Papers with gradient ascent
Balancing Forget Quality and Model Utility: A Reverse KL-Divergence Knowledge Distillation Approach for Better Unlearning in LLMs (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing methods for unlearning large language models struggle with forget quality and model utility, leading to over-unlearning or partial unlearning. |
| Approach: | They propose a method that uses reverse KL-divergence based knowledge distillation for unlearning to achieve significant forget quality while maintaining model utility. |
| Outcome: | The proposed method outperforms existing methods in forget quality and model utility with larger unlearning datasets. |
EFUF: Efficient Fine-Grained Unlearning Framework for Mitigating Hallucinations in Multimodal Large Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to eliminate hallucinations require expensive human annotation . hallucination in multimodal large language models poses unique challenges for current research . |
| Approach: | They propose a fine-grained unlearning framework that performs gradient ascent to eliminate hallucinations without paired data. |
| Outcome: | The proposed method reduces hallucinations while preserving quality with modest computational overhead. |
Jailbreaking Prompt Attack: A Controllable Adversarial Attack against Diffusion Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Text-to-image (T2I) models can be used to generate harmful content such as sexually explicit, unfaithful, and misleading or Not-Safe-for-Work (NSFW) images. |
| Approach: | They propose a more practical and universal attack that does not require the presence of a target model. |
| Outcome: | The proposed attack bypasses both text and image safety checkers while preserving high semantic alignment with the target prompt. |
Disentangling Biased Knowledge from Reasoning in Large Language Models via Machine Unlearning (2025.acl-long)
Copied to clipboard
Zheyuan Liu, Suraj Maharjan, Fanyou Wu, Rahil Parikh, Belhassen Bayar, Srinivasan H. Sengamedu, Meng Jiang
| Challenge: | Existing approaches to disentangle biased knowledge from reasoning are sub-optimal . entangled data makes curation difficult, leading to inclusion of sensitive, toxic data. |
| Approach: | They propose a framework that selectively removes biased knowledge while preserving reasoning abilities. |
| Outcome: | The proposed framework improves fairness accuracy by 14.7% and reasoning performance by 62.6% across multiple LLMs. |
Unlearning Traces the Influential Training Data of Language Models (2024.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) exhibit remarkable abilities without being explicitly trained for such tasks. |
| Approach: | They propose a method that unlearns a test dataset and evaluates the unlearned model on training datasets. |
| Outcome: | The proposed method resembles UnTrac, while being efficient for massive training datasets. |
SGPVT: Self-Generated Proximal Visual Tokens for Mitigating Proximal Collateral Damage in MLLM Unlearning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches focus on general utility metrics, overlooking the preservation of semantically related concepts. |
| Approach: | They propose a method that introduces self-generated proximal visual tokens to prevent forgetting vulnerability. |
| Outcome: | The proposed framework outperforms existing methods in preserving semantically related concepts while achieving effective target unlearning. |
Machine Unlearning of Pre-trained Large Language Models (2024.acl-long)
Copied to clipboard
| Challenge: | Using curated datasets, we establish a robust benchmark for unlearning performance, demonstrating that these methods are over 105 times more computationally efficient than retraining. |
| Approach: | They propose a framework for machine unlearning in pre-trained LLMs and integrate gradient ascent with gradient descent on in-distribution data to achieve robustness. |
| Outcome: | The proposed framework is over 105 times more efficient than retraining on in-distribution data and provides detailed guidelines for efficient hyperparameter tuning in the unlearning process. |
Knowledge Unlearning for Mitigating Privacy Risks in Language Models (2023.acl-long)
Copied to clipboard
| Challenge: | Recent work shows that an adversary can extract training data from Pretrained Language Models including Personally Identifiable Information (PII) such as names, phone numbers, and email addresses. |
| Approach: | They propose to use knowledge unlearning to reduce privacy risks for LMs by performing gradient ascent on target token sequences instead of trying to unlearn all the data at once. |
| Outcome: | The proposed method can give a stronger empirical privacy guarantee in scenarios where the data vulnerable to extraction attacks are known a priori while being much more efficient and robust. |
CLEAR: Character Unlearning in Textual and Visual Modalities (2025.findings-acl)
Copied to clipboard
Alexey Dontsov, Dmitrii Korzh, Alexey Zhavoronkin, Boris Mikheev, Denis Bobkov, Aibek Alanov, Oleg Rogov, Ivan Oseledets, Elena Tutubalina
| Challenge: | Existing methods for removing private or hazardous data from deep learning models are focused on single-modality models. |
| Approach: | They propose CLEAR, the first open-source benchmark specifically for MMU. CLEAR contains 200 fictitious individuals and 3,700 images linked with corresponding question-answer pairs. |
| Outcome: | The proposed benchmarks show that unlearning both modalities outperform single-modality approaches. |
Rectifying Belief Space via Unlearning to Harness LLMs’ Reasoning (2025.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) exhibit sophisticated reasoning yet still generate incorrect answers. |
| Approach: | They propose a belief space rectification framework that suppresses spurious beliefs and enhances true ones to reduce erroneous reasoning and generalization. |
| Outcome: | The proposed framework reduces erroneous reasoning and improves generalization on three QA datasets and three LLMs. |