Dr. Yushun Dong’s RAI Lab lands six papers at EMNLP 2026
Department of Computer Science
Dr. Yushun Dong’s research team (RAI Lab) has had six papers accepted at the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP), including four papers in the Main Conference and two papers in Findings. The accepted work spans trustworthy and secure large language models, multimodal hallucination mitigation, reliable LLM-based evaluation, reasoning distillation, and large language models for finance, two of which come from the undergrad research intern students. Notably, RAI Lab take the training of undergrad research interns seriously: we provide thoughtful and comprehensive academic guidance to research interns who are deeply committed to their research. Many outstanding undergrad research intern alumni in RAI Lab have published top-tier research papers—often their very first top-tier research publications in their academic career!
An Embarrassingly Simple Detector for Model Extraction Attacks in Large Language Model API Traffic
This paper is led by his PhD student Shuze Liu. This work studies a practical security challenge for hosted large language model APIs: detecting model extraction attacks when individual malicious queries can look benign. The paper formulates extraction monitoring as benign-calibrated traffic-window distribution testing and introduces a simple detector based on maximum mean discrepancy (MMD) over semantic query embeddings. The detector requires only historical benign traffic for calibration and is evaluated across fourteen attacker–normal query pairs from four extraction scenarios. It achieves a 0.3% benign false-positive rate, 100.0% true-positive rate on pure attacker traffic, and strong performance under mixed multi-user traffic, providing a simple and deployment-oriented baseline for protecting LLM services from model extraction.
Let Them Steal: Trapping Large Language Model Extraction Attacks with Knowledge Honeypot
This paper is led by his PhD student Yuyang Dai. This work introduces Knowledge Trap, a new defense that redirects model extraction attacks toward low-transferability knowledge rather than simply blocking suspicious queries or degrading model outputs. The method constructs a Honeypot Knowledge Graph and uses breadcrumb-guided exploration to consume an attacker’s limited query budget on knowledge with negligible downstream utility. Experiments across medical, financial, and legal domains show that the approach reduces surrogate-model agreement while preserving legitimate-user performance, suggesting a new direction for defending LLM intellectual property by manipulating an attacker’s knowledge-space traversal.
From Rubrics to Reliable Scores: Evidence-Grounded Text Evaluation with LLM Judges
This paper is led by an undergrad research intern student under his supervision. The work addresses the reliability of rubric-based text evaluation with large language models and introduces RULERS, an inference-time framework for translating human rubric intent into a stable and auditable scoring protocol. RULERS fixes a task-level rubric specification, grounds judgments in structured evidence, and calibrates model signals to human score boundaries. Across four rubric-governed benchmarks and multiple frozen LLM backbones, the framework improves agreement with human scores in most evaluated settings, better matches empirical score distributions, and remains more stable under semantically equivalent changes to rubric wording.
Self-Correction Can Amplify Hallucinations: Fact-Level Repair with Graph-Based Evidence Routing in Multimodal Generation
This paper is led by a PhD student under his supervision. The paper studies a failure mode of multimodal self-correction: feedback generated jointly from the input and an already hallucinated output can reinforce unsupported claims rather than repair them. It introduces TIGER, an inference-time framework that independently extracts an observation graph from the input and a claim graph from the generated output, ranks individual claims by graph-conditioned support and conflict risk, and locally repairs high-risk facts while keeping the backbone frozen. Experiments across image-to-text, image-plus-text-to-text, audio-to-text, and video-to-text generation show that TIGER reduces unsupported content while preserving task quality, with additional evidence from a multi-source crisis-reporting case study.
FinanceLLM: A Survey of Large Language Models in Finance
This paper is led by an undergrad research intern student under his supervision. This survey provides a structured view of the rapidly growing literature on large language models in finance. It introduces a unified taxonomy that organizes financial LLM approaches by computational complexity, proposes a three-stage evaluation framework spanning laboratory benchmarking, human evaluation, and production readiness, and connects deployment approaches to operational financial contexts. The paper also highlights open challenges including validation, data contamination, and operational compliance, offering a practical framework for understanding how financial LLM research can move from academic evaluation toward real-world use.
TeacherGRPO: Closing the Capacity Gap in Reasoning Distillation via Teacher Alignment
This paper is led by a PhD student under his supervision. The work studies the capacity mismatch that arises when increasingly powerful teacher models produce reasoning distributions that smaller student models cannot effectively approximate. It proposes Teacher Alignment, which adapts the teacher toward the student’s distribution before distillation, and introduces TeacherGRPO, a reinforcement-learning approach based on Group Relative Policy Optimization. Its curriculum selective alignment and importance-adaptive length regularization are designed to preserve high-value reasoning while reducing the teacher–student distribution gap. Extensive experiments across reasoning benchmarks and distillation methods show consistent improvements over existing approaches.
Together, these six papers reflect the RAI Lab’s broad research agenda in responsible and trustworthy AI, with contributions spanning LLM security, reliable evaluation, multimodal generation, efficient reasoning, and domain-specific language-model applications.