Papers by Ziyu Zhao

16 papers
C2DLM: Causal Concept-Guided Diffusion Large Language Models (2026.findings-acl)

Copied to clipboard

Challenge: Autoregressive (AR) and diffusion language models (DLMs) suffer from insufficient reasoning capabilities.
Approach: They propose a fully connected Diffusion Language Model that uses a concept-level causal graph to guide attention to learn causal relationships between concepts.
Outcome: The proposed model achieves a 12% improvement and 3.2 training speedup on the COT-OrderPerturb task, along with an average gain of 1.31% across six downstream reasoning tasks.
Dropping Experts, Recombining Neurons: Retraining-Free Pruning for Sparse Mixture-of-Experts LLMs (2025.findings-emnlp)

Copied to clipboard

Challenge: Sparse Mixture-of-Experts (SMoE) architectures require loading all expert parameters . previous work focused on expert pruning and merging but focused on neuron-level structure .
Approach: They propose a task-agnostic framework for expert pruning and reconstruction . it prunes redundant experts using router statistics, then decomposes them into neuron-level expert segments .
Outcome: The proposed framework reduces the number of experts and memory usage, making it easier to deploy.
StructuThink: Reasoning with Task Transition Knowledge for Autonomous LLM-Based Agents (2025.findings-emnlp)

Copied to clipboard

Challenge: StructuThink framework enhances LLMs' ability to ground decisions in domain-specific scenarios.
Approach: They propose a knowledge-structured reasoning framework that enhances LLM-based agents with explicit decision constraints.
Outcome: The proposed framework achieves higher task success rates and more efficient action sequences than baseline methods.
A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models (2025.findings-emnlp)

Copied to clipboard

Challenge: Sparse Autoencoders (SAEs) can disentangle complex features into more interpretable components.
Approach: They propose to use Sparse Autoencoders to disentangle LLM features into more interpretable components.
Outcome: The proposed method disentangles complex features into more interpretable components.
LoraRetriever: Input-Aware LoRA Retrieval and Composition for Mixed Tasks in the Wild (2024.findings-acl)

Copied to clipboard

Challenge: Low-Rank Adaptation (LoRA) is an effective yet efficient solution for fine-tuning large language models.
Approach: They propose a low-rank Adaptation framework that retrieves and composes multiple LoRAs according to input prompts.
Outcome: Experimental results show that LoraRetriever outperforms baselines in terms of performance and versatility.
SciMDR: Advancing Scientific Multimodal Document Reasoning (2026.acl-long)

Copied to clipboard

Challenge: Current models struggle to provide reliable assistance in real-world scientific workflows because evidence is distributed across long, multimodal documents.
Approach: They propose a framework for QA Synthesis and document-scale regrounding that generates faithful, isolated QA pairs and reasoning on focused segments.
Outcome: The proposed framework achieves significant improvements across multiple QA benchmarks, particularly in tasks requiring complex document-level reasoning.
Cross-lingual Multimodal Sentiment Analysis for Low-Resource Languages via Language Family Disentanglement and Rethinking Transfer (2025.findings-acl)

Copied to clipboard

Challenge: Existing multimodal sentiment analysis methods are limited to textual data and cannot handle multimodal scenarios.
Approach: They propose a transfer learning framework that allows cross-lingual and cross-modal alignments and a language family disentanglement module that enhances the sharing of language universals within families.
Outcome: The proposed method is superior to existing methods and can handle low-resource languages.
Unveiling the Deficiencies of Pre-trained Text-and-Layout Models in Real-world Visually-rich Document Information Extraction (2026.findings-eacl)

Copied to clipboard

Challenge: PTLMs have shown remarkable success in multiple information extraction tasks . however, their performance in real-world scenarios falls short of expectations .
Approach: They propose to use an entity-centric dataset to evaluate PTLMs' performance . they find that inadequate annotations in benchmark datasets lead to spurious correlations .
Outcome: The proposed dataset disentangles the falsely-coupled segment and entity annotations that arises from the block-level annotation of FUNSD.
On Domain-Adaptive Post-Training for Multimodal Large Language Models (2025.findings-emnlp)

Copied to clipboard

Challenge: Adapting general multimodal large language models to specific domains is important for practical applications.
Approach: They investigate domain adaptation of multimodal large language models via post-training . they develop a generate-then-filter pipeline that curates diverse visual instruction tasks .
Outcome: The proposed model outperforms existing models in domain adaptation by combining data from open-source models with training pipelines.
H-LegalKI: A Hierarchical Legal Knowledge Integration Framework for Legal Community Question Answering (2024.findings-emnlp)

Copied to clipboard

Challenge: Legal question answering (LQA) aims to bridge the gap between limited availability of legal professionals and the extensive volume of legal issues.
Approach: They propose a legal knowledge retriever and a hierarchical legal knowledge integration framework to address multiple user-specific circumstances.
Outcome: The proposed framework outperforms baselines on the legal community question-answering dataset.
KoCo-Bench: Can Large Language Models Leverage Domain Knowledge in Software Development? (2026.acl-long)

Copied to clipboard

Challenge: Existing domain-specific code benchmarks focus on assessing what knowledge LLMs possess rather than how they acquire and apply new knowledge.
Approach: They propose a benchmark to evaluate domain specialization methods in real-world software development.
Outcome: KOCO-bench is a new benchmark for evaluating domain specialization methods in real-world software development.
LRBench and Judge-R1: Principled Evaluation and Training of LLM-Based Judges for Long-Context Reasoning (2026.findings-acl)

Copied to clipboard

Challenge: Existing benchmarks for evaluating large language models (LLMs) under long contexts are underexplored.
Approach: They propose a large-scale benchmark for evaluating large language models (LLMs) that combines reinforcement learning with multi-turn search to enable grounded and principle-aware evaluation.
Outcome: The proposed model outperforms single-turn baselines across domains and principles.
SC2: Towards Enhancing Content Preservation and Style Consistency in Long Text Style Transfer (2024.acl-long)

Copied to clipboard

Challenge: Existing methods for short TST are difficult to implement and can cause content degradation.
Approach: They propose a method to vary the style polarity of text while preserving semantic content.
Outcome: The proposed method improves over baselines and is highly efficient.
SocioBench: Modeling Human Behavior in Sociological Surveys with Large Language Models (2025.emnlp-main)

Copied to clipboard

Challenge: Large language models (LLMs) lack large-scale, systematically constructed benchmarks for evaluating their alignment with real-world social attitudes.
Approach: They propose a benchmark to assess LLMs' alignment with real-world social attitudes . they find LLM models achieve only 30–40% accuracy when simulating individuals .
Outcome: The proposed benchmark shows that LLMs achieve only 30% accuracy when simulating individuals in complex survey scenarios.
OS Agents: A Survey on MLLM-based Agents for Computer, Phone and Browser Use (2025.acl-long)

Copied to clipboard

Challenge: a new generation of (M)LLMs is enabling the creation of superintelligent AI assistants . OS Agents can complete tasks autonomously and have the potential to significantly enhance the lives of billions of users worldwide.
Approach: They propose to build OS Agents that operate within operating systems' GUIs and GUIs . they examine evaluation metrics and benchmarks to identify promising directions .
Outcome: The proposed agents are based on operating systems (OS) and operating systems frameworks.
CAT: Causal Attention Tuning For Injecting Fine-grained Causal Knowledge into Large Language Models (2025.emnlp-main)

Copied to clipboard

Challenge: Existing fine-tuning paradigms focus on aligning LLMs with task-specific objectives.
Approach: They propose a pipeline that leverages human priors to automatically generate token-level causal signals and introduce the Re-Attention mechanism to guide training.
Outcome: The proposed pipeline achieves an average improvement of 5.76% on the STG dataset and 1.56% on downstream tasks.

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