Papers by Han Qiu

35 papers
Course-Correction: Safety Alignment Using Synthetic Preferences (2024.emnlp-industry)

Copied to clipboard

Challenge: Recent studies show that large language models generate harmful content, but the potential for generating harmful content is an escalating concern.
Approach: They propose to fine-tune LLMs with preference learning to emphasize the preference for timely course-correction by using an automated pipeline.
Outcome: The proposed model improves course-correction skills without affecting general performance and resists jailbreak attacks.
Embodied Executable Policy Learning with Language-based Scene Summarization (2024.naacl-long)

Copied to clipboard

Challenge: Existing Large Language models with text inputs lack the capability to evolve with non-expert interactions with environments.
Approach: They propose a novel learning paradigm that generates robots’ executable actions in the form of text, derived solely from visual observations.
Outcome: The proposed learning paradigm surpasses baselines and can adapt to the target tasks effectively.
The Side Effects of Being Smart: Safety Risks in MLLMs’ Multi-Image Reasoning (2026.acl-long)

Copied to clipboard

Challenge: Recent advances in multimodal reasoning may pose new safety risks . evaluators neglect reasoningbased safety, where harm emerges only through MLLMs .
Approach: They introduce a benchmark for multi-image reasoning safety that includes 2,676 instances . they find that models with more advanced multi- image reasoning are more vulnerable .
Outcome: The proposed benchmark consists of 2,676 instances covering 9 multi-image relations . the results show that models with more advanced multi- image reasoning are more vulnerable .
LeakDojo: Decoding the Leakage Threats of RAG Systems (2026.findings-acl)

Copied to clipboard

Challenge: Existing studies have failed to assess RAG leakage risks for large language models . constructing and maintaining highquality RAG knowledge databases has become increasingly costly .
Approach: They propose a framework for controlled evaluation of RAG leakage using query generation and adversarial instructions.
Outcome: The proposed framework compares six existing attacks across fourteen LLMs, four datasets, and diverse RAG systems.
Meta-Learning Adversarial Domain Adaptation Network for Few-Shot Text Classification (2021.findings-acl)

Copied to clipboard

Challenge: Existing approaches for few-shot text classification rely on exploitation of lexical features and distributional signatures on training data, while neglecting to strengthen the model's ability to adapt to new tasks.
Approach: They propose a meta-learning framework integrated with an adversarial domain adaptation network to improve the model's adaptive ability and generate high-quality text embedding for new classes.
Outcome: The proposed framework outperforms the state-of-the-art models on four datasets and shows clear superiority over existing models.
“I’ve Decided to Leak”: Probing Internals Behind Prompt Leakage Intents (2025.emnlp-main)

Copied to clipboard

Challenge: Large language models (LLMs) exhibit prompt leakage vulnerabilities, raising intellectual property and confidentiality concerns.
Approach: They use probing techniques to capture LLMs’ intent-related internal representations and show that they internalize prompt leakage intents in their hidden states before generating tokens.
Outcome: The proposed probes achieve 90%+ AUROC across all tested models, even when applied to new system prompts and attacks.
pEBR: A Probabilistic Approach to Embedding Based Retrieval (2025.emnlp-industry)

Copied to clipboard

Challenge: Existing embedding-based retrieval systems rely on heuristic and suboptimal cutoffs for item retrieval.
Approach: They propose a probabilistic Embedding-Based Retrieval framework that learns a shared semantic representation space for both queries and items.
Outcome: The proposed framework improves retrieval precision and recall, and ablation studies show it captures the differences between head-to-tail queries.
Walking in Others’ Shoes: How Perspective-Taking Guides Large Language Models in Reducing Toxicity and Bias (2024.emnlp-main)

Copied to clipboard

Challenge: Existing prompting methods that require white-box access to the model or substantial training fail to simultaneously lessen toxicity and bias.
Approach: They propose a strategy that encourages LLMs to integrate diverse human perspectives and self-regulate their responses by incorporating diverse human viewpoints.
Outcome: The proposed approach can significantly diminish toxicity (up to 89%) and bias (up 73%) in LLMs’ responses.
Multi-Agent Collaboration via Cross-Team Orchestration (2025.findings-acl)

Copied to clipboard

Challenge: Large Language Models (LLMs) have significantly impacted various domains, especially through organized LLM-driven autonomous agents.
Approach: They propose a framework that enables orchestrated teams to jointly propose various task-oriented solutions and interact with their insights in a self-independence while cross-team collaboration environment for superior solutions generation.
Outcome: Experiments show that the framework can generate better software quality compared to state-of-the-art frameworks.
LASA: Language-Agnostic Semantic Alignment at the Semantic Bottleneck for LLM Safety (2026.acl-long)

Copied to clipboard

Challenge: Large language models (LLMs) have demonstrated better safety performance in high-resource languages than in low-resourced languages.
Approach: They propose language-agnostic semantic alignment (LASA) which anchors safety alignment directly in semantic bottlenecks.
Outcome: The proposed approach significantly improves safety across all languages: average attack success rate drops from 24.7% to 2.8% on LLaMA-3.1-8B-Instruct and remains within 3–4% across Qwen2.5 and Qwend3 Instruct models (7B–32B).
Can Language Models Follow Multiple Turns of Entangled Instructions? (2025.findings-emnlp)

Copied to clipboard

Challenge: Despite of significant achievements in improving instruction-following capabilities of large language models, the ability to process multiple potentially entangled or conflicting instructions remains a considerable challenge.
Approach: They construct multi-turn instruction with 1.1K high-quality multi-turned conversations using the human-in-the-loop approach and examine their capabilities.
Outcome: The proposed model shows that it is difficult to integrate multiple turns and balance competing objectives when instructions intersect or conflict.
SynSetExpan: An Iterative Framework for Joint Entity Set Expansion and Synonym Discovery (2020.emnlp-main)

Copied to clipboard

Challenge: Entity set expansion and synonym discovery are two critical NLP tasks that are often performed separately, without exploring their interdependencies.
Approach: They propose a framework that enables two tasks to mutually enhance each other by including popular entities’ infrequent synonyms into the set, which boosts set expansion recall.
Outcome: The proposed framework can be used to enhance two NLP tasks by including popular entities’ infrequent synonyms into the set, which boosts set expansion recall.
PersLEARN: Research Training through the Lens of Perspective Cultivation (2023.acl-demo)

Copied to clipboard

Challenge: PersLEARN is a tool designed to facilitate the cultivation of scientific perspectives . junior researchers struggle to identify the perspectives reflected in the literature and struggle to develop their own viewpoints.
Approach: They propose a tool to facilitate the cultivation of scientific perspectives by interacting with a prompt-based model and allowing students to develop their own perspectives explicitly.
Outcome: The proposed tool outperforms baseline approaches across multiple domains of literature from different perspectives.
An MRC Framework for Semantic Role Labeling (2022.coling-1)

Copied to clipboard

Challenge: Existing work on semantic role labels ignores the semantic connection between the two tasks . et al. (2010) defined two types of semantic roles: core roles and non-core roles.
Approach: They propose to use machine reading comprehension to bridge the gap between these two tasks . they formalize predicate disambiguation as multiple-choice machine reading understanding .
Outcome: The proposed framework achieves state-of-the-art or comparable results to previous work . it uses the descriptions of candidate senses of a given predicate as options to select the correct sense .
Speculating LLMs’ Chinese Training Data Pollution from Their Tokens (2025.emnlp-main)

Copied to clipboard

Challenge: Experiments on GPT and other 23 LLMs indicate that tokens widely exist while GPT’s vocabulary behaves the worst: more than 23% long Chinese tokens (i.e., a token with more than two Chinese characters) are either porn or online gambling.
Approach: They propose to locate Polluted Chinese (PoC) tokens in LLMs and build a PoC token detector to label them in vocabularies by considering each token’s semantics and related contents from the search engines.
Outcome: The proposed method predicts that the ratio of “*” related webpages in GPT-4o's training data is around 0.5%.
Exploring Multimodal Challenges in Toxic Chinese Detection: Taxonomy, Benchmark, and Findings (2025.findings-acl)

Copied to clipboard

Challenge: Recent studies show that character substitutions in toxic Chinese text can confuse state-of-the-art LLMs.
Approach: They propose a taxonomy of 3 perturbation strategies and 8 specific approaches in Chinese text to assess if they can detect perturbed Chinese toxic contents.
Outcome: The proposed model can detect perturbed Chinese text with 8 different approaches . the proposed model is compared with 9 other LLMs from the US and China .
Arithmetic Control of LLMs for Diverse User Preferences: Directional Preference Alignment with Multi-Objective Rewards (2024.acl-long)

Copied to clipboard

Challenge: Reinforcement Learning from Human Feedback (RLHF) relies on scalar rewards to capture user preferences.
Approach: They propose a framework that integrates multi-objective reward modeling to represent diverse preference profiles.
Outcome: The proposed method improves performance across reward objectives and targets.
TOXIFRENCH: Benchmarking and Enhancing Language Models via CoT Fine-Tuning for French Toxicity Detection (2026.findings-acl)

Copied to clipboard

Challenge: toxicity detection in French remains underdeveloped due to the lack of culturally relevant, human-annotated, large-scale datasets.
Approach: They propose a method that generalizes French online comments using a semi-automated annotation pipeline that reduces manual labeling to only 10% through high-confidence LLM-based pre-annotation and human verification.
Outcome: The proposed model outperforms GPT-4o and DeepSeek-R1 on the benchmark while maintaining cross-lingual capabilities.
Understanding the Dark Side of LLMs’ Intrinsic Self-Correction (2025.acl-long)

Copied to clipboard

Challenge: Recent studies show that LLMs’ intrinsic self-correction fails without oracle labels as feedback.
Approach: They propose to use one simple task and three complex tasks with state-of-the-art LLMs like ChatGPT, Llama, and DeepSeek to interpret LLM's intrinsic self-correction.
Outcome: The proposed methods reveal the dark side of LLMs’ intrinsic self-correction for different tasks, especially for those failure cases.
GLIMPSE: Do Large Vision-Language Models Truly Think With Videos or Just Glimpse at Them? (2025.emnlp-main)

Copied to clipboard

Challenge: Existing video benchmarks often resemble image-based questions with scans of only a few key frames, without deep temporal reasoning.
Approach: They propose a video benchmark to assess whether large vision-language models can genuinely think with videos rather than perform superficial frame-level analysis.
Outcome: The proposed benchmark consists of 3,269 videos and over 4,342 highly visual-centric questions across 11 categories, including Trajectory Analysis, Temporal Reasoning, and Forensics Detection.
Failure makes the agent stronger: Enhancing Accuracy through Structured Reflection for Reliable Tool Interactions (2026.findings-acl)

Copied to clipboard

Challenge: Existing approaches to self-reflection rely on heuristic prompting or unidirectional reasoning traces.
Approach: They propose a structured reflection method that transforms the "from error to repair" process into a first-class, controllable, and trainable action.
Outcome: The proposed method improves multi-turn tool-call success rates and error recovery while reducing redundant calls.
When Audio and Text Disagree: Revealing Text Bias in Large Audio-Language Models (2025.emnlp-main)

Copied to clipboard

Challenge: Large Audio-Language Models (LALMs) are augmented with the ability to perceive audio, but their reliability when faced with conflicting inputs remains largely unexplored.
Approach: They examine how LALMs prioritize information when presented with inconsistent audio-text pairs.
Outcome: The proposed models display a significant bias toward textual input when presented with inconsistent audio-text pairs.
TaxoClass: Hierarchical Multi-Label Text Classification Using Only Class Names (2021.naacl-main)

Copied to clipboard

Challenge: Hierarchical multi-label text classification (HMTC) aims to assign each text document to a set of relevant classes from a taxonomy.
Approach: They propose to conduct HMTC based on only class surface names as supervision signals to mimic human experts.
Outcome: The proposed framework outperforms the best existing method by 25% on two challenging datasets.
Shoes-ACOSI: A Dataset for Aspect-Based Sentiment Analysis with Implicit Opinion Extraction (2024.findings-emnlp)

Copied to clipboard

Challenge: Prior work in ABSA has investigated opinion extraction as an important subtask, but these works only label concise, *explicitly*-stated opinion spans.
Approach: They propose a new ABSA dataset with implicit opinion span annotations . they use paragraph-length inputs and prompted-LLM baselines to evaluate the dataset .
Outcome: The proposed dataset presents significant challenges for fully-supervised models and LLMs.
Manual Evaluation Matters: Reviewing Test Protocols of Distantly Supervised Relation Extraction (2021.findings-acl)

Copied to clipboard

Challenge: Distantly supervised relation extraction (RE) has attracted much attention in the past few years . previous methods to evaluate models manually or directly on autolabeled data have produced inaccurate evaluations .
Approach: They propose to use distant supervision to generate large-scale autolabeled data . they build manually-annotated test sets for two DS-RE datasets and evaluate models .
Outcome: The proposed method produces 53% wrong labels at the entity pair level in the popular NYT10 dataset.
Think While You Write: Hypothesis Verification Promotes Faithful Knowledge-to-Text Generation (2024.findings-naacl)

Copied to clipboard

Challenge: Knowledge-to-text generators often struggle to faithfully generate descriptions for input facts . we propose a decoding-only method to reduce hallucinations .
Approach: They propose a decoding-only method to generate accurate descriptions for input facts . they use a Natural Language Inference model as the model and replace it with a task-specific HVM .
Outcome: The proposed method improves faithfulness with minimal impact on quality and in/out-of-distribution evaluations.
PKU-SafeRLHF: Towards Multi-Level Safety Alignment for LLMs with Human Preference (2025.acl-long)

Copied to clipboard

Challenge: Using large-scale annotation data, large language models can generate noise, errors and biases, leading to unexpected behaviours.
Approach: They propose a dataset to promote safety alignment in large language models . they separate helpfulness and harmlessness annotations for question-answering pairs .
Outcome: The proposed dataset provides 44.6k prompts and 265k question-answer pairs with safety meta-labels for 19 harm categories and three severity levels, with answers generated by Llama-family models.
Revisiting the Reliability of Language Models in Instruction-Following (2026.acl-long)

Copied to clipboard

Challenge: Several benchmarks have been proposed to measure instruction-following accuracy, but these scores do not translate to reliable services in real-world use.
Approach: They propose a new metric reliable@k and develop an automated pipeline to generate cousin prompts.
Outcome: The proposed model can be instantiated with cousin prompts and generates high-quality cousin prompt data.
Eliciting In-context Retrieval and Reasoning for Long-context Large Language Models (2025.findings-acl)

Copied to clipboard

Challenge: Existing benchmarks like LOFT often overestimate LCLM performance by providing overly simplified contexts.
Approach: They propose to use retrieval-attention-probing to filter and de-noise long contexts during decoding and joint retrieval head training alongside the generation head to improve LCLM performance.
Outcome: The proposed approach outperforms RAG and GPT-4-Turbo on most tasks despite being a much smaller model.
New Terms, New Toxicity: Consensus-based Chinese Neologism Toxicity Detection via Search-Augmented LLMs (2026.acl-long)

Copied to clipboard

Challenge: Neologisms can foster new linguistic consensus by stabilizing shared meanings and usage in common communicative norms.
Approach: They propose a taxonomy that captures the origins and consensus-verification criteria of toxic neologisms . they propose 'SeTox' framework that integrates real-time web context for naeologim detection .
Outcome: The proposed framework outperforms large-scale models in detecting neologism toxicity.
Transfer Knowledge from Natural Language to Electrocardiography: Can We Detect Cardiovascular Disease Through Language Models? (2023.findings-eacl)

Copied to clipboard

Challenge: Recent advances in Large Language Models (LLMs) have shown powerful ability in various downstream applications.
Approach: They propose an approach for cardiovascular disease diagnosis and automatic ECG diagnosis report generation.
Outcome: The proposed approach generates high-quality cardiac diagnosis reports and achieves competitive zero-shot classification performance even compared with supervised baselines.
Can Brain Signals Reveal Inner Alignment with Human Languages? (2023.findings-emnlp)

Copied to clipboard

Challenge: Brain Signals, such as Electroencephalography, and human languages have been explored independently for many downstream tasks, however, the connection between them has not been well explored.
Approach: They introduce a multimodal transformer alignment model to observe coordinated representations between EEG and language.
Outcome: The proposed method achieved an F1-score improvement of 1.7% on ZuCo and 9.3% on Zuco datasets for sentiment analysis, and 7.4% on ZuCO for relation detection.
IsOBS: An Information System for Oracle Bone Script (2020.emnlp-demos)

Copied to clipboard

Challenge: Oracle bone script (OBS) documents are the oldest continuously-used writing system in the world and are important for linguistic and historical research.
Approach: They construct an information system for OBS to symbolize, serialize, and store OBS data at the character-level using efficient databases and retrieval modules.
Outcome: The proposed system symbolizes, serializes, and stores OBS data at the character-level, based on efficient databases and retrieval modules.
Reward Generalization in RLHF: A Topological Perspective (2025.findings-acl)

Copied to clipboard

Challenge: Existing alignment methods share a common topology of information flow, but their alternatives have not been thoroughly explored.
Approach: They propose a theory of reward generalization in reinforcement learning from human feedback . they propose induced Bayesian networks to model the impact of dataset topologies on reward generalisation .
Outcome: The proposed method achieves an average win rate of 65% on three NLP tasks.

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