimal computational overhead. |
Similar Papers
When and Why Does Bias Mitigation Work? (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Neural models exploit shallow surface features to perform language understanding tasks, rather than learning the deeper language understanding and reasoning skills that practitioners desire. |
| Approach: | They propose to use model debiasing techniques to pressure models away from spurious features and to use them to learn useful representations instead. |
| Outcome: | The proposed methods increase models' reliance on hidden biases instead of learning robust features that help them solve a task. |
Self-Diagnosis and Self-Debiasing: A Proposal for Reducing Corpus-Based Bias in NLP (2021.tacl-1)
Copied to clipboard
| Challenge: | Pretrained language models pick up and reproduce undesirable biases when trained on large, unfiltered crawls from the Internet. |
| Approach: | They propose a decoding algorithm that, given only a textual description of the undesired behavior, reduces the probability of a language model producing problematic text. |
| Outcome: | The proposed approach reduces the probability of a language model producing problematic text by giving only a textual description of the undesired behavior. |
Mitigating Biases in Language Models via Bias Unlearning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent debiasing approaches target different demographic groups, harming fairness and discrimination. |
| Approach: | They propose a model debiasing framework which targets stereotypes by unlearning stereotype forgetting and anti-stereotype retention. |
| Outcome: | The proposed framework outperforms existing methods in mitigating bias while retaining language modeling capabilities. |
Fixing Model Bugs with Natural Language Patches (2022.emnlp-main)
Copied to clipboard
| Challenge: | a growing body of research focused on using language to give instructions, supervision and even inductive biases to models instead of relying exclusively on labeled examples. |
| Approach: | They explore natural language patches that provide corrective feedback at the right level of abstraction. |
| Outcome: | The proposed model improves accuracy on real data by 1–4 accuracy points on different slices of a sentiment analysis dataset and F1 by 7 points on a relation extraction dataset. |
Mitigating Shortcut Learning via Smart Data Augmentation based on Large Language Model (2025.coling-main)
Copied to clipboard
| Challenge: | Existing methods to improve shortcut learning performance are limited by manual definition of shortcuts and inherent confirmation bias during model training. |
| Approach: | They propose a method of Smart Data Augmentation based on Large Language Models to identify shortcuts and generate their anti-shortcut counterparts. |
| Outcome: | The proposed method shows an improvement of 5.61% across various natural language processing tasks. |
A Prompt Array Keeps the Bias Away: Debiasing Vision-Language Models with Adversarial Learning (2022.aacl-main)
Copied to clipboard
| Challenge: | Large-scale, pretrained vision-language models are growing in popularity due to impressive performance on downstream tasks with minimal finetuning. |
| Approach: | They propose to apply ranking metrics to image-text representations to investigate bias measures and debiasing methods to reduce various bias measures. |
| Outcome: | The proposed model reduces bias measures with minimal degradation to image-text representations. |
Debiasing Masks: A New Framework for Shortcut Mitigation in NLU (2022.emnlp-main)
Copied to clipboard
| Challenge: | Debiasing language models from unwanted behaviors in natural language understanding datasets is a topic with increasing interest in the NLP community. |
| Approach: | They propose a method to debiase language models from unwanted behaviors in NLU tasks by identifying pruning masks that can be applied to a finetuned model. |
| Outcome: | The proposed method shows superior performance and performance over standard methods. |
C2PO: Diagnosing and Disentangling Bias Shortcuts in LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to solve large language models address stereotypical and structural biases in isolation . however, prior paradigms address these in isolation, often at the expense of exacerbating the other . |
| Approach: | They propose a framework to tackle latent spurious feature correlations within input that drive erroneous reasoning shortcuts. |
| Outcome: | The proposed framework mitigates stereotypical and structural biases while preserving robust general reasoning capabilities. |
CRISP: Persistent Concept Unlearning via Sparse Autoencoders (2026.acl-long)
Copied to clipboard
| Challenge: | Recent work has explored sparse autoencoders (SAEs) to perform precise interventions on monosemantic features, but most SAE-based methods operate at inference time, which does not create persistent changes in the model’s parameters. |
| Approach: | They propose a parameter-efficient method for persistent concept unlearning using SAEs that automatically identifies salient SAE features across multiple layers and suppresses their activations. |
| Outcome: | The proposed method outperforms previous methods on safety-critical unlearning tasks from the WMDP benchmark, successfully removing harmful knowledge while preserving general and in-domain capabilities. |
FIND: Human-in-the-Loop Debugging Deep Text Classifiers (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing models are limited in the number of available datasets and lack the necessary tools to improve them. |
| Approach: | They propose a framework which enables humans to debug deep learning text classifiers by disabling irrelevant hidden features. |
| Outcome: | Experiments show that using FIND, humans can improve CNN text classifiers trained on different types of imperfect datasets. |