Papers by Xinyue Li

21 papers
Temporal Token Matters: Investigating and Interpreting the Consistency of Temporal Ordering in Large Language Models (2026.findings-acl)

Copied to clipboard

Challenge: Large Language Models (LLMs) exhibit notable deficiencies in temporal reasoning . phrasing changes can lead LLMs to produce inconsistent outputs .
Approach: They investigate the mechanistic interpretability of temporal ordering within event temporal reasoning . they identify a sparse subset of attention heads that are causally responsible for reasoning outcomes .
Outcome: The proposed model outperforms other models in a variety of tasks and is validated by intervention-based experiments.
Automating eHMI Action Design with LLMs for Automated Vehicle Communication (2025.findings-emnlp)

Copied to clipboard

Challenge: Currently, eHMIs employ predefined text messages and manually designed actions to perform these messages . this limits the real-world deployment of ehMIs, where adaptability in dynamic scenarios is essential.
Approach: They propose a pipeline that integrates large language models and 3D renderers to generate executable actions for controlling eHMIs and rendering action clips.
Outcome: The proposed pipeline integrates large language models and 3D renderers to generate executable actions for controlling eHMIs and rendering action clips.
PSSAT: A Perturbed Semantic Structure Awareness Transferring Method for Perturbation-Robust Slot Filling (2022.coling-1)

Copied to clipboard

Challenge: Existing slot filling models memorize inherent patterns of entities and contexts from training data.
Approach: They propose a perturbed semantic structure awareness transferring method for slot filling models . they use two MLM-based training strategies to learn contextual semantic structure and word distribution .
Outcome: The proposed method outperforms existing methods and gains strong generalization while preventing model from memorizing inherent patterns of entities and contexts.
LLMSurgeon: Diagnosing Data Mixture of Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: a lack of transparency in large language models makes auditing their "digital DNA" difficult.
Approach: They propose a framework that casts DMS as an inverse problem under label-shift assumption . they propose LLMScan, a recipe-verifiable evaluation suite built from open-source LLMs .
Outcome: The proposed framework casts DMS as an inverse problem under label-shift assumption . compared with existing frameworks, it recovers domain mixtures with high fidelity .
See2Refine: Vision-Language Feedback Improves LLM-Based eHMI Action Designers (2026.acl-long)

Copied to clipboard

Challenge: External Human-Machine Interfaces (eHMIs) are emerging as promising solutions to address this communication gap.
Approach: They propose a framework that uses vision-language models (VLMs) for perceptual evaluation as automated visual feedback to improve an LLM-based eHMI action designer.
Outcome: The proposed framework outperforms prompt-only LLM designers and manually specified baselines in three eHMI modalities and multiple LLM model sizes.
EduBench: A Comprehensive Benchmarking Dataset for Evaluating Large Language Models in Diverse Educational Scenarios (2026.acl-long)

Copied to clipboard

Challenge: Existing benchmarks that focus on knowledge-intensive tasks do not reflect diverse educational scenarios.
Approach: They propose a benchmark that incorporates 9 major scenarios and 4,000 educational contexts.
Outcome: The proposed model performs comparable to state-of-the-art large models on the test set.
Constructing Your Model’s Value Distinction: Towards LLM Alignment with Anchor Words Tuning (2025.findings-emnlp)

Copied to clipboard

Challenge: a study of large language models (LLMs) shows that they can generate outputs that are honest, positive, harmless, etc.
Approach: They propose a method that amplifies logits difference between positive and negative tokens . they propose to use the logits gap to generate positive and positive tokens after alignment .
Outcome: The proposed method achieves effective alignment, but requires fewer computational resources compared to training-time alignment methods.
ModSCAN: Measuring Stereotypical Bias in Large Vision-Language Models from Vision and Language Modalities (2024.emnlp-main)

Copied to clipboard

Challenge: Large vision-language models have been widely used but stereotypical biases are unexplored.
Approach: They propose a framework to SCAN stereotypical bias within large vision-language models . they examine stereotype biases with respect to gender and race in three scenarios .
Outcome: The proposed framework can reduce stereotypical biases in large vision-language models . the currently popular models show significant stereotype biase .
Linear Steerability in Language Models: When It Emerges and How It Evolves (2025.findings-emnlp)

Copied to clipboard

Challenge: a new framework for steering language models reveals how concepts become linearly separable as training progresses .
Approach: They propose a framework to analyze steerability in language models by using hidden state and representation analysis.
Outcome: The proposed framework reveals how steerability evolves over training . concepts become linearly separable as training progresses, the framework shows .
The Art of (Mis)alignment: How Fine-Tuning Methods Effectively Misalign and Realign LLMs in Post-Training (2026.findings-acl)

Copied to clipboard

Challenge: Misaligned large language models can magnify harm by exploiting them to undermine safety . et al., 2022b; Bai e.t., 2023): misalignment, realignment and model-specific resistance are important .
Approach: They evaluate four methods to identify a mechanism asymmetry between attack and defense . they find that ORPO is most effective for misalignment, but DPO excels in realignment .
Outcome: The proposed methods show a mechanism asymmetry between attack and defense . the proposed methods excel in realignment, but at the expense of model utility .
MUCAR: Benchmarking Multilingual Cross-Modal Ambiguity Resolution for Multimodal Large Language Models (2025.emnlp-main)

Copied to clipboard

Challenge: Existing multimodal benchmarks overlook linguistic and visual ambiguities, authors say . ambiguity resolution between modalities is lacking in multimodal large language models .
Approach: They propose a benchmark to evaluate multimodal ambiguity resolution across multilingual and cross-modal scenarios.
Outcome: a new benchmark evaluates multimodal ambiguity resolution across multilingual and cross-modal scenarios . the benchmark shows that MLLMs can resolve ambiguities in image-text alignment . however, existing benchmarks often overlook linguistic and visual ambiguties .
How Do LLMs and VLMs Understand Viewpoint Rotation Without Vision? An Interpretability Study (2026.acl-long)

Copied to clipboard

Challenge: Existing studies on spatial intelligence from the perspective of visual-spatial intelligence have not explored whether visual intelligence alone is sufficient to endow models with spatial intelligence.
Approach: They propose to use a linguistic perspective to investigate spatial intelligence from a theoretical perspective.
Outcome: The proposed model performs poorly on the proposed dataset while human can easily achieve 100% accuracy.
Open Schrödinger’s Closed Box: Identifying Retrieval Augmented Generation in API-Accessible Large Language Model Services (2026.acl-long)

Copied to clipboard

Challenge: Large language models (LLMs) are powerful at question-answering but prone to hallucinations due to limited domain-specific or up-to-date knowledge.
Approach: They propose a framework for IDentifying RAG properties in LLM services that integrates LLMs with retrieval systems and adds an external retriever and knowledge database to mitigate hallucinations.
Outcome: The proposed framework detects RAG-enhanced LLMs with 99.97% accuracy with partial or no optional knowledge and nearly 100% when the LLM and database are known.
RAPO: An Adaptive Ranking Paradigm for Bilingual Lexicon Induction (2022.emnlp-main)

Copied to clipboard

Challenge: Existing approaches focus on minimizing distances between words in aligned pairs, while suffering from low discriminative capability to distinguish the relative orders between positive and negative candidates.
Approach: They propose a ranking-oriented induction model to learn personalized mapping function for each word.
Outcome: The proposed model can learn personalized mapping function for each word on public datasets including rich-resource and low-resourced languages.
Think in Safety: Unveiling and Mitigating Safety Alignment Collapse in Multimodal Large Reasoning Model (2025.emnlp-main)

Copied to clipboard

Challenge: Existing Large Reasoning Models have demonstrated broad application potential, yet their safety and reliability remain critical concerns.
Approach: They conduct a safety evaluation of 13 MLRMs across 5 benchmarks and examine their safety performance.
Outcome: The proposed model improves safety on jailbreak and safety-awareness benchmarks.
AIRepr: An Analyst-Inspector Framework for Evaluating Reproducibility of LLMs in Data Science (2025.findings-emnlp)

Copied to clipboard

Challenge: Large language models are increasingly used to automate data analysis, but data science tasks often admit multiple statistically valid solutions.
Approach: They propose a framework to evaluate LLM-generated code and assess its reproducibility . they introduce two reproducibility-enhancing prompting strategies and benchmark them against standard prompting .
Outcome: The proposed framework improves reproducibility of large language models . it provides a foundation for transparent, reliable, and efficient human–AI collaboration in data science.
Multi-task Learning of Spoken Language Understanding by Integrating N-Best Hypotheses with Hierarchical Attention (2020.coling-industry)

Copied to clipboard

Challenge: Existing methods to integrate hypotheses into speech recognition systems are noisy and can cause information loss.
Approach: They propose to integrate hypotheses into multi-task learning and transfer learning to improve performance.
Outcome: The proposed model improves domain and intent classification by 19% and 37% compared to current methods . the proposed model could recover transcription and rewrite the query for a better understanding .
Can LLMs See Without Pixels? Benchmarking Spatial Intelligence from Textual Descriptions (2026.findings-acl)

Copied to clipboard

Challenge: Existing advances in Spatial Intelligence rely on vision-Language Models . however, a critical question remains: does spatial understanding originate from visual encoders?
Approach: They propose to evaluate the SI performance of Large Language Models without pixel-level input.
Outcome: The proposed benchmark challenges large language models to perform symbolic reasoning rather than visual pattern matching.
Faithful-First Reasoning, Planning, and Acting for Multimodal LLMs (2026.findings-acl)

Copied to clipboard

Challenge: Existing efforts to improve task accuracy or enrich COT generation are lacking in multimodal large language models.
Approach: They propose a Faithful-First Reasoning, Planning, and Acting framework that evaluates faithfulness of intermediate reasoning and uses it to plan and execute faithfulness-aware actions during inference.
Outcome: The proposed framework improves perceptual faithfulness by up to 24% over prompt-based and tool-augmented reasoning frameworks without degrading task accuracy.
How Numerical Precision Affects Arithmetical Reasoning Capabilities of LLMs (2025.findings-acl)

Copied to clipboard

Challenge: Despite the success of transformer-based large language models, understanding and enhancing their mathematical capabilities remains a significant challenge.
Approach: They propose to use numerical precision as a key factor that influences LLMs' effectiveness in arithmetical tasks to determine their effectiveness.
Outcome: The proposed models perform better in arithmetic tasks than transformer-based models with standard numerical precision.
Knowledge Injection Exists in MoE? Exploring Expert-Aware Contrast Decoding in MoE for Mitigating LLMs’ Hallucinations (2026.acl-long)

Copied to clipboard

Challenge: Existing methods to mitigate hallucinations include prompt engineering and model optimization, but lack domain generalization and potential errors in fine-tuning data may exacerbate the hallucism.
Approach: They propose an expert-aware adaptive contrast decoding that uses expert differences in MoE’s higher layers to mitigate hallucinations on QA tasks.
Outcome: The proposed method outperforms baseline models on four datasets Large language models (LLMs) show strong performance but suffer from hallucinations, limiting their application.

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