Papers by Mozhi Zhang
InferAligner: Inference-Time Alignment for Harmlessness through Cross-Model Guidance (2024.emnlp-main)
Copied to clipboard
Pengyu Wang, Dong Zhang, Linyang Li, Chenkun Tan, Xinghao Wang, Mozhi Zhang, Ke Ren, Botian Jiang, Xipeng Qiu
| Challenge: | Existing methods for enhancing harmlessness and helpfulness of large language models (LLMs) involve complex and resource-intensive training processes. |
| Approach: | They propose a method that decouples harmlessness from helpfulness during inference phase. |
| Outcome: | The proposed method significantly reduces the attack success rate (ASR) of harmful instructions and jailbreak instructions while maintaining almost unchanged performance in downstream tasks. |
MetaAlign: Align Large Language Models with Diverse Preferences during Inference Time (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing methods to align large language models with human preferences often result in a static alignment that cannot account for the diversity of human preferences in practical applications. |
| Approach: | They propose a method to help large language models dynamically align with various explicit or implicit preferences specified at inference time. |
| Outcome: | The proposed method can help LLMs dynamically align with various explicit or implicit preferences specified at the inference stage, validating the feasibility of MetaAlign. |
Calibrating the Confidence of Large Language Models by Eliciting Fidelity (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large language models with RLHF and RLAIF have good alignment but exhibit overconfidence post-alignment. |
| Approach: | They propose a plug-and-play method to estimate the confidence of large language models. |
| Outcome: | The proposed method has shown good calibration performance on 6 RLHF-LMs on four MCQA datasets. |
A Dataset and Baselines for Multilingual Reply Suggestion (2021.acl-long)
Copied to clipboard
| Challenge: | Reply suggestion models help users process emails and chats faster. |
| Approach: | They present a multilingual reply suggestion dataset with ten languages . they build a generation model and a retrieval model as baselines for MRS . |
| Outcome: | The proposed model complements existing benchmarks for cross-lingual generalization . the model has different strengths in the English monolingual setting and requires different strategies to generalize across languages. |
Interactive Refinement of Cross-Lingual Word Embeddings (2020.emnlp-main)
Copied to clipboard
| Challenge: | Cross-lingual word embeddings transfer knowledge between languages to models trained on resource-rich languages can predict in low-resource languages. |
| Approach: | They propose an interactive system to quickly refine cross-lingual word embeddings for a given classification problem. |
| Outcome: | The proposed system improves on identifying health-related text in four low-resource languages. |
Why Overfitting Isn’t Always Bad: Retrofitting Cross-Lingual Word Embeddings to Dictionaries (2020.acl-main)
Copied to clipboard
| Challenge: | Recent studies only evaluate cross-lingual word embeddings on bilingual lexicon induction (BLI) however, underfitting can hinder generalization to other downstream tasks. |
| Approach: | They retrofit cross-lingual word embeddings to the training dictionary and a synthetic dictionary to improve their results. |
| Outcome: | The proposed method improves accuracy on two downstream tasks, despite underfitting the training dictionary. |
Are Girls Neko or Shōjo? Cross-Lingual Alignment of Non-Isomorphic Embeddings with Iterative Normalization (P19-1)
Copied to clipboard
| Challenge: | Cross-lingual word embeddings (CLWE) are used to perform multilingual natural language processing tasks. |
| Approach: | They propose a method that transforms monolingual embeddings to make orthogonal alignment easier by simultaneously enforcing that (1) individual word vectors are unit length, and (2) each language’s average vector is zero. |
| Outcome: | The proposed method improves translation accuracy of three CLWE methods, with the largest improvement observed on English-Japanese (2% to 44% test accuracy). |