Papers with MLP layers
On Localizing and Deleting Toxic Memories in Large Language Models (2025.findings-naacl)
Copied to clipboard
Anubrata Das, Manoj Kumar, Ninareh Mehrabi, Anil Ramakrishna, Anna Rumshisky, Kai-Wei Chang, Aram Galstyan, Morteza Ziyadi, Rahul Gupta
| Challenge: | Existing methods to reduce toxic generation in large language models are not fully understood. |
| Approach: | They propose to understand the mechanisms that drive toxic generation in large language models by using memory localization to reduce toxic generation. |
| Outcome: | The proposed method reduces toxic generation from 62.86% to 28.61%, but it also improves generation quality. |
Making Large Language Models Efficient Dense Retrievers (2026.acl-long)
Copied to clipboard
| Challenge: | Recent studies have shown that fine-tuning large language models for dense retrieval yields strong performance, but their substantial parameter counts make them computationally inefficient. |
| Approach: | They propose a framework for developing efficient retrievers that performs coarse-to-fine compression through a coarse-grained coarse-tuning strategy. |
| Outcome: | The proposed framework reduces model size and inference cost while preserving performance of full-size models. |
Preference Tuning For Toxicity Mitigation Generalizes Across Languages (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Detoxifying multilingual Large Language Models (LLMs) has become crucial due to their increasing global use. |
| Approach: | They propose to use English preference tuning to study cross-lingual detoxification of LLMs. |
| Outcome: | The proposed method reduces toxicity in multilingual LLMs by reducing the probability of mGPT-1.3B generating toxic continuations across 17 languages. |
How Does DPO Reduce Toxicity? A Mechanistic Neuron-Level Analysis (2025.emnlp-main)
Copied to clipboard
| Challenge: | Direct Preference Optimization (DPO) is a popular choice of safety fine-tuning algorithms, but prior explanations of its effects only account for dampened toxic neurons in the MLP layers. |
| Approach: | They analysed four language models and found that toxic neurons only account for 2.5% to 24% of DPO’s effects across models. |
| Outcome: | The proposed method outperforms DPO in reducing toxicity while preserving perplexity, without requiring any weight updates. |