| Challenge: | Existing methods for erasing human-interpretable concepts from neural representations that assume linearity are not fully understood. |
| Approach: | They define linear guardedness as the inability of an adversary to predict the concept directly from the representation . they show that a log-linear model can be constructed that indirectly recovers the concept . |
| Outcome: | The proposed model can be constructed that indirectly recovers the erased concept in some cases. |
Similar Papers
Adversarial Concept Erasure in Kernel Space (2022.emnlp-main)
Copied to clipboard
| Challenge: | Large neural networks in NLP produce real-valued representations that encode the bit of human language that they were trained on. |
| Approach: | They propose a kernelization of the recently-proposed linear concept-removal objective and propose to remove linear subspaces from the representation space. |
| Outcome: | The proposed kernelization protects against the ability of nonlinear adversaries to recover the concept. |
On Transferability of Bias Mitigation Effects in Language Model Fine-Tuning (2021.naacl-main)
Copied to clipboard
| Challenge: | PTLMs can exhibit biases against protected groups in a host of modeling tasks . but, fine-tuned LMs may propagate bias to downstream classifiers . |
| Approach: | They propose to use upstream bias mitigation techniques to reduce bias on downstream tasks by fine-tuning an upstream model and applying it to a downstream model. |
| Outcome: | The proposed model reduces bias on hate speech detection, toxicity detection and coreference resolution tasks over bias factors. |
Safety Alignment via Constrained Knowledge Unlearning (2025.acl-long)
Copied to clipboard
Zesheng Shi, Yucheng Zhou, Jing Li, Yuxin Jin, Yu Li, Daojing He, Fangming Liu, Saleh Alharbi, Jun Yu, Min Zhang
| Challenge: | Existing defense mechanisms have not fully deleted harmful knowledge in large language models (LLMs) Existing methods to address safety alignment have not completely deleted harmful information in LLMs. |
| Approach: | They propose a safety alignment strategy that uses scoring neurons to identify useful knowledge in LLMs and pruning the gradients of neurons in U to preserve beneficial information. |
| Outcome: | The proposed method significantly improves model safety while maintaining utility compared to existing methods. |
Gold Doesn’t Always Glitter: Spectral Removal of Linear and Nonlinear Guarded Attribute Information (2023.eacl-main)
Copied to clipboard
| Challenge: | Spectral Attribute removaL is a method to remove private or guarded information from neural representations. |
| Approach: | They propose a method to remove guarded or private information from neural representations by matrix decomposition. |
| Outcome: | The proposed method retains better main task performance after removing guarded information compared to previous work. |
Linear Steerability in Language Models: When It Emerges and How It Evolves (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a new framework for steering language models reveals how concepts become linearly separable as training progresses . |
| Approach: | They propose a framework to analyze steerability in language models by using hidden state and representation analysis. |
| Outcome: | The proposed framework reveals how steerability evolves over training . concepts become linearly separable as training progresses, the framework shows . |
Steering Safely or Off a Cliff? Rethinking Specificity and Robustness in Inference-Time Interventions (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing studies have shown that model steering can preserve fluency and unrelated abilities, but it fails to preserve robustness specificity. |
| Approach: | They propose a framework that distinguishes three dimensions of specificity: general, control, and robustness. |
| Outcome: | The proposed framework distinguishes three dimensions of specificity: general (preserving fluency and unrelated abilities), control (preserving related control properties), and robustness (preserving control properties under distribution shifts). |
Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection (2020.acl-main)
Copied to clipboard
| Challenge: | Word embeddings, pre-trained language models, and deep learning methods are becoming effective for text classification. |
| Approach: | They propose a method for removing information from neural representations using null-space projection. |
| Outcome: | The proposed method mitigates bias in word embeddings and increases fairness in multi-class classification. |
Language Concept Erasure for Language-invariant Dense Retrieval (2024.emnlp-main)
Copied to clipboard
| Challenge: | Multilingual models aim for language-invariant representations but still encode language identity. |
| Approach: | They propose a multi-task learning framework that induces language invariance in multilingual retrieval by reducing language-specific signals in the embedding space. |
| Outcome: | The proposed learning framework improves language-invariant dense retrieval over baselines on English retrieval data and general multilingual corpora. |
Do Neural Language Models Overcome Reporting Bias? (2020.coling-main)
Copied to clipboard
| Challenge: | Recent studies show that pre-trained language models can overcome reporting bias by estimating the plausibility of rare but unspoken facts. |
| Approach: | They revisit the experiments conducted by Gordon and Van Durme (2013) . they find that pre-trained language models overestimate the very rare . |
| Outcome: | The proposed approach overestimates the rare at the expense of the rare, while minimizing reporting bias. |
Upstream Mitigation Is Not All You Need: Testing the Bias Transfer Hypothesis in Pre-Trained Language Models (2022.acl-long)
Copied to clipboard
| Challenge: | Large language models and other massively pre-trained "foundation" models can easily adapt to a wide variety of downstream tasks in a process called finetuning. |
| Approach: | They propose to use the bias transfer hypothesis to reduce social biases internalized by large language models during pre-training into harmful task-specific behavior after fine-tuning. |
| Outcome: | The bias transfer hypothesis is the theory that social biases internalized by large language models during pre-training transfer into harmful task-specific behavior after fine-tuning. |