Papers by Zi Wang

18 papers
Select-Then-Decompose: From Empirical Analysis to Adaptive Selection Strategy for Task Decomposition in Large Language Models (2025.emnlp-main)

Copied to clipboard

Challenge: Existing task decomposition methods focus on memory, tool usage, and feedback mechanisms, but they often overlook the trade-off between performance and cost.
Approach: They propose a strategy that selects the most suitable decomposition approach based on task characteristics and enhances the reliability of the results through a verification module.
Outcome: The proposed strategy is based on categories of approaches, characteristics of tasks, and configuration of decomposition and execution models.
Adaptive Feature-based Low-Rank Compression of Large Language Models via Bayesian Optimization (2024.findings-emnlp)

Copied to clipboard

Challenge: Large language models require a balance between efficiency and performance.
Approach: They propose a low-rank compression technique that reduces non-essential parameters by decomposing weight matrices into products of two low-ranked matrici.
Outcome: The proposed method outperforms existing pruning and low-rank compression techniques in maintaining model performance at the same compression ratio.
On Evaluating Multilingual Compositional Generalization with Translated Datasets (2023.acl-long)

Copied to clipboard

Challenge: a growing amount of research investigating compositional generalization in NLP is done on English . a critical semantic distortion is a limitation of the translation of datasets .
Approach: They propose to translate a dataset for evaluating compositional generalization in semantic parsing.
Outcome: The proposed benchmarks show that the translation of the MCWQ dataset suffers from semantic distortion.
“Yes, My LoRD.” Guiding Language Model Extraction with Locality Reinforced Distillation (2025.acl-long)

Copied to clipboard

Challenge: Existing methods for model extraction attacks on large language models are inadequate . existing methods neglect the inconsistency between training tasks and LLM alignment .
Approach: They propose a model extraction algorithm that uses a policy-gradient-style training task to guide the crafting of preference for the local model.
Outcome: The proposed algorithm reduces query complexity while mitigating watermark protection . it can extract various state-of-the-art commercial LLMs while minimizing query complexity .
OPT-Tree: Speculative Decoding with Adaptive Draft Tree Structure (2025.tacl-1)

Copied to clipboard

Challenge: Autoregressive language models generate one token in one step, limiting inference efficiency . Existing methods do not adapt to different situations to maximize acceptance length . speculative decoding has shown great potential for lossless acceleration .
Approach: They propose an algorithm to construct adaptive and scalable draft trees for autoregressive language models.
Outcome: Experimental results show that OPT-Tree outperforms existing draft trees and achieves speed-up ratio of up to 3.2 compared with autoregressive decoding.
Text Meets Topology: Rethinking Out-of-distribution Detection in Text-Rich Networks (2025.emnlp-main)

Copied to clipboard

Challenge: Existing methods for out-of-distribution (OOD) detection ignore textual-structural diversity . text-rich networks (TrNs) represent complex interplay between textual content and relational structures .
Approach: They propose a framework for evaluating out-of-distribution detection in text-rich networks . they propose augmentations, structural shifts, and domain-based divisions to model interplay .
Outcome: Experiments on 11 datasets show the framework is effective in out-of-distribution detection.
ForestCast: Open-Ended Event Forecasting with Semantic News Forest (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing approaches and datasets overlook the complex relationships among events . current research lacks comprehensive evaluation methods to evaluate OEEF .
Approach: They propose a prediction pipeline that extracts forecast-relevant events from news data . forestcast organizes news events into a story tree and predicts subsequent events along each path .
Outcome: The proposed pipeline extracts forecast-relevant events from news data and predicts subsequent events along each path.
Event-Content-Oriented Dialogue Generation in Short Video (2024.naacl-long)

Copied to clipboard

Challenge: Existing multi-modal dialogue models are limited to incapacity of reading visual information and multi-dimensional interactions.
Approach: They propose a novel event-oriented video-dialogue dataset called SportsVD to overcome these challenges by generating human-like response according to event contents in the video and related external knowledge.
Outcome: The proposed method outperforms existing methods on SportsVD and other baselines under several automatic metrics.
Beyond Prompt Engineering: A Systematic Analysis of Prompt Lexical Sensitivity and Its Impacts on Quality (2026.findings-acl)

Copied to clipboard

Challenge: Existing studies on prompt engineering have focused on optimizing models for performance under stylistic perturbations.
Approach: They conduct the first analysis of n-gram token-level mechanisms . they find that higher average performance is inherently associated with lower variance and greater stability.
Outcome: The proposed model reduces the variance of the generated code by 40% . the proposed model is based on a large-scale dataset of 132,000 prompt variants .
ESDM: Early Sensing Depression Model in Social Media Streams (2024.lrec-main)

Copied to clipboard

Challenge: Existing approaches to use social media data for depression detection are based on traditional risk detection (TRD) and early risk detection of depression (ERD).
Approach: They propose a model that uses two modules: classification with partial information module (CPI) and decision for classification moment module (DMC) and an early detection loss function.
Outcome: The proposed model outperforms benchmarks in both accuracy and accuracy with evolving partial data.
Hint-Based Training for Non-Autoregressive Machine Translation (D19-1)

Copied to clipboard

Challenge: AutoRegressive Translation models have to generate tokens sequentially during decoding and thus suffer from high inference latency.
Approach: They propose to use hidden states and word alignments to help train NART models.
Outcome: The proposed model improves on the WMT14 En-De and De-En datasets but is faster in inference than the current models.
End-to-End Learnable Psychiatric Scale Guided Risky Post Screening for Depression Detection on Social Media (2025.emnlp-main)

Copied to clipboard

Challenge: Existing methods to detect depression from social media posting history are limited by frozen screening models and lack of learning.
Approach: They propose to use a frozen screening model to train a risky post detection model with psychiatric scales to enable a learnable end-to-end learning process.
Outcome: The proposed model outperforms several strong baseline methods and qualitative analysis confirms that it better captures users’ mental states than others.
Making Every Step Effective: Jailbreaking Large Vision-Language Models Through Hierarchical KV Equalization (2025.findings-emnlp)

Copied to clipboard

Challenge: HKVE selectively accepts gradient optimization results based on the distribution of attention scores across different layers, ensuring that every optimization step positively contributes to the attack.
Approach: They propose a framework that selectively accepts gradient optimization results based on the distribution of attention scores across different layers and selectively takes them into account when calculating the attack success rate.
Outcome: The proposed framework outperforms existing methods by achieving success rates of 75.08% on MiniGPT4, 85.84% on LLaVA and 81.00% on Qwen-VL.
ToxicChat: Unveiling Hidden Challenges of Toxicity Detection in Real-World User-AI Conversation (2023.findings-emnlp)

Copied to clipboard

Challenge: toxicity detection has been largely based on social media content, leaving the unique challenges inherent to real-world user-AI interactions insufficiently explored.
Approach: They propose a benchmark to detect toxicity in real-world user-AI conversations . they compare existing models with social media content to find toxicity .
Outcome: The proposed benchmark reveals that existing models fail to recognize toxicity in real-world user-AI conversations.
LSEG: A Fine-tuning Free Method for NL2FOL via Logic-Structure and Entropy Guided Inference Controlling (2026.findings-acl)

Copied to clipboard

Challenge: Large language models struggle with natural language to first order logic (NL2FOL) translation due to logical hallucination.
Approach: They propose a fine-tuning free framework to correct hidden state deviation by leveraging logical stability across logic preserving perturbations of the input.
Outcome: The proposed framework improves logical consistency during inference and improves accuracy over baselines.
SOM-NCSCM : An Efficient Neural Chinese Sentence Compression Model Enhanced with Self-Organizing Map (2021.emnlp-main)

Copied to clipboard

Challenge: Sentence Compression (SC) is an important natural language processing task . it aims to shorten sentences while preserving the original meanings of the words . improvements on Chinese SC models are still lacking due to several difficulties .
Approach: They propose a neural Chinese SC model enhanced with a Self-Organizing Map from Chinese colloquial sentences from a real-life question answering system.
Outcome: The proposed model achieves a promising F1 score of 89.655 and BLEU4 score of 70.116 . it improves the performance of the whole neural Chinese SC model in a valid manner .
Balancing Forget Quality and Model Utility: A Reverse KL-Divergence Knowledge Distillation Approach for Better Unlearning in LLMs (2025.naacl-long)

Copied to clipboard

Challenge: Existing methods for unlearning large language models struggle with forget quality and model utility, leading to over-unlearning or partial unlearning.
Approach: They propose a method that uses reverse KL-divergence based knowledge distillation for unlearning to achieve significant forget quality while maintaining model utility.
Outcome: The proposed method outperforms existing methods in forget quality and model utility with larger unlearning datasets.
CoLA: Compute-Efficient Pre-Training of LLMs via Low-Rank Activation (2025.emnlp-main)

Copied to clipboard

Challenge: Large foundation models have become huge, but they consume computational resources in pretraining.
Approach: They propose to replace full-size layers with compute-efficient auto-encoders that enforce low-rank activations throughout training.
Outcome: The proposed method reduces the computing cost by 2pmbtimes and improves training throughput by 1.86pmtime.

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