Papers by Haiyang Xu

23 papers
MIBench: Evaluating Multimodal Large Language Models over Multiple Images (2024.emnlp-main)

Copied to clipboard

Challenge: Existing benchmarks and MLLMs focus on single-image input scenarios, leaving performance of ML models when handling multiple images underexplored.
Approach: They propose a benchmark to evaluate fine-grained abilities of multimodal large language models in multi-image scenarios.
Outcome: The proposed benchmark categorizes the multi-image abilities into three scenarios: MII, MKS and MIC.
How Alignment and Jailbreak Work: Explain LLM Safety through Intermediate Hidden States (2024.findings-emnlp)

Copied to clipboard

Challenge: Large language models (LLMs) rely on safety alignment to avoid malicious user inputs.
Approach: They employ weak classifiers to explain LLM safety through the intermediate hidden states.
Outcome: The proposed model can identify malicious and normal inputs and detect malicious ones without jailbreak.
Experience-driven Multi-turn Reinforcement Learning for GUI Agents (2026.acl-long)

Copied to clipboard

Challenge: GUI agents have demonstrated remarkable progress in automating complex user interface interactions . training such agents for long-horizon tasks remains challenging due to limited rewards and prohibitive costs.
Approach: They propose a method that leverages expert trajectories as environment experiences for on-policy multi-turn training.
Outcome: The proposed method achieves significant gains over the base model with 1K public trajectories as RL experiences . it achieves competitive performance against strong baselines such as UI-TARS-7B and GPT-4o .
Vision Language Pre-training by Contrastive Learning with Cross-Modal Similarity Regulation (2023.acl-long)

Copied to clipboard

Challenge: Large-scale pre-trained vision-language models have recently achieved tremendous success on a wide range of cross-modal tasks.
Approach: They propose a new framework for a semantically-aware contrastive learning that minimizes the MI between false negative and positive samples .
Outcome: The proposed framework minimizes the MI between false negative samples and positive samples even though they share similar semantics.
mPLUG-DocOwl 1.5: Unified Structure Learning for OCR-free Document Understanding (2024.findings-emnlp)

Copied to clipboard

Challenge: Existing Multimodal Large Language Models lack general structure understanding abilities for text-rich document images.
Approach: They propose to use unified structure learning to boost the performance of MLLMs by encoding structure information into text-rich images.
Outcome: The proposed model achieves state-of-the-art on 10 visual document understanding benchmarks.
mPLUG-DocOwl2: High-resolution Compressing for OCR-free Multi-page Document Understanding (2025.acl-long)

Copied to clipboard

Challenge: Multimodal Large Language Models (MLLMs) have improved document understanding performance but generate thousands of visual tokens for a single document image, leading to excessive GPU memory and slower inference times.
Approach: They propose a high-resolution document compression module to generate 324 tokens for a single document image.
Outcome: The proposed module reduces first token latency by more than 50% and improves document comprehension performance.
E2E-VLP: End-to-End Vision-Language Pre-training Enhanced by Visual Learning (2021.acl-long)

Copied to clipboard

Challenge: Existing vision-language pre-training methods use a two-step training procedure to learn visual features from image-text pairs.
Approach: They propose a vision-language pre-trained model for V+L understanding and generation using a unified Transformer framework.
Outcome: The proposed model can learn visual representation and semantic alignments between image and text on visual-text pairs and on visual processing tasks.
Unifying Latent and Lexicon Representations for Effective Video-Text Retrieval (2024.lrec-main)

Copied to clipboard

Challenge: Existing methods for video-text retrieval capture fine-grained semantic concepts . however, they lack the ability to capture finer-grain concepts such as objects and actions.
Approach: They propose a dual-encoder architecture for fast video-text retrieval that learns lexicon representations to capture fine-grained semantics.
Outcome: The proposed framework outperforms existing methods with 4.8% and 8.2% improvement on MSR-VTT and DiDeMo respectively.
UReader: Universal OCR-free Visually-situated Language Understanding with Multimodal Large Language Model (2023.findings-emnlp)

Copied to clipboard

Challenge: Existing studies for visually-situated language understanding have shown shallow zero-shot visual text recognition ability when fed a low-resolution image with salient text information.
Approach: They propose a model for universal OCR-free visually-situated language understanding based on the Multimodal Large Language Model (MLLM) their model is jointly finetuned on a wide range of visually situated language understanding tasks via a unified instruction format.
Outcome: The proposed model achieves state-of-the-art ocr-free performance in 8 out of 10 visually-situated language understanding tasks across 5 domains: documents, tables, charts, natural images, and webpage screenshots.
Semantics-enhanced Cross-modal Masked Image Modeling for Vision-Language Pre-training (2024.lrec-main)

Copied to clipboard

Challenge: Existing methods for vision-language pre-training lack high-level semantics and text is not sufficiently involved in masked modeling.
Approach: They propose a semantics-enhanced cross-modal MIM framework for vision-language representation learning that harvests high-level semantics from global image features via self-supervised agreement learning and transfers them to local patch encodings by sharing the encode space.
Outcome: The proposed model achieves state-of-the-art or competitive performance on multiple vision-language tasks.
DeepKE: A Deep Learning Based Knowledge Extraction Toolkit for Knowledge Base Population (2022.emnlp-demos)

Copied to clipboard

Challenge: Existing knowledge extraction tools are not complete due to emerging entities and relations in real-world applications.
Approach: They propose an open-source knowledge extraction toolkit DeepKE that supports low-resource, document-level and multimodal scenarios in the knowledge base population.
Outcome: The proposed toolkit supports low-resource, document-level and multimodal scenarios in the knowledge base population.
TinyChart: Efficient Chart Understanding with Program-of-Thoughts Learning and Visual Token Merging (2024.emnlp-main)

Copied to clipboard

Challenge: Recent studies have shown that multimodal large language models can be useful for chart understanding, but their size limits their use in resource-constrained environments.
Approach: They propose an efficient multimodal large language model with only 3B parameters for chart understanding.
Outcome: The proposed model outperforms several chart-understanding MLLMs with up to 13B parameters on ChartQA, Chart-to-Text, Chart to Table, OpenCQA, and ChartX.
DEMO: Reframing Dialogue Interaction with Fine-grained Element Modeling (2025.findings-acl)

Copied to clipboard

Challenge: Large language models (LLMs) enabled dialogue systems are one of the central modes in human-machine interaction.
Approach: They propose a benchmark task for dialogue element MOdeling and Element Awareness and a new benchmark for dialogue agent interaction that allows the agent to model dialogue elements via imitation learning.
Outcome: The proposed agent performs well in both dialogue element modeling and out-of-domain tasks.
AgentOCR: Reimagining Agent History via Optical Self-Compression (2026.acl-long)

Copied to clipboard

Challenge: Recent advances in large language models (LLMs) have enabled agentic systems trained with reinforcement learning over multi-turn interaction, but practical deployment is bottlenecked by rapidly growing textual histories that inflate token and memory costs.
Approach: They propose a framework that represents the accumulated observation-action history as a compact rendered image.
Outcome: The proposed framework preserves over 95% of text-based agent performance while significantly reducing token consumption (>50%), yielding consistent token and memory efficiency.
ModelScope-Agent: Building Your Customizable Agent System with Open-source Large Language Models (2023.emnlp-demo)

Copied to clipboard

Challenge: Large language models (LLMs) have demonstrated remarkable capabilities to comprehend human intentions, engage in reasoning, and design planning-like behavior.
Approach: They propose a framework that equips large language models with tool-use capabilities . they propose LLaMA and Chat-GLM as controllers, and a model-based agent framework .
Outcome: The proposed framework equips open-source LLMs with tool-use capabilities . it provides a user-friendly system library with a customizable engine design .
Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning (2023.acl-short)

Copied to clipboard

Challenge: Parameter-efficient fine-tuning only optimizes a few task-specific parameters with frozen pre-trained model.
Approach: They propose to optimize a prefix vector inserted into Transformer layers to optimize the prefix . they propose to use a gate mechanism to adjust the prefixed to each layer .
Outcome: The proposed approach improves on the SuperGLUE and NER datasets.
Transforming Visual Scene Graphs to Image Captions (2023.acl-long)

Copied to clipboard

Challenge: Existing approaches to generate captions using image captioning are based on multi-head attention (MHA)
Approach: They propose to transform scene graphs into more descriptive captions by using multi-head attention to build a Graph Neural Network (GNN) . they construct a Mixture-of-Expert (MOE)-based decoder where each expert is built on MHA for discriminating the graph embeddings to generate different kinds of words.
Outcome: The proposed framework can generate captions from multiple visual features and objects . it is based on a mixture-of-expert (MOE)-based decoder based upon MHA .
Leave No Document Behind: Benchmarking Long-Context LLMs with Extended Multi-Doc QA (2024.emnlp-main)

Copied to clipboard

Challenge: Existing benchmarks for evaluating long-context language models employ irrelevant noise texts to artificially extend the length of test cases, diverging from the real-world scenarios of long-constituency applications.
Approach: They propose a long-context benchmark, Loong, aligning with realistic scenarios through extended multi-document question answering (QA) .
Outcome: The proposed model can scale up the context window of large language models to perform in-depth analysis of multiple long documents.
EcomScriptBench: A Multi-task Benchmark for E-commerce Script Planning via Step-wise Intention-Driven Product Association (2025.acl-long)

Copied to clipboard

Challenge: Goal-oriented script planning is used by humans to plan for typical activities . however, this capability remains underexplored due to several challenges .
Approach: They propose a framework that enables product-enriched scripts by associating products with each step based on the semantic similarity between the actions and their purchase intentions.
Outcome: The proposed framework can generate product-enriched scripts from 2.4 million scripts . human annotations are conducted to provide gold labels for a sampled subset .
mPLUG: Effective and Efficient Vision-Language Learning by Cross-modal Skip-connections (2022.emnlp-main)

Copied to clipboard

Challenge: Existing pre-trained vision-language models suffer from inefficiency and linguistic signal overwhelmed by long visual sequences in cross-modal alignment.
Approach: They propose a vision-language foundation model with cross-modal skip-connections that can be pre-trained end-to-end on large-scale image-text pairs with both discriminative and generative objectives.
Outcome: The proposed model achieves state-of-the-art results on a wide range of vision-language downstream tasks, including image captioning, image-text retrieval, visual grounding and visual question answering.
Decoding Scientific Experimental Images: The SPUR Benchmark for Perception, Understanding, and Reasoning (2026.acl-long)

Copied to clipboard

Challenge: Xu and Peng, 2025) . . SPUR is a comprehensive benchmark for scientific experimental image perception, understanding, and reasoning, comprising 4,264 question-answering (QA) pairs derived from 1,084 expert-curated images.
Approach: They propose to use 4,264 question-answering (QA) pairs derived from 1,084 expert-curated images to evaluate the visual perception of multimodal large language models (MLLMs) . they also propose to utilize cross-panel relation understanding to evaluate MLLM’s ability to decipher intricate cross-panel relations.
Outcome: The proposed model is based on 4,264 question-answering pairs derived from 1,084 expert-curated images.
TRIPS: Efficient Vision-and-Language Pre-training with Text-Relevant Image Patch Selection (2022.emnlp-main)

Copied to clipboard

Challenge: Existing vision-and-language pre-training models suffer from long visual sequences . experimental results show that TRIPS gains a speedup of 40% over previous similar VLP models .
Approach: They propose an efficient vision-and-language pre-training model with text-relevant image patch selection, TRIPS, which reduces the visual sequence progressively with a text-guided patch-selection layer in the visual backbone for efficient training and inference.
Outcome: The proposed model can speed up training and inference by 40% over previous models.
Neural Topic Modeling with Bidirectional Adversarial Training (2020.acl-main)

Copied to clipboard

Challenge: Recent studies have shown that neural topic models for automatic topic extraction avoid complicated mathematical derivations for model inference.
Approach: They propose a bidirectional adversarial topic model which uses a generator and an encoder to infer topic distribution.
Outcome: The proposed model outperforms baselines and competitive models in three benchmark corpora.

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