Papers by Yujun Cai

19 papers
GraphCache: Message Passing as Caching for Sentence-Level Relation Extraction (2022.findings-naacl)

Copied to clipboard

Challenge: Existing work only encodes entity types and textual context within individual instances, which limits the performance of sentence-level relation extraction (RE).
Approach: They propose a module that aggregates the features from sentences to learn global representations of properties and augments local features within individual sentences.
Outcome: The proposed module can learn global representations of properties from sentences and augment local features within individual sentences.
Mapping the Minds of LLMs: A Graph-Based Analysis of Reasoning LLMs (2025.emnlp-main)

Copied to clipboard

Challenge: Large Reasoning Models (LRMs) often display unstable behaviors, e.g., hallucinating unsupported premises, overthinking simple tasks, and displaying higher sensitivity to prompt variations.
Approach: They propose a graph-based analytical framework that clusters long, verbose CoT outputs into semantically coherent reasoning steps, then constructs directed reasoning graphs to capture contextual and logical dependencies among these steps.
Outcome: The proposed framework enables quantitative evaluation of internal reasoning structure and quality beyond conventional metrics and provides practical insights for prompt engineering and cognitive analysis of LLMs.
ReCon: Active Defense against Large Vision-Language Model Jailbreaks via Reverse Safety Concept Injection (2026.findings-acl)

Copied to clipboard

Challenge: Existing defense strategies neglect visual threats and lack of fine-grained specificity regarding specific attack semantics.
Approach: They propose a black-box defense framework that maps unsafe concepts to fine-grained, constructive Safe Concepts.
Outcome: a new black-box defense framework enhances robustness against jailbreak attacks . it maps detected unsafe concepts to fine-grained, constructive Safe Concepts . the proposed framework is available for free at http://www.epa.org/recon/ .
Understanding GUI Agent Localization Biases through Logit Sharpness (2025.findings-emnlp)

Copied to clipboard

Challenge: Multimodal large language models often exhibit hallucinations that compromise reliability . despite promising performance, these models often display systematic localization errors .
Approach: They propose a framework that categorizes model predictions into four distinct types . they propose metric that evaluates alignment between semantic continuity and logits distribution .
Outcome: The proposed framework categorizes model predictions into four different types . it reveals nuanced failure modes beyond traditional accuracy metrics .
Primacy Effect of ChatGPT (2023.emnlp-main)

Copied to clipboard

Challenge: Existing machine learning models may lead to poor performance in discriminative natural language understanding tasks.
Approach: They propose to use ChatGPT to query large amounts of human-written text to find the answer to a question.
Outcome: The proposed model has a high chance to select labels at earlier positions as the answer.
VistaWise: Building Cost-Effective Agent with Cross-Modal Knowledge Graph for Minecraft (2025.emnlp-main)

Copied to clipboard

Challenge: Large language models (LLMs) have shown significant promise in embodied decision-making tasks within virtual open-world environments, but lack domain-specific knowledge.
Approach: They propose a cost-effective agent framework that integrates cross-modal domain knowledge and finetunes a dedicated object detection model for visual analysis.
Outcome: The proposed framework reduces the requirement for domain-specific training data from millions of samples to a few hundred.
Should We Rely on Entity Mentions for Relation Extraction? Debiasing Relation Extraction with Counterfactual Analysis (2022.naacl-main)

Copied to clipboard

Challenge: Existing studies rely on entity information for sentence-level relation extraction (RE) but this can leak superficial and spurious clues of relations.
Approach: They propose to use entity mentions to extract relations from textual context . they use a causal graph to model dependencies between variables in RE models .
Outcome: The proposed method yields significant gains on both effectiveness and generalization for RE.
MRFD: Multi-Region Fusion Decoding with Self-Consistency for Mitigating Hallucinations in LVLMs (2025.findings-emnlp)

Copied to clipboard

Challenge: Large Vision-Language Models often produce hallucinations due to the limited ability to verify information in different regions of the image.
Approach: a new decoding method improves factual grounding by modeling inter-region consistency . the method identifies salient regions using cross-attention and generates initial responses for each .
Outcome: a training-free decoding method reduces hallucinations and improves response consistency . the proposed method generates initial responses for each region and weights reliability weights among responses .
SemVink: Advancing VLMs’ Semantic Understanding of Optical Illusions via Visual Global Thinking (2025.emnlp-main)

Copied to clipboard

Challenge: Vision-language models excel in semantic tasks but fail at detecting hidden content . current architectures prioritize abstract reasoning over low-level visual operations .
Approach: They propose a benchmark to test vision-language models that can detect hidden content . they propose HC-Bench to scale images to low resolutions to unlock 99% accuracy .
Outcome: HC-Bench shows that leading VLMs achieve near-zero accuracy even with explicit prompting . et al.: current models prioritize abstract reasoning over low-level visual operations . they urge a shift toward hybrid models bridging gap between computational vision and human cognition .
Vulnerability of LLMs to Vertically Aligned Text Manipulations (2025.acl-long)

Copied to clipboard

Challenge: Recent research shows that vertical text input significantly degrades the accuracy of large language models (LLMs) in text classification tasks.
Approach: They investigate the impact of vertical text input on the performance of LLMs . they find that chain of thought reasoning does not help LLM recognize vertical input .
Outcome: The proposed model can significantly mislead models, posing a risk of bypassing detection in real-world scenarios involving harmful or sensitive information.
Dynamic Infilling Anchors for Format-Constrained Generation in Diffusion Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: Diffusion large language models (dLLMs) offer bidirectional attention and parallel generation . fixed anchors can enforce constraints, but they often impose rigid spans, leading to truncated reasoning .
Approach: They propose a method that dynamically estimates end-anchor positions to adjust generation length before iterative infilling.
Outcome: The proposed method improves format compliance and answer accuracy on GSM8K and MATH.
Tricking Retrievers with Influential Tokens: An Efficient Black-Box Corpus Poisoning Attack (2025.naacl-long)

Copied to clipboard

Challenge: Existing methods for crafting adversarial passages are slow and computationally expensive, requiring either access to retriever’s gradients or large computational resources.
Approach: They propose a method that leverages two key characteristics of retrievers: insensitivity to token order and bias towards influential tokens to generate effective adversarial passages.
Outcome: The proposed method achieves superior efficiency and scalability compared to existing methods while maintaining comparable or better attack success rates across multiple datasets.
Con-ReCall: Detecting Pre-training Data in LLMs via Contrastive Decoding (2025.coling-main)

Copied to clipboard

Challenge: Existing methods analyze training data with member and non-member contexts, overlooking potential insights from both member and not-member.
Approach: They propose a method that leverages asymmetric distributional shifts induced by member and non-member contexts through contrastive decoding to enhance membership inference.
Outcome: The proposed approach outperforms the current state-of-the-art on the WikiMIA benchmark and is robust against various text manipulation techniques.
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.
DiMo-GUI: Advancing Test-time Scaling in GUI Grounding via Modality-Aware Visual Reasoning (2025.emnlp-main)

Copied to clipboard

Challenge: DiMo-GUI is a training-free framework for GUI grounding that splits input into textual elements and iconic elements, allowing the model to reason over each modality independently using general-purpose vision-language models.
Approach: They propose a training-free framework for GUI grounding that leverages two core strategies: dynamic visual grounding and modality-aware optimization.
Outcome: The proposed framework splits the input into textual elements and iconic elements, allowing the model to reason over each modality independently using general-purpose vision-language models.
DRS: Deep Question Reformulation With Structured Output (2025.findings-acl)

Copied to clipboard

Challenge: Existing models like GPT-3 and Instruct-GPT lack the ability to reformulate unanswerable questions.
Approach: They propose a zero-shot method that combines the strengths of LLMs with a DFS-based algorithm to iteratively explore potential entity combinations and constrain outputs using predefined entities.
Outcome: The proposed method outperforms all baselines, including the GPT-3.5 model, on the unanswerable question reformulation task.
Beyond Query Bias: Candidate-Aware Iterative Refinement for Zero-Shot Composed Image Retrieval (2026.findings-acl)

Copied to clipboard

Challenge: Existing methods to retrieve target images suffer from inherent cognitive bias due to unknown candidate distribution.
Approach: They propose a training-free framework that reframes ZS-CIR as a self-correcting process . they propose to use retrieved results as feedback to perceive the candidate distribution .
Outcome: Experiments on public benchmarks show that CoRR outperforms other SOTA methods.
Mitigating Coordinate Prediction Bias from Positional Encoding Failures (2026.findings-acl)

Copied to clipboard

Challenge: Multimodal Large Language Models (MLLMs) excel at general vision-language tasks, but precise coordinate prediction remains a challenge.
Approach: They propose a training-free, inference-time correction method to correct VPEs . they isolate position-unconditioned tendencies by shuffling VPE and use it to steer digit decoding .
Outcome: The proposed method is training-free, inference-time correction method . it effectively rectifies coordinate drift, yielding consistent improvements without retraining .
VideoStir: Understanding Long Videos via Spatio-Temporally Structured and Intent-Aware RAG (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for retrieval-augmented generation (RAG) to long videos are limited by limited context windows and flatten videos into independent segments.
Approach: They propose a structured and intent-aware long-video RAG framework that structures a video as a spatio-temporal graph and then performs multi-hop retrieval to aggregate evidence across distant yet contextually related events.
Outcome: The proposed framework is competitive with state-of-the-art baselines without auxiliary information.

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