Papers with adversarial approaches
Semi-supervised Domain Adaptation for Dependency Parsing via Improved Contextualized Word Representations (2020.coling-main)
Copied to clipboard
| Challenge: | Recent advances in deep neural network models have improved parsing performance on in-domain texts . however, the problem is to improve performance on out-of-domain text data when there is only a small-scale out-domain labeled data. |
| Approach: | They propose to use adversarial learning and fine-tuning BERT to improve contextualized word representations on out-of-domain texts. |
| Outcome: | The proposed models achieve consistent improvement and fine-tune BERT processes boost parsing accuracy by a large margin. |
Learning Implicit Text Generation via Feature Matching (2020.acl-main)
Copied to clipboard
Inkit Padhi, Pierre Dognin, Ke Bai, Cícero Nogueira dos Santos, Vijil Chenthamarakshan, Youssef Mroueh, Payel Das
| Challenge: | Generative feature matching networks are an approach for training implicit generative models for images . a novel formulation of GFMN for unconditional sequence generation is proposed . |
| Approach: | They propose a new GFMN formulation that performs token level feature matching on pre-trained neural networks. |
| Outcome: | The proposed method outperforms adversarial approaches for text generation and style transfer. |
Is Multihop QA in DiRe Condition? Measuring and Reducing Disconnected Reasoning (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing models exploit dataset artifacts to produce correct answers without connecting information across multiple facts. |
| Approach: | They formalize disconnected reasoning across subsets of supporting facts to reduce disconnected reasoning . they propose an automatic transformation of existing datasets that reduces disconnected reasoning. |
| Outcome: | The proposed model-agnostic probe reduces disconnected reasoning in a reading comprehension setting. |
Adversarial Tokenization (2025.acl-long)
Copied to clipboard
| Challenge: | Current LLM pipelines account for only one possible tokenization for a given string . authors: noncanonical tokenizations can evade LLM safety while still generating meaningful responses. |
| Approach: | They show that LLM pipelines account for only one possible tokenization for a given string . they show that tokenizers retain semantic understanding of other tokenizations . authors propose an exploit that can be exploited to evade safety and alignment restrictions . |
| Outcome: | The proposed exploit exploits a previously unknown vulnerability in subword models. |