Papers by Yun Zhao

19 papers
Tracing the Roots: A Multi-Agent Framework for Uncovering Data Lineage in Post-Training LLMs (2026.acl-long)

Copied to clipboard

Challenge: High-quality post-training data is the primary engine driving LLM capabilities . datasets are often treated as isolated artifacts, overlooking their true developmental context .
Approach: They propose a framework to reconstruct the evolutionary graph of dataset development using data lineage.
Outcome: The proposed framework characterizes domain-specific structural patterns in Math-oriented datasets and general-domain corpora.
Advancing Vision-Language Models with Adapter Ensemble Strategies (2024.findings-emnlp)

Copied to clipboard

Challenge: CLIP revolutes vision-language pretraining by using contrastive learning on paired web data.
Approach: They propose to combine a "adapter ensemble" with traditional machine learning techniques to augment large-scale pretrained vision-language models.
Outcome: The proposed model outperforms baselines and derives improvement when the number of ensemble parameters increases.
Encoding Spreadsheets for Large Language Models (2024.emnlp-main)

Copied to clipboard

Challenge: Spreadsheets are characterized by their extensive two-dimensional grids, flexible layouts, and varied formatting options, which pose significant challenges for large language models (LLMs).
Approach: They propose a structural-anchor-based compression, inverse index translation, and data-format-aware aggregation module to compress spreadsheets effectively.
Outcome: The proposed method outperforms the existing model in GPT4 and achieves a state-of-the-art 78.9% F1 score.
TRAM: Benchmarking Temporal Reasoning for Large Language Models (2024.findings-acl)

Copied to clipboard

Challenge: despite advances in natural language processing, temporal reasoning is still a challenge . despite advancements in NLP, current language models have yet to reach human level in this domain .
Approach: They propose a temporal reasoning benchmark that measures time-related temporal aspects of events . they evaluate popular LLMs like GPT-4 and Llama2 in zero-shot and few-shot scenarios .
Outcome: The proposed model outperforms human models in a few-shot and zero-shot scenarios . the best-performing model lags significantly behind human models, the authors say .
CondenseFlow: Scalable Latent Space Collaboration via Semantic Compression for Multi-Agent Systems (2026.findings-acl)

Copied to clipboard

Challenge: Full-state latent communication in LLMs suffers from memory overhead scaling linearly with collaboration rounds.
Approach: They propose a lightweight module that uses learnable semantic probes to compress KV caches into fixed-size representations.
Outcome: The proposed module reduces KV cache memory by over 99% and inference latency by approximately 20% on seven benchmarks spanning six models . it outperforms text-based methods by 1.7 percentage points on average across all configurations while outperforming existing methods by 1.7%.
Composite Backdoor Attacks Against Large Language Models (2024.findings-naacl)

Copied to clipboard

Challenge: Large language models (LLMs) have demonstrated superior performance on various tasks, but untrustworthy third-party LLMs may covertly introduce vulnerabilities for downstream tasks.
Approach: They propose a composite backdoor attack that scatters multiple trigger keys in different prompt components.
Outcome: The proposed attack achieves 100% Attack Success Rate (ASR) with a False Triggered Rate (FTR) below 2.06% and negligible model accuracy degradation.
ToM: Leveraging Tree-oriented MapReduce for Long-Context Reasoning in Large Language Models (2025.emnlp-main)

Copied to clipboard

Challenge: Experimental results show ToM outperforms existing divide-and-conquer frameworks . RAG relies on similarity-based rankings to retrieve and reason over chunks based on logical coherence .
Approach: They propose a Tree-oriented MapReduce framework for long-context reasoning . it leverages the hierarchical structure of long documents by constructing a DocTree .
Outcome: Experimental results show that ToM outperforms existing divide-and-conquer frameworks and RAGs . the proposed framework improves logical coherence and long-context reasoning on 70B+ LLMs compared to existing approaches .
Metacognitive Prompting Improves Understanding in Large Language Models (2024.naacl-long)

Copied to clipboard

Challenge: Recent advances in prompting have enhanced reasoning in logic-intensive tasks for LLMs, yet the nuanced understanding abilities of these models remain underexplored.
Approach: They propose a strategy inspired by human introspective reasoning processes to enhance LLMs' understanding abilities.
Outcome: The proposed method outperforms chain-of-thought prompting and its advanced versions on ten natural language understanding (NLU) datasets.
PharmMT: A Neural Machine Translation Approach to Simplify Prescription Directions (2020.findings-emnlp)

Copied to clipboard

Challenge: a novel machine translation-based approach to simplify prescription directions is proposed . the language used by physicians and health professionals includes medical jargon and implicit directives .
Approach: They propose a machine translation-based approach to automatically and reliably simplify prescription directions into patient-friendly language.
Outcome: The proposed system achieves a BLEU score of 60.27 over 530K prescriptions from a large mail-order pharmacy.
CLeVeR: Multi-modal Contrastive Learning for Vulnerability Code Representation (2025.findings-acl)

Copied to clipboard

Challenge: Existing methods for detecting code capture the overall semantics of the code rather than its intrinsic vulnerability-specific semantics.
Approach: They propose an approach that leverages contrastive learning to generate precise vulnerability code representations under the supervision of vulnerability descriptions.
Outcome: The proposed approach outperforms state-of-the-art methods in vulnerability detection tasks by 11.85% and 13.61%.
DORM: Preference Data Weights Optimization for Reward Modeling in LLM Alignment (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing approaches to align large language models with human preferences are noisy and varying in importance of preference samples.
Approach: a new method enhances reward modeling by learning to dynamically weigh preference data.
Outcome: a new method improves the performance of large language models with human preferences . it initializes data importance and iteratively refines them to maximize validation performance.
Representation-Guided Parameter-Efficient LLM Unlearning (2026.findings-acl)

Copied to clipboard

Challenge: Existing methods to unlearning large language models often memorize sensitive or harmful information, but they struggle with the forget-retain trade-off due to the polysemantic nature of LLMs parameters.
Approach: They propose a representation-guided low-rank unlearning approach that leverages the geometric properties of representation spaces to achieve robust and precise unlearning.
Outcome: The proposed approach outperforms state-of-the-art models on TOFU and WMDP benchmarks while maintaining higher model utility.
LexicalAT: Lexical-Based Adversarial Reinforcement Training for Robust Sentiment Classification (D19-1)

Copied to clipboard

Challenge: Existing text classification models are fragile and sensitive to simple perturbations.
Approach: They propose a generator-classifier adversarial training approach to improve classification models . they use a large-scale lexical knowledge base to generate attacking examples .
Outcome: The proposed approach outperforms strong baselines and reduces test errors on neural networks.
From Parameters to Performance: A Data-Driven Study on LLM Structure and Development (2025.emnlp-main)

Copied to clipboard

Challenge: Large language models have revolutionized a wide range of domains, driving significant advancements in both technology and real-world applications.
Approach: They present a large-scale dataset encompassing diverse open-source LLM structures and their performance across multiple benchmarks.
Outcome: The proposed model validates the relationship between structural configurations and performance across multiple benchmarks and further corroborates the findings using mechanistic interpretability techniques.
Beyond Static Rules: Automated Discovery of Latent Vulnerabilities in Text-to-SQL (2026.findings-acl)

Copied to clipboard

Challenge: Large Language Models (LLMs) have been successful in Text-to-SQL tasks, but their deployment in real-world environments is hindered by latent reliability issues.
Approach: They propose a framework to autonomously uncover latent failure patterns in LLM-based Text-to-SQL generation.
Outcome: The proposed framework uncovers a substantial number of failure cases on state-of-the-art open-source LLMs.
SportQA: A Benchmark for Sports Understanding in Large Language Models (2024.naacl-long)

Copied to clipboard

Challenge: SportQA is a benchmark specifically designed for evaluating Large Language Models (LLMs) sports knowledge is characterized by its fast pace, variety of types, abundance of strategies, and rich player narratives .
Approach: They propose a benchmark specifically designed for evaluating Large Language Models in the context of sports understanding.
Outcome: The proposed benchmark aims to bridge the gap between existing and specialized benchmarks in sports understanding.
MEGen: Generative Backdoor into Large Language Models via Model Editing (2025.findings-acl)

Copied to clipboard

Challenge: Existing methods for training large language models are limited to yes-or-no discriminative tasks, leading users to underestimate the potential risks.
Approach: They propose an editing-based generative backdoor that expands the backdoor to generative tasks in a unified format of any text-to-any text.
Outcome: The proposed model achieves high attack success rate by adjusting only a small set of local parameters with few-shot samples.
CoViPAL: Layer-wise Contextualized Visual Token Pruning for Large Vision-Language Models (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing methods to prune redundant vision tokens struggle in shallow layers due to the lack of contextual information.
Approach: They propose a layer-wise contextualized visual token pruning method that uses a plug-and-play Pruning Module to prune redundant vision tokens.
Outcome: The proposed method outperforms training-free pruning methods under equal token budgets and surpasses training based methods with comparable supervision.

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