Papers by Defu Lian

30 papers
What Makes a Good Reasoning Chain? Uncovering Structural Patterns in Long Chain-of-Thought Reasoning (2025.emnlp-main)

Copied to clipboard

Challenge: Recent advances in reasoning with large language models have popularized Long Chain-of-Thought (LCoT) a framework that converts sequential LCoTs into hierarchical tree structures enables deeper structural analysis of LLM reasoning.
Approach: They propose a framework that converts sequential LCoTs into hierarchical tree structures and enables deeper structural analysis of LLM reasoning.
Outcome: The proposed framework can be used to analyze LLM reasoning in a variety of tasks and models.
On the Vulnerability of Safety Alignment in Open-Access LLMs (2024.findings-acl)

Copied to clipboard

Challenge: Large language models (LLMs) are susceptible to malicious exploitation, but are often rejected and limited harmfulness is limited.
Approach: They propose two types of reverse alignment techniques: reverse supervised fine-tuning (RSFT) and reverse preference optimization (RPO).
Outcome: The proposed methods can significantly enhance the success rate and harmfulness of jailbreak attacks, but they face high rejection rates and limited harmfulness.
SPARD: Self-Paced Curriculum for RL Alignment via Integrating Reward Dynamics and Data Utility (2026.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) are shifting the focus from single verifiable tasks toward complex, open-ended real-world scenarios.
Approach: They propose a framework that automatically adjusts reward weights and data importance to synchronize learning intent with data utility for optimal performance.
Outcome: The proposed framework improves model capabilities across all domains and scales.
Evaluating Readability and Faithfulness of Concept-based Explanations (2024.emnlp-main)

Copied to clipboard

Challenge: Existing methods for evaluating concepts from different perspectives lack a unified formalization.
Approach: They propose a formal definition of concepts generalizing to diverse concept-based explanations’ settings and apply it to other types of explanations or tasks.
Outcome: Extensive experimental analysis was carried out to determine the evaluation measures for explanation evaluation measures.
M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation (2024.findings-acl)

Copied to clipboard

Challenge: In this paper, we introduce a new embedding model for semantic retrieval of more than 100 working languages.
Approach: They propose a new embedding model that supports multi-lingual, cross-lingual and long-document retrieval . they propose integrating relevance scores from different retrieval functionalities into the teacher signal .
Outcome: The proposed model exhibits superior performance on multilingual, cross-lingual, and long-document retrieval benchmarks.
Mitigating the Language Mismatch and Repetition Issues in LLM-based Machine Translation via Model Editing (2024.emnlp-main)

Copied to clipboard

Challenge: Existing studies have focused on using Large Language Models to improve translation quality . language mismatch and repetition are two of the main problems with LLMs .
Approach: They propose to leverage model editing methods to reduce language mismatch and repetition . they propose to fetch intersections of locating results under different language settings .
Outcome: The proposed methods reduce language mismatch and repetition ratios and enhance translation quality in most cases.
MegaPairs: Massive Data Synthesis for Universal Multimodal Retrieval (2025.acl-long)

Copied to clipboard

Challenge: despite the growing demand for multimodal retrieval, there is a lack of training data.
Approach: They propose a data synthesis method that leverages vision language models and open-domain images to generate high-quality data.
Outcome: The proposed method outperforms baseline models on 70 more datasets and can scale up.
ReasonEmbed: Enhanced Text Embeddings for Reasoning-Intensive Document Retrieval (2026.acl-long)

Copied to clipboard

Challenge: Recent studies suggest that traditional retrievers struggle with reasoningintensive tasks such as personal assistants and scientific research.
Approach: They propose a new data synthesis method that overcomes the triviality problem prevalent in previous synthetic datasets and propose 'ReMixer', a data fusion method that generates 82K high-quality training samples.
Outcome: The proposed model outperforms existing models on reasoning-intensive retrieval tasks.
Mitigate Negative Transfer with Similarity Heuristic Lifelong Prompt Tuning (2024.findings-acl)

Copied to clipboard

Challenge: Existing approaches to improve the performance of lifelong prompt tuning are limited due to inherent challenges.
Approach: They propose a similarity heuristic lifelong prompt tuning framework that partitions tasks into two distinct subsets by harnessing a learnable similarity measure.
Outcome: The proposed framework outperforms state-of-the-art techniques in lifelong learning benchmarks and demonstrates robustness against negative transfer in diverse task sequences.
Matching-oriented Embedding Quantization For Ad-hoc Retrieval (2021.emnlp-main)

Copied to clipboard

Challenge: Product quantization (PQ) is a widely used technique for ad-hoc retrieval.
Approach: They propose a match-oriented product quantization with a multinoulli contrastive loss objective.
Outcome: The proposed method maximizes matching probability of query and ground-truth key, compared with previous methods on non-supervised datasets.
Internal Value Alignment in Large Language Models through Controlled Value Vector Activation (2025.acl-long)

Copied to clipboard

Challenge: Existing LLMs do not possess consistent values, but many have been developed to align them at the behavioral level, including supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF).
Approach: They propose a Controlled Value Vector Activation method that directly aligns the internal values of Large Language Models by interpreting how a value is encoded in their latent representations.
Outcome: The proposed method achieves highest success rate across 10 basic values without hurting model performance and fluency, and ensures target values even with opposite and potentially malicious input prompts.
Towards Anytime Fine-tuning: Continually Pre-trained Language Models with Hypernetwork Prompts (2023.findings-emnlp)

Copied to clipboard

Challenge: Continual pre-training has been used for a multitude of domains and tasks . a continually pre-trained model can show a non-decreasing performance on unseen domains .
Approach: They propose a method that generates domain-specific prompts by agreement and disagreement losses.
Outcome: The proposed method achieves improvements of 3.57% and 3.4% on two real-world datasets.
Investigating Data Variance in Evaluations of Automatic Machine Translation Metrics (2022.findings-acl)

Copied to clipboard

Challenge: Current evaluation methods focus on one dataset, e.g., Newstest dataset in each year’s WMT Metrics Shared Task.
Approach: They propose to use a single dataset to evaluate the performance of automatic translation metrics.
Outcome: The results show that the rankings of metrics vary when the evaluation is conducted on different datasets.
Learning to Substitute Spans towards Improving Compositional Generalization (2023.acl-long)

Copied to clipboard

Challenge: despite the rising prevalence of neural sequence models, there is a deficiency in compositional generalization.
Approach: They propose a compositional augmentation strategy that enables multi-grained composition of substructures in the whole training set.
Outcome: The proposed strategy outperforms existing strategies on three compositional generalization benchmarks.
Learning from Emptiness: De-biasing Listwise Rerankers with Content-Agnostic Probability Calibration (2026.acl-short)

Copied to clipboard

Challenge: Existing methods for listwise reranking exhibit intrinsic position bias . existing methods are constrained by an inherent trade-off between efficiency and flexibility .
Approach: They propose a training-free framework that mechanically decouples positional bias from ranking decisions.
Outcome: a training-free framework decouples position bias from ranking decisions . evaluations show it outperforms training-based methods and outperformed expensive methods .
Let Retrievers Think Before Action: Thought-Augmented Embedding for Dense Retrieval (2026.findings-acl)

Copied to clipboard

Challenge: Large language models have demonstrated that explicit step-by-step thinking can substantially improve performance on complex tasks.
Approach: They propose a model that generates preliminary thoughts for input queries before document retrieval.
Outcome: The proposed model generates preliminary thoughts for input queries before document retrieval.
Understanding and Patching Compositional Reasoning in LLMs (2024.findings-acl)

Copied to clipboard

Challenge: LLMs have marked a revolutonary shift, yet they falter when faced with compositional reasoning tasks.
Approach: They propose a lightweight method to patch compositional reasoning errors via editing the located MHSA modules in LLMs.
Outcome: The proposed method can be used to patch compositional reasoning errors using MHSA modules located within the layers of the LLMs.
Llama2Vec: Unsupervised Adaptation of Large Language Models for Dense Retrieval (2024.acl-long)

Copied to clipboard

Challenge: Dense retrieval requires discriminative embeddings to represent the semantic relationship between query and document.
Approach: They propose an unsupervised approach that performs unsupervised adaptation of large language models for dense retrieval.
Outcome: The proposed model improves on a variety of dense retrieval benchmarks and is available on github.
Visualizing the Relationship Between Encoded Linguistic Information and Task Performance (2022.findings-acl)

Copied to clipboard

Challenge: Recent studies show that encoding more syntactic information does not lead to better performance.
Approach: They propose a method to optimize pareto-optimal models by formalizing it as a multi-objective optimization problem.
Outcome: The proposed method is better than a baseline method on two NLP tasks.
Assessing Dialogue Systems with Distribution Distances (2021.findings-acl)

Copied to clipboard

Challenge: Existing evaluation metrics focus on turnlevel quality, which is not well suited for open-end dialogue tasks.
Approach: They propose to measure the performance of a dialogue system by computing the distributionwise distance between its generated conversations and real-world conversations.
Outcome: The proposed metrics correlate better with human judgments than existing metrics on dialogue systems.
AIR-Bench: Automated Heterogeneous Information Retrieval Benchmark (2025.acl-long)

Copied to clipboard

Challenge: Evaluation benchmarks based on predefined domains and human-labeled data face limitations in addressing evaluation needs for emerging domains.
Approach: They propose an automated information retrieval benchmark based on predefined domains and human-labeled data . AIR-Bench is automated and Heterogeneous with three key features .
Outcome: The proposed benchmarks are based on predefined domains and human-labeled data.
Boosting Long-Context Information Seeking via Query-Guided Activation Refilling (2025.acl-long)

Copied to clipboard

Challenge: Existing methods for processing long contexts are ineffective due to their inherent context window limitations and the computational burden of extensive key-value activations.
Approach: They propose a method for processing long context information-seeking tasks via query-guided ACtivation REfilling (ACRE) a bi-layer KV Cache is constructed where the layer-1 cache compactly captures global information and the layer-2 cache provides detailed, localized information.
Outcome: The proposed method achieves significant improvements in both performance and efficiency on a variety of long-context information-seeking datasets.
RAPID: Efficient Retrieval-Augmented Long Text Generation with Writing Planning and Information Discovery (2025.findings-acl)

Copied to clipboard

Challenge: Existing methods for knowledge-intensive long texts struggle with issues like hallucinations, topic incoherence, and significant latency.
Approach: They propose a retrieval-augmented long text generation framework with writing P**lanning and I**nformation to address these challenges.
Outcome: The proposed framework outperforms state-of-the-art methods on a freshWiki-2024 dataset.
Benchmarking and Enabling Efficient Chinese Medical Retrieval via Asymmetric Encoders (2026.acl-long)

Copied to clipboard

Challenge: Effective medical text retrieval requires high accuracy and low latency.
Approach: They propose a benchmark for medical text retrieval in Chinese using a symmetric architecture . CARE is a lightweight encoder with an LLM-based encoder for offline document encoding .
Outcome: The proposed benchmark surpasses state-of-the-art symmetric models on CMedTEB . it matches high retrieval quality without increasing latency, and it performs well on a single GPU .
Reinforced IR: A Self-Boosting Framework For Domain-Adapted Information Retrieval (2025.acl-long)

Copied to clipboard

Challenge: Existing retrieval methods struggle with highly specialized situations that require extensive domain expertise.
Approach: They propose a method that integrates additional information from an LLM-based generator to enhance query performance and train the retriever to better discriminate the relevant documents identified by the generator.
Outcome: The proposed method outperforms existing domain adaptation methods by a large margin and leads to substantial improvements in retrieval quality across a wide range of application scenarios.
Benchmarking and Improving Compositional Generalization of Multi-aspect Controllable Text Generation (2024.acl-long)

Copied to clipboard

Challenge: Existing MCTG methods face a noticeable performance drop in compositional testing.
Approach: They propose a benchmark to evaluate compositional generalization of MCTG methods by combining multi-aspect labeled datasets and a crafted three-dimensional evaluation protocol.
Outcome: The proposed framework improves compositional generalization performance by 3.64% and 94.4% in compositional testing.
Efficient Data Labeling by Hierarchical Crowdsourcing with Large Language Models (2025.coling-main)

Copied to clipboard

Challenge: Large language models (LLMs) have been gaining attention for their impressive performance in in-context dialogues.
Approach: They propose a hierarchical framework that leverages multiple LLMs for efficient data labeling under budget constraints.
Outcome: The proposed framework outperforms human labelers and GPT-4 in terms of accuracy and efficiency.
Exploring the Choice Behavior of Large Language Models (2025.findings-acl)

Copied to clipboard

Challenge: Large Language Models (LLMs) are increasingly being adopted across various domains where they help to make choices.
Approach: They construct a virtual QA platform that includes three different experimental conditions, with four models from GPT and Llama series participating in repeated experiments.
Outcome: The proposed model includes three experimental conditions and four models from GPT and Llama series.
EquivPruner: Boosting Efficiency and Quality in LLM-Based Search via Action Pruning (2026.acl-long)

Copied to clipboard

Challenge: Existing semantic similarity methods struggle to accurately identify semantically equivalent steps in domain-specific contexts like mathematical reasoning.
Approach: They propose a simple yet effective approach that identifies and prunes semantically equivalent actions during LLM reasoning search.
Outcome: The proposed approach reduces token consumption while improving reasoning efficiency and accuracy on Qwen2.5-Math-7B-Instruct on GSM8K.
Any Information Is Just Worth One Single Screenshot: Unifying Search With Visualized Information Retrieval (2025.acl-long)

Copied to clipboard

Challenge: Existing multimodal retrieval models are lacking in visual representations of multimodal data.
Approach: They propose a visualized information retrieval paradigm where multimodal information is represented by a unified visual format called Screenshots for various retrieval applications.
Outcome: The proposed model is based on a large dataset of screenshots from diverse sources . it is compared with existing models and lays a solid foundation for the new model .

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