Papers by Tianwei Zhang

20 papers
Course-Correction: Safety Alignment Using Synthetic Preferences (2024.emnlp-industry)

Copied to clipboard

Challenge: Recent studies show that large language models generate harmful content, but the potential for generating harmful content is an escalating concern.
Approach: They propose to fine-tune LLMs with preference learning to emphasize the preference for timely course-correction by using an automated pipeline.
Outcome: The proposed model improves course-correction skills without affecting general performance and resists jailbreak attacks.
GNN-SL: Sequence Labeling Based on Nearest Examples via GNN (2023.findings-acl)

Copied to clipboard

Challenge: Existing sequence labeling algorithms can be decomposed into two parts .
Approach: They propose a graph neural networks sequence labeling (GNN-SL) that augments the vanilla SL model output with similar tagging examples retrieved from the whole training set.
Outcome: The proposed model performs well on three sequence labeling tasks.
Multi-target Backdoor Attacks for Code Pre-trained Models (2023.acl-long)

Copied to clipboard

Challenge: Existing work for backdoor attacks on neural code models insert triggers into task-specific data for code-related downstream tasks, limiting the scope of attacks.
Approach: They propose task-agnostic backdoor attacks for code pre-trained models . they use two learning strategies to implant backdoors into code understanding and generation models - Poisoned Seq2Seq learning and token representation learning .
Outcome: The proposed model is pre-trained with two learning strategies to support the multi-target attack of downstream code understanding and generation tasks.
TeamLoRA: Boosting Low-Rank Adaptation with Expert Collaboration and Competition (2025.acl-long)

Copied to clipboard

Challenge: Existing methods for fine-tuning are resource-efficient, but performance often falls short . a new approach, TeamLoRA, integrates collaborative and competitive modules to improve performance.
Approach: They propose to introduce task-specific LoRA as domain experts to improve learning efficiency . teamLoRA integrates collaborative and competition modules to improve model learning .
Outcome: Experiments show that TeamLoRA improves performance in multi-task learning . teamLorea integrates collaborative and competitive modules to improve performance .
LeakDojo: Decoding the Leakage Threats of RAG Systems (2026.findings-acl)

Copied to clipboard

Challenge: Existing studies have failed to assess RAG leakage risks for large language models . constructing and maintaining highquality RAG knowledge databases has become increasingly costly .
Approach: They propose a framework for controlled evaluation of RAG leakage using query generation and adversarial instructions.
Outcome: The proposed framework compares six existing attacks across fourteen LLMs, four datasets, and diverse RAG systems.
CoMoL: Efficient Mixture of LoRA Experts via Dynamic Core Space Merging (2026.findings-acl)

Copied to clipboard

Challenge: Existing PEFT methods suffer from limited parameter efficiency and coarse-grained adaptation due to proliferation of LoRA experts and instance-level routing.
Approach: They propose a new MoE-LoRA framework that incorporates expert diversity, parameter efficiency, and fine-grained adaptation.
Outcome: The proposed framework outperforms existing methods on multiple tasks while maintaining parameter efficiency.
Walking in Others’ Shoes: How Perspective-Taking Guides Large Language Models in Reducing Toxicity and Bias (2024.emnlp-main)

Copied to clipboard

Challenge: Existing prompting methods that require white-box access to the model or substantial training fail to simultaneously lessen toxicity and bias.
Approach: They propose a strategy that encourages LLMs to integrate diverse human perspectives and self-regulate their responses by incorporating diverse human viewpoints.
Outcome: The proposed approach can significantly diminish toxicity (up to 89%) and bias (up 73%) in LLMs’ responses.
MoA: Heterogeneous Mixture of Adapters for Parameter-Efficient Fine-Tuning of Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for parameter-efficient fine-tuning (PEFT) are limited by computational costs and performance degradation.
Approach: They propose a method that integrates Low-Rank Adaptation and Mixture-of-Experts (MoE) they propose combining expert load imbalance and representation collapse to improve LLM performance .
Outcome: The proposed method outperforms homogeneous MoE-LoRA architectures in performance and parameter efficiency.
Vision-Language Models Mistake Head Orientation for Gaze Direction: Nonverbal Conversation Cues (2026.findings-acl)

Copied to clipboard

Challenge: Where someone looks is a nonverbal communication cue that children and adults readily use.
Approach: They used 1,360 real-world photos to construct evaluation stimuli for Vision-Language Models (VLMs) they found a substantial performance gap between VLMs and humans .
Outcome: The proposed model outperforms existing models in predicting gaze direction using head orientation rather than eye appearance.
Speculating LLMs’ Chinese Training Data Pollution from Their Tokens (2025.emnlp-main)

Copied to clipboard

Challenge: Experiments on GPT and other 23 LLMs indicate that tokens widely exist while GPT’s vocabulary behaves the worst: more than 23% long Chinese tokens (i.e., a token with more than two Chinese characters) are either porn or online gambling.
Approach: They propose to locate Polluted Chinese (PoC) tokens in LLMs and build a PoC token detector to label them in vocabularies by considering each token’s semantics and related contents from the search engines.
Outcome: The proposed method predicts that the ratio of “*” related webpages in GPT-4o's training data is around 0.5%.
Exploring Multimodal Challenges in Toxic Chinese Detection: Taxonomy, Benchmark, and Findings (2025.findings-acl)

Copied to clipboard

Challenge: Recent studies show that character substitutions in toxic Chinese text can confuse state-of-the-art LLMs.
Approach: They propose a taxonomy of 3 perturbation strategies and 8 specific approaches in Chinese text to assess if they can detect perturbed Chinese toxic contents.
Outcome: The proposed model can detect perturbed Chinese text with 8 different approaches . the proposed model is compared with 9 other LLMs from the US and China .
kFolden: k-Fold Ensemble for Out-Of-Distribution Detection (2021.emnlp-main)

Copied to clipboard

Challenge: Existing studies studying OOD detection in NLP often rely on external data to diversify model predictions.
Approach: They propose a framework which mimics OOD detection behavior without external data . they take text classification as an archetype and compare them to existing datasets .
Outcome: The proposed framework can resolve in- and out-distribution examples in a natural way using existing datasets.
Fast Nearest Neighbor Machine Translation (2022.findings-acl)

Copied to clipboard

Challenge: Fast kNN-MT uses the entire corpus as the datastore for the nearest neighbor search . knn-MT is two-orders slower than vanilla MT models .
Approach: They propose a fast kNN-MT model that uses the entire corpus as the datastore for nearest neighbor search.
Outcome: The proposed model is two-orders faster than kNN-MT and is only two times slower than the standard model.
Understanding the Dark Side of LLMs’ Intrinsic Self-Correction (2025.acl-long)

Copied to clipboard

Challenge: Recent studies show that LLMs’ intrinsic self-correction fails without oracle labels as feedback.
Approach: They propose to use one simple task and three complex tasks with state-of-the-art LLMs like ChatGPT, Llama, and DeepSeek to interpret LLM's intrinsic self-correction.
Outcome: The proposed methods reveal the dark side of LLMs’ intrinsic self-correction for different tasks, especially for those failure cases.
When Audio and Text Disagree: Revealing Text Bias in Large Audio-Language Models (2025.emnlp-main)

Copied to clipboard

Challenge: Large Audio-Language Models (LALMs) are augmented with the ability to perceive audio, but their reliability when faced with conflicting inputs remains largely unexplored.
Approach: They examine how LALMs prioritize information when presented with inconsistent audio-text pairs.
Outcome: The proposed models display a significant bias toward textual input when presented with inconsistent audio-text pairs.
Triggerless Backdoor Attack for NLP Tasks with Clean Labels (2022.naacl-main)

Copied to clipboard

Challenge: Backdoor attacks are a new threat to neural natural language processing models due to the fragility and lack of interpretability of NLP models.
Approach: They propose a method to perform backdoor attacks without an external trigger . they propose to use clean-labeled examples to generate poisoned clean-labelled examples .
Outcome: The proposed strategy is effective and hard to defend due to its triggerless nature.
GPT-NER: Named Entity Recognition via Large Language Models (2025.findings-naacl)

Copied to clipboard

Challenge: Large-scale language models (LLMs) have shown impressive ability for in-context learning with limited training data.
Approach: They propose a novel sequence labeling task that transforms a sequence labeled as a text-generation task into a self-verification task that LLMs can adapt to.
Outcome: The proposed model performs better on NER than supervised models on a variety of tasks . the proposed model can be easily adapted by LLMs to generate a text sequence .
Layer-wise Model Pruning based on Mutual Information (2021.emnlp-main)

Copied to clipboard

Challenge: In spite of impressive results of neural networks, the huge model size has hindered their applications in cases where computation and memory resources are limited.
Approach: They propose a method for layer-wise pruning using mutual information based feature selection in SVMs and logistic regression.
Outcome: The proposed pruning strategy offers greater speedup and higher performance than weight-based pruning methods.
Sentence Similarity Based on Contexts (2022.tacl-1)

Copied to clipboard

Challenge: Existing methods to measure sentence similarity face limited dataset size and training-test gap . existing methods lack large-scale labeled datasets with labeles that are labor-intensive and expensive .
Approach: They propose a framework that measures sentence similarity by comparing probabilities of generating two sentences given the same context.
Outcome: The proposed framework achieves significant performance boosts over baselines under supervised and unsupervised settings.
Text Classification via Large Language Models (2023.findings-emnlp)

Copied to clipboard

Challenge: Large-scale Language Models (LLMs) have shown the ability for in-context learning.
Approach: They propose a progressive reasoning strategy tailored to addressing complex linguistic phenomena such as intensification, contrast, irony and limited number of tokens allowed in in-context learning.
Outcome: The proposed model performs better on 4 out of 5 widely-used text-classification benchmarks, while demonstrating comparable performance to SOTA on MR.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations