MMErroR: A Benchmark for Erroneous Reasoning in Vision-Language Models

Yang Shi, Yifeng Xie, Minzhe Guo, Liangsi Lu, Mingxuan Huang, Jingchao Wang, Zhihong Zhu, Boyan Xu†, Zhiqi Huang

• We introduce MMErroR, a multi-modal benchmark for process-level error diagnosis in vision-language models, comprising 1,997 curated samples across 24 subdomains and four reasoning error types, revealing that even the strongest evaluated VLM achieves only 66.65% accuracy in fine-grained error-type classification.

Rose-SQL: Role-State Evolution Guided Structured Reasoning for Multi-Turn Text-to-SQL

Le Zhou, Feng Yao, Fengcai Qiao, Bo Xu, Fangyuan Wang, Boyan Xu†

• We propose Rose-SQL, a training-free framework that adapts small-scale reasoning models to multi-turn Text-to-SQL through Role-State evolution, using gain dependency analysis, structural isomorphism checks, and augmented hierarchical reasoning to track dialogue changes and guide SQL generation, achieving state-of-the-art performance on SParC and CoSQL without task-specific fine-tuning.

SAM-NER: Semantic Archetype Mediation for Zero-Shot Named Entity Recognition

Ruichu Cai, Juntao Gan, Miao Mai, Zhifeng Hao, Boyan Xu†

• We propose SAM-NER, a three-stage zero-shot NER framework that mitigates semantic drift through semantic archetype mediation, combining cooperative entity discovery, universal archetype projection, and definition-guided calibration to achieve state-of-the-art cross-domain performance on CrossNER without relying on external knowledge bases.

SERE: Structural Example Retrieval for Enhancing LLMs in Event Causality Identification

Zhifeng Hao, Zhongjie Chen, Junhao Lu, Shengyin Yu, Guimin Hu, Keli Zhang, Ruichu Cai, Boyan Xu†

• We propose SERE, a structural example retrieval framework that enhances LLMs for event causality identification by jointly leveraging conceptual path similarity, syntactic tree similarity, and causal pattern filtering to retrieve structurally aligned demonstrations, mitigating causal hallucination and improving performance across multiple ECI datasets.

GenLink: Generation-Driven Schema-Linking via Multi-Model Learning for Text-to-SQL

Zhifeng Hao, Junqi Huang, Shaobin Shi, Ruichu Cai, Boyan Xu†

• We propose GenLink, a generation-driven schema-linking framework that integrates multiple small language models to infer implicit schema links through SQL generation and self-consistency, achieving strong cross-domain Text-to-SQL performance on BIRD and Spider with execution accuracies of 67.34%, 89.7%, and 87.8%.

Chat2DB: Chatting to the Database with Interactive Agent Assisted Language Models

Boyan Xu *, Yuyuan Cai *, Shaobin Shi, Zhifeng Hao, Ruichu Cai

• We propose Chat2DB, an interactive conversational database system that enhances Text-to-SQL parsers through agent-assisted question-schema linking and adaptive retraining, enabling users to collaboratively refine schema selection, improve SQL generation accuracy, and query domain-specific databases through natural language.

Handling Missing Entities in Zero-Shot Named Entity Recognition: Integrated Recall and Retrieval Augmentation

Ruichu Cai, Junhao Lu, Zhongjie Chen, Boyan Xu†, Zhifeng Hao

• We introduce IRRA, a novel two-stage framework that first uses recall-augmented entity extraction on perturbed data to boost recall and then applies retrieval-augmented generation for type correction, achieving state-of-the-art zero-shot cross-domain NER performance.

Track-SQL: Enhancing Generative Language Models with Dual-Extractive Modules for Schema and Context Tracking in Multi-turn Text-to-SQL

Bingfeng chen, Shaobin Shi, yongqi luo, Boyan Xu†, Ruichu Cai, Zhifeng Hao

• We propose Track-SQL, a dual-extractive framework that augments generative LMs with a semantic-enhanced schema extractor and a schema-aware context extractor to track schema and context changes in multi-turn Text-to-SQL, achieving state-of-the-art results on SparC and CoSQL with execution accuracy gains of 7.1% and 9.55%.

S^2IT: Stepwise Syntax Integration Tuning for Large Language Models in Aspect Sentiment Quad Prediction

Bingfeng chen, Chenjie Qiu, Yifeng Xie, Boyan Xu†, Ruichu Cai, Zhifeng Hao

• We propose S^2IT, a novel Stepwise Syntax Integration Tuning framework that progressively incorporates global and local syntactic structure knowledge into LLMs to significantly enhance performance on Aspect Sentiment Quad Prediction.

Dr.ECI: Infusing Large Language Models with Causal Knowledge for Decomposed Reasoning in Event Causality Identification

Ruichu Cai, Shengyin Yu, Jiahao Zhang, Wei Chen, Boyan Xu†, Keli Zhang

• We propose Dr.ECI, a multi-agent decomposed reasoning framework for Event Causality Identification that employs specialized discovery agents (Causal Explorer, Mediator Detector) and reasoning agents (Direct and Indirect Reasoners) to capture implicit, indirect, and generalized causal structures, achieving state-of-the-art performance.

CACA: Context-Aware Cross-Attention Network for Extractive Aspect Sentiment Quad Prediction

Bingfeng Chen, Haoran Xu, Yongqi Luo, Boyan Xu†, Ruichu Cai, Zhifeng Hao

• We propose CACA, an extractive ASQP framework that employs a Context-Aware Cross-Attention Network—alternating updates of explicit and implicit representations—and contrastive learning to align aspects and opinions for implicit term prediction, achieving superior results on three benchmarks.

S^2GSL: Incorporating Segment to Syntactic Enhanced Graph Structure Learning for Aspect-based Sentiment Analysis

Bingfeng Chen, Qihan Ouyang, Yongqi Luo, Boyan Xu†, Ruichu Cai, Zhifeng Hao (ACL 2024 Main)

• We introduce S^2GSL, a novel graph-structure learning framework for ABSA that combines segment-aware semantic graph learning with syntax-based latent graph learning—and a self-adaptive aggregation network to filter irrelevant contexts and fuse diverse structures—achieving superior results on four benchmarks.