Papers by Zhengxuan Wu
pyvene: A Library for Understanding and Improving PyTorch Models via Interventions (2024.naacl-demo)
Copied to clipboard
Zhengxuan Wu, Atticus Geiger, Aryaman Arora, Jing Huang, Zheng Wang, Noah Goodman, Christopher Manning, Christopher Potts
| Challenge: | Existing libraries are often project-based, but pyvene provides a unified and extensible framework for performing interventions on neural models and sharing the intervened upon models with others. |
| Approach: | They propose an open-source Python library that supports customizable interventions on a range of different PyTorch modules. |
| Outcome: | The proposed framework provides a unified and extensible framework for performing interventions on neural models and sharing the intervened upon models with others. |
RAVEL: Evaluating Interpretability Methods on Disentangling Language Model Representations (2024.acl-long)
Copied to clipboard
| Challenge: | Existing methods to disentangle individual neurons from multiple high-level concepts are not yet benchmarked. |
| Approach: | They propose a method of Multi-task Distributed Alignment Search that allows to find distributed representations satisfying multiple causal criteria. |
| Outcome: | The proposed method achieves state-of-the-art on the target language model with Llama2-7B . |
MQuAKE: Assessing Knowledge Editing in Language Models via Multi-Hop Questions (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for retraining from scratch are limited and only work on the recall of edited facts. |
| Approach: | They propose a benchmark method that allows users to ask multi-hop questions to assess whether edited models correctly answer questions where the answer should change as an entailed consequence of edited facts. |
| Outcome: | The proposed method outperforms existing models and scales well with LLMs (up to 175B) it is based on a memory-based approach that stores all edited facts externally while prompting the language model iteratively to generate answers consistent with the edited facts. |
Dynabench: Rethinking Benchmarking in NLP (2021.naacl-main)
Copied to clipboard
Douwe Kiela, Max Bartolo, Yixin Nie, Divyansh Kaushik, Atticus Geiger, Zhengxuan Wu, Bertie Vidgen, Grusha Prasad, Amanpreet Singh, Pratik Ringshia, Zhiyi Ma, Tristan Thrush, Sebastian Riedel, Zeerak Waseem, Pontus Stenetorp, Robin Jia, Mohit Bansal, Christopher Potts, Adina Williams
| Challenge: | Dynabench is an open-source platform for dynamic dataset creation and model benchmarking. |
| Approach: | They propose an open-source platform for dynamic dataset creation and model benchmarking. |
| Outcome: | The proposed platform can be used to create models that fail on simple challenges and falter in real-world scenarios. |
Inducing Character-level Structure in Subword-based Language Models with Type-level Interchange Intervention Training (2023.findings-acl)
Copied to clipboard
| Challenge: | Language models operating on subword units are challenging for character-level manipulations, authors say . authors develop a framework to learn robust character representations inside subword-based models . |
| Approach: | They propose a causal intervention framework to learn robust character representations inside subword-based language models. |
| Outcome: | The proposed model outperforms character-level models on more complex tasks . it improves robustness on unseen token sequences and leads to human-interpretable representations of characters. |
Oolong: Investigating What Makes Transfer Learning Hard with Controlled Studies (2023.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) require vast datasets for pretraining, making it difficult to train LLMs from scratch for lowresource languages. |
| Approach: | They propose to transform a language of the GLUE benchmark and then fine tune a pretrained model on that dataset. |
| Outcome: | The proposed models recover from syntactic-style shifts, but cannot recover from vocabulary misalignment and embedding matrix re-initialization, even with continued pretraining on 15 million tokens. |
DynaSent: A Dynamic Benchmark for Sentiment Analysis (2021.acl-long)
Copied to clipboard
| Challenge: | Sentiment analysis is an early success story for NLP, in both a technical and an industrial sense. |
| Approach: | They propose to combine naturally occurring sentences with sentences created using the open-source Dynabench Platform, which facilities human-and-model-in-the-loop dataset creation. |
| Outcome: | The proposed model is more coherent than comparable models and motivates training models from scratch over successive fine-tuning. |
Causal Distillation for Language Models (2022.naacl-main)
Copied to clipboard
Zhengxuan Wu, Atticus Geiger, Joshua Rozner, Elisa Kreiss, Hanson Lu, Thomas Icard, Christopher Potts, Noah Goodman
| Challenge: | Distillation efforts have led to language models that are more compact and efficient without serious drops in performance. |
| Approach: | They propose to augment distillation with a third objective that encourages the student model to imitate the causal dynamics of the teacher through a distillation interchange intervention training objective (DIITO). |
| Outcome: | The proposed method lowers perplexity on the WikiText-103M corpus and improves on the GLUE benchmark, SQuAD, and CoNLL-2003. |
Dancing in Chains: Reconciling Instruction Following and Faithfulness in Language Models (2024.emnlp-main)
Copied to clipboard
Zhengxuan Wu, Yuhao Zhang, Peng Qi, Yumo Xu, Rujun Han, Yian Zhang, Jifan Chen, Bonan Min, Zhiheng Huang
| Challenge: | Modern language models fail to follow human instructions while being faithful . a trade-off exists between instruction following and faithfulness when training LMs . |
| Approach: | They propose a method that relies on Reject-sampling by Self-instruct with Continued Fine-tuning to train LMs to follow human instructions while being faithful. |
| Outcome: | The proposed method outperforms vanilla MTL with high-quality data, but with significantly smaller data. |
MAR: Matching-Augmented Reasoning for Enhancing Visual-based Entity Question Answering (2024.emnlp-main)
Copied to clipboard
| Challenge: | Multimodal large language models (MLLMs) struggle with visual-based entity questions (VEQA) MLLM can identify A, but may refrain from answering due to privacy concerns. |
| Approach: | They propose a method that uses vector representations to analyze visual-based entity questions (VEQA) they use visual cues and textual information to integrate visual cus and visual information . |
| Outcome: | The proposed method significantly improves visual-based entity question answering (VEQA) it can identify faces, names, and alignments within visual objects, and then derive the answer over this matching graph. |