Papers by Guanhua Zhang

25 papers
The Elephant in the Room: Exploring the Role of Neutral Words in Language Model Group-Agnostic Debiasing (2025.findings-acl)

Copied to clipboard

Challenge: Large Language Models (LLMs) are increasingly integrated into our daily lives, raising ethical concerns, especially about perpetuating stereotypes.
Approach: They propose a method that incorporates a neutral word semantics-based loss function to alleviate the deterioration of the LMS during debiasing.
Outcome: The proposed method alleviates the deterioration of the Language Modeling Score (LMS) by incorporating a neutral word semantics-based loss function.
Zero-Shot Cross-Lingual Transfer of Neural Machine Translation with Multilingual Pretrained Encoders (2021.emnlp-main)

Copied to clipboard

Challenge: Existing work on improving cross-lingual transferability of NMT model is under-explored.
Approach: They propose a model that leverages a multilingual pretrained encoder to improve cross-lingual transferability.
Outcome: The proposed model outperforms mBART and m2m-100 on a zero-shot cross-lingual transfer task.
Tag-Instruct: Controlled Instruction Complexity Enhancement through Structure-based Augmentation (2025.findings-acl)

Copied to clipboard

Challenge: High-quality instruction data is crucial for developing large language models (LLMs), yet existing approaches struggle to effectively control instruction complexity.
Approach: They propose a framework that compresses instructions into a compact tag space and enhances complexity through RL-guided tag expansion.
Outcome: The proposed framework outperforms existing methods in the evaluation of instruction complexity augmentation and semantic compression of text into a compact tag space.
Are AI-Generated Text Detectors Robust to Adversarial Perturbations? (2024.acl-long)

Copied to clipboard

Challenge: Existing detectors for AI-generated text lack robustness against adversarial perturbations, with even minor changes in characters or words causing a reversal in distinguishing between human-created and AI-generated text.
Approach: They propose a siamese calibration technique to train the model to make equally confident predictions under different noise, which improves the model’s robustness against adversarial perturbations.
Outcome: The proposed detector outperforms baseline methods on four datasets and is more generalizable in cross-domain, cross-genre, and mixed-source scenarios.
LLMs Trust Humans More, That’s a Problem! Unveiling and Mitigating the Authority Bias in Retrieval-Augmented Generation (2025.acl-long)

Copied to clipboard

Challenge: Large language models (LLMs) generate outputs that stray from user input or contravene established knowledge.
Approach: They propose a new phenomenon, Authority Bias, where LLMs favor one knowledge source over the other . they propose atomic information that generates conflicts and a Conflict Detection Enhanced Query framework .
Outcome: The proposed framework reduces Authority bias in large language models . it detects conflicts, performs credibility assessment on conflicting paragraphs, and detects perturbed text .
From Word to World: Can Large Language Models be Implicit Text-based World Models? (2026.acl-long)

Copied to clipboard

Challenge: Agentic learning increasingly hinges on interaction, yet real-world experience is expensive, limited, and often irreversible at inference time.
Approach: They propose a framework that reframes language modeling as next-state prediction under interaction.
Outcome: The proposed framework evaluates world models in text-based environments . it shows that sufficiently trained models capture coherent environment dynamics .
PlanGPT: Enhancing Urban Planning with a Tailored Agent Framework (2025.acl-industry)

Copied to clipboard

Challenge: Empirical tests demonstrate that PlanGPT framework has achieved advanced performance, providing comprehensive support that significantly enhances professional planning efficiency.
Approach: They propose a specialized AI agent framework tailored for urban and spatial planning that integrates a customized local database retrieval system and domain-specific knowledge activation capabilities.
Outcome: Empirical tests show that PlanGPT framework significantly improves planning efficiency . it integrates a customized database retrieval system, domain-specific knowledge activation capabilities, and advanced tool orchestration mechanisms.
Learning to Think on Hypergraph: HyperCoT for Structure-Guided N-ary Knowledge Graph Completion (2026.acl-long)

Copied to clipboard

Challenge: Existing methods to solve knowledge hypergraph link prediction problem are limited by their ability to generate chain-of-thought (CoT) representations.
Approach: They propose a structure-aware approach that models multi-hop structural reasoning as a depth-sensitive progressive evidence accumulation process.
Outcome: Experiments on three real-world datasets show that HyperCoT outperforms strong n-ary KGC baselines while yielding interpretable multi-hop reasoning traces.
Reinforcement Learning on Pre-Training Data (2026.acl-long)

Copied to clipboard

Challenge: Recent progress in large language models is driven by scaling of training compute through pre-training with nexttoken prediction (NTP) or post-training (RL) Pre-training using NTP enables models to acquire extensive knowledge and skills from general data, but it suffers from data inefficiency and catastrophic forgetting in continual learning settings.
Approach: They propose to scale training compute through pre-training with next-token prediction (NTP) or post-training by scaling reinforcement learning (RL) to improve learning from general data.
Outcome: Experiments on multiple benchmarks and models show that the proposed approach improves continual pre-training and provides a strong foundation for post-training on Qwen3-8B-Base.
Can LLMs Hear the Dogwhistle? (2026.findings-acl)

Copied to clipboard

Challenge: Existing safety benchmarks focus on explicitly harmful content, but ignore context-dependent expressions such as dogwhistles.
Approach: They propose a benchmark for evaluating LLM safety under dogwhistle-driven prompts . their findings expose a blind spot in current safety evaluation practices .
Outcome: The proposed benchmark compared safety performance with toxic terms using dogwhistle-driven prompts.
No More Stale Feedback: Co-Evolving Critics for Open-World Agent Learning (2026.acl-long)

Copied to clipboard

Challenge: Current methods for training Large Language Model agents rely on static or offline critic models, which fail to adapt as the policy evolves.
Approach: They propose a framework that integrates a critique and a policy to optimize the policy and critic through a synchronized co-evolutionary loop.
Outcome: The proposed framework yields more stable training and higher long-horizon task success across open-world environments.
VFA: Empowering Multilingual MLLMs via Vision-Free Adaptation (2026.acl-long)

Copied to clipboard

Challenge: Multimodal large language models have advanced rapidly, yet most remain English-centric . scaling multilingual multimodal instruction tuning is limited by the scarcity and high cost of non-English image–text supervision.
Approach: They propose a framework that decouples multilingual language enhancement from visual alignment by composing complementary task vectors over a shared LLM backbone.
Outcome: The proposed framework achieves competitive performance with a fully multimodally trained model using less than 2% of the text data.
Selection Bias Explorations and Debias Methods for Natural Language Sentence Matching Datasets (P19-1)

Copied to clipboard

Challenge: Natural Language Sentence Matching (NLSM) is a popular NLP task.
Approach: They propose to use QuoraQP to train and evaluate NLSM models using a selection bias framework.
Outcome: The proposed framework can improve generalization ability of trained models and give more trustworthy evaluation results for real-world adoptions.
FinSafetyBench: Evaluating LLM Safety in Real-World Financial Scenarios (2026.findings-acl)

Copied to clipboard

Challenge: Existing large language models (LLMs) are prone to misuse and misinformation, posing serious compliance risks.
Approach: They propose a bilingual red-teaming benchmark to test an LLM’s refusal of requests that violate financial compliance.
Outcome: The proposed benchmark is based on real-world financial crime cases and ethical violations and includes 14 subcategories covering financial crimes and ethical breaches.
Advancing the Robustness of Large Language Models through Self-Denoised Smoothing (2024.naacl-short)

Copied to clipboard

Challenge: Existing adversarial attacks can cause LLMs to make wrong predictions on downstream tasks or generate harmful content misaligned with human values.
Approach: They propose to use randomized smoothing to add noise to the input and then make predictions based on these denoised versions.
Outcome: The proposed method surpasses existing methods in both empirical and certified robustness in defending against adversarial perturbations for both downstream tasks and human alignments (i.e., jailbreak attacks).
Ex3: Automatic Novel Writing by Extracting, Excelsior and Expanding (2024.acl-long)

Copied to clipboard

Challenge: Generating long-term texts using artificial intelligence has always been a challenge . however, the generated novels exhibit poor logical coherence and appeal in their plots and deficiencies in character and event depiction, ultimately compromising the overall narrative quality.
Approach: They propose a method for extracting excelsior and expanding from novel data to generate arbitrarily long novels using large language models.
Outcome: The proposed method produces high-quality long-form novels with a high level of logical coherence and appeal despite the use of large language models.
InstructDiff: Domain-Adaptive Data Selection via Contrastive Entropy for Efficient LLM Fine-Tuning (2026.acl-long)

Copied to clipboard

Challenge: Existing data selection methods suffer from severe domain specificity . existing methods for general instruction-following fail on reasoning tasks .
Approach: They propose a framework that operationalizes contrastive entropy as a domain-adaptive selection criterion through warmup calibration, bi-directional NLL filtering, and entropic-based ranking.
Outcome: Experiments show that InstructDiff outperforms baseline training on reasoning tasks while using only 10% of the data.
Self-DC: When to Reason and When to Act? Self Divide-and-Conquer for Compositional Unknown Questions (2025.naacl-long)

Copied to clipboard

Challenge: Existing studies focus on leveraging internal knowledge of Large Language Models (LLMs) to answer known questions.
Approach: They propose a framework that allows LLMs to choose between internal and external knowledge . they use a dataset to analyze compositional questions that are composed of unknown sub-questions .
Outcome: The proposed framework can achieve comparable or even better performance with much fewer external calls compared with several strong baselines.
PlanGPT-VL: Enhancing Urban Planning with Domain-Specific Vision-Language Models (2025.emnlp-industry)

Copied to clipboard

Challenge: Existing Vision-Language Models (VLMs) fail to analyze planning maps . specialized visual representations of land use zones, transportation networks, and development policies are needed to interpret complex planning maps.
Approach: They propose a domain-specific VLM tailored for urban planning maps that employs three innovations: PlanAnno-V framework for high-quality VQA data synthesis, Critical Point Thinking (CPT) and PlanBench-V benchmark for systematic evaluation.
Outcome: The new model outperforms general-purpose VLMs on planning map interpretation tasks.
Towards Making the Most of Cross-Lingual Transfer for Zero-Shot Neural Machine Translation (2022.acl-long)

Copied to clipboard

Challenge: Existing unsupervised neural machine translation systems can degrade when labeled data is limited.
Approach: They propose a multilingual pretraining and multilingual fine-tuning for facilitating cross-lingual transfer in zero-shot translation using a parallel dataset.
Outcome: The proposed model outperforms state-of-the-art models on many-to-English translation by over 7.2 and 5.0 BLEU.
FANNO: Augmenting High-Quality Instruction Data with Open-Sourced LLMs Only (2025.findings-acl)

Copied to clipboard

Challenge: Recent studies explore approaches to synthesize instruction data with open-sourced LLMs but require high-quality human-crafted seed data.
Approach: They propose an end-to-end framework to synthesize high-quality instruction data with open-sourced LLMs and sampled unlabeled documents, eliminating the need for seed data.
Outcome: The proposed framework synthesizes high-quality instruction data with open-sourced LLMs and sampled unlabeled documents, eliminating the need for seed data.
PaSa: An LLM Agent for Comprehensive Academic Paper Search (2025.acl-long)

Copied to clipboard

Challenge: We introduce PaSa, an advanced Paper Search agent powered by large language models . despite being trained on synthetic data, PaSA outperforms existing baselines on RealScholarQuery .
Approach: They introduce PaSa, an advanced Paper Search agent powered by large language models . they optimize PaSA using a synthetic dataset, AutoScholarQuery, which includes 35k fine-grained queries .
Outcome: The paper analyzes the performance of a paper search agent using a synthetic dataset . it significantly outperforms existing benchmarks on RealScholarQuery .
Demographics Should Not Be the Reason of Toxicity: Mitigating Discrimination in Text Classifications with Instance Weighting (2020.acl-main)

Copied to clipboard

Challenge: Recent research has found that text classification datasets contain certain unintended biases, such as text containing demographic identity-terms that are more likely to be abusive.
Approach: They propose a model-agnostic debiasing framework that recovers the non-discrimination distribution using instance weighting, which does not require extra resources or annotations apart from a pre-defined set of demographic identity-terms.
Outcome: The proposed framework alleviates the unintended biases without hurting models’ generalization ability.
LayAlign: Enhancing Multilingual Reasoning in Large Language Models via Layer-Wise Adaptive Fusion and Alignment Strategy (2025.findings-naacl)

Copied to clipboard

Challenge: Large language models (LLMs) are pretrained on multilingual corpora but exhibit suboptimal performance on low-resource languages.
Approach: They propose a framework that integrates representations from all encoder layers and an adaptive fusion-enhanced attention mechanism to enable layer-wise interaction between the LLM and the multilingual encoder.
Outcome: Experiments on multilingual reasoning tasks show that the proposed framework outperforms baselines.
Rhombus: Incentivizing Coordination in Parallel Thinking through Reinforcement Learning (2026.findings-acl)

Copied to clipboard

Challenge: Parallel thinking is a promising avenue for scaling test-time compute in Large Language Models . however, coordinating the exploration and aggregation stages remains challenging .
Approach: They propose a parallel thinking framework that explicitly incentivizes coordination between components via end-to-end reinforcement learning.
Outcome: The proposed framework improves accuracy by 6.0% over long chain-of-thought baselines while reducing wall-clock latency by 39.4% under matched token budgets.

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