Papers by Haifeng Sun

39 papers
ERNIE-Doc: A Retrospective Long-Document Modeling Transformer (2021.acl-long)

Copied to clipboard

Challenge: Existing models for document-level language pretraining are not suitable for long documents due to their quadratically increasing memory and time consumption.
Approach: They propose a document-level language pretraining model based on Recurrence Transformers.
Outcome: The proposed model outperforms existing models on language understanding tasks.
Erasing Without Remembering: Implicit Knowledge Forgetting in Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: a new method for unlearning large language models is proposed to improve the performance of large language model models.
Approach: They propose a probability perturbation-based unlearning paradigm that allows models to forget implicit knowledge in large language models with a focus on generalisation.
Outcome: The proposed model improves unlearning vanilla target data while forgetting implicit knowledge.
CORD: Bridging the Audio–Text Reasoning Gap via Weighted On-policy Cross-modal Distillation (2026.findings-acl)

Copied to clipboard

Challenge: Large Audio Language Models (LALMs) exhibit a degradation in knowledge and reasoning capabilities . empirical results show that CORD significantly bridges the audio–text performance gap .
Approach: They propose a framework that performs online cross-modal self-distillation to bridge the acoustic-semantic gap between LALMs and text-based models.
Outcome: The proposed framework bridges the acoustic-semantic gap between LALMs and text-based models . it employs on-policy reverse KL divergence with importance-aware weighting .
Rethinking Smoothness for Fast and Adaptable Entity Alignment Decoding (2025.findings-naacl)

Copied to clipboard

Challenge: Existing methods for integrating knowledge graphs rely on entity and relation embeddings . Fig. 1 shows how to decode knowledge graph in under 6 seconds .
Approach: They propose a framework that only utilizes entity embeddings to decode knowledge graphs.
Outcome: The proposed framework reconstructs KG representation by maximizing smoothness of entity embeddings.
Uncertainty-Aware Routing for Principled Alignment with MoE Dynamics (2026.acl-long)

Copied to clipboard

Challenge: Mixture-of-Experts (MoE) is a cornerstone for scaling LLMs, yet its training dynamics remain poorly understood, often leading to sub-optimal specialization.
Approach: They propose to use Helmholtz Free Energy and Router Entropy to study the MoE lifecycle and identify a universal Three-Stage Phase Transition .
Outcome: The proposed model reduces perplexity and improves expert distinctiveness, offering a principled path toward thermodynamically aligned computation.
Distributional Clarity: The Hidden Driver of RL-Friendliness in Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: RL-friendly models exhibit intra-class compactness and inter-class separation in probability assignments . under identical training, Qwen models achieve substantial gains, while others like Llama yield limited improvements.
Approach: They propose a method to quantify distributional clarity in probability space . they show distributional clearness is a trainable property underlying RL-Friendliness .
Outcome: The proposed model families achieve substantial gains under identical training, while others like Llama yield limited improvements.
VALU: A Benchmark for Video Anomaly Temporal Localization and Understanding at Multiple Semantic Levels (2026.acl-long)

Copied to clipboard

Challenge: Recent advances in Video Large Language Models (Video-LLMs) enhance the ability of VAU models to describe and interpret anomalies.
Approach: They propose a benchmark that explicitly defines anomalies across five semantic levels and provides detailed temporal boundaries and detailed textual descriptions for each.
Outcome: The proposed benchmark defines anomalies across five semantic levels and provides detailed descriptions for each.
MDR: Model-Specific Demonstration Retrieval at Inference Time for In-Context Learning (2024.naacl-long)

Copied to clipboard

Challenge: Existing methods for retrieval-based in-context learning ignore model biases and fail to retrieve the most appropriate demonstrations for different LLMs.
Approach: They propose a model-specific demonstration retrieval method that considers the biases of different LLMs at inference time.
Outcome: The proposed method improves performance on seen and unseen tasks with multi-scale inference LLMs by up to 41.2%.
ClusterAttn: KV Cache Compression under Intrinsic Attention Clustering (2025.acl-long)

Copied to clipboard

Challenge: Existing methods for sparse attention apply the same pattern across different attention heads and inputs, but fail to capture the intrinsic attention clustering in large language models.
Approach: They propose a training-free sparse attention method that provides an efficient prompt cache compression scheme under intrinsic attention clustering for efficient LLM inference.
Outcome: The proposed method reduces memory usage by 10%–65% and increases throughput by 2.6–4.8 times with no accuracy loss.
Modeling Aspect Correlation for Aspect-based Sentiment Analysis via Recurrent Inverse Learning Guidance (2022.coling-1)

Copied to clipboard

Challenge: Existing methods to learn complex sentence with multiple aspects do not consider correlation between aspects to distinguish overlapped feature.
Approach: They propose a method that uses aspect correlation to improve aspect correlation modeling . they use Recurrent Mechanism to improve the joint representation of aspects .
Outcome: The proposed method is state-of-the-art in multiaspect scenarios.
Unveiling Internal Reasoning Modes in LLMs: A Deep Dive into Latent Reasoning vs. Factual Shortcuts with Attribute Rate Ratio (2025.emnlp-main)

Copied to clipboard

Challenge: Existing research in multi-hop questions has identified two reasoning modes, but has not investigated how these modes differ during inference.
Approach: They propose a classification metric that compares latent reasoning and factual shortcuts in multi-hop questions.
Outcome: The proposed metric achieves 90% accuracy on the proposed datasets and demonstrates effectiveness in RAG conflict scenarios.
SSS: Editing Factual Knowledge in Language Models towards Semantic Sparse Space (2024.findings-acl)

Copied to clipboard

Challenge: Existing methods to modify LMs suffer from sub-optimal locality, where irrelevant neighborhood examples can be adversely influenced.
Approach: They propose to use a model editing method to modify specific examples in LMs to improve locality and reasoning capability by directing the hidden state of edit example towards spaces where semantics are sparse.
Outcome: The proposed method improves locality and reasoning capability on two datasets.
Investigating Capsule Network and Semantic Feature on Hyperplanes for Text Classification (D19-1)

Copied to clipboard

Challenge: Various neural networks are designed for text classification on the basis of word embedding, but polysemy is a fundamental feature of the natural language, which brings challenges to text classification.
Approach: They propose to use capsule networks to construct the vectorized representation of semantics and utilize hyperplanes to decompose each capsule to acquire the specific senses.
Outcome: The proposed model extracts more discriminative semantic features and yields significant performance gain compared to baseline methods.
Clip-Tuning: Towards Derivative-free Prompt Learning with a Mixture of Rewards (2022.findings-emnlp)

Copied to clipboard

Challenge: Existing work does not take full advantage of over-parameterized characteristics of large pre-trained language models.
Approach: They propose a method that uses frozen "thinned" networks to obtain a mixture of rewards and advance the derivative-free prompt learning.
Outcome: The proposed method outperforms previous gradient-free prompt learning methods and achieves parity with gradient-based counterparts on seven language understanding benchmarks under few-shot settings.
A Diverse and Effective Retrieval-Based Debt Collection System with Expert Knowledge (2025.naacl-industry)

Copied to clipboard

Challenge: Existing debt collection systems lack script diversity, contextual relevance and coherence due to their complexity.
Approach: They propose a script library based on real-world debt collection conversations and a retrieval based response system for contextual relevance.
Outcome: The proposed system improves script diversity and responds to debtor-collector conversations better through knowledge distillation.
ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora (2021.emnlp-main)

Copied to clipboard

Challenge: Existing methods for pretraining cross-lingual models are limited in their size due to the limited amount of parallel corpora.
Approach: They propose a method that encourages the model to align multiple languages with monolingual corpora to overcome the constraint of the parallel corpus size.
Outcome: The proposed method outperforms existing cross-lingual models and delivers new state-of-the-art results in various cross-linguistic downstream tasks.
Inner Thinking Transformer: Leveraging Dynamic Depth Scaling to Foster Adaptive Internal Thinking (2025.acl-long)

Copied to clipboard

Challenge: Large language models face inherent performance bottlenecks under parameter constraints . challenging tokens induce abrupt gradient spikes across layers, exposing stress points .
Approach: They propose an inner thinking transformer that reimagines layer computations as implicit thinking steps.
Outcome: Empirical results show that ITT outperforms Transformer/Loop variants in 11 benchmarks.
Curriculum Learning Meets Weakly Supervised Multimodal Correlation Learning (2022.emnlp-main)

Copied to clipboard

Challenge: Existing studies have used the correlation information stored in samples for self-supervised learning, but they feed the training pairs in a random order without consideration of difficulty.
Approach: They propose to inject curriculum learning into weakly supervised multimodal correlation learning by scoring and feeding pairs according to difficulty.
Outcome: The proposed model achieves state-of-the-art on multimodal sentiment analysis without human annotation.
RecStream: Graph-aware Stream Management for Concurrent Recommendation Model Online Serving (2025.coling-industry)

Copied to clipboard

Challenge: Existing systems that use recommendation models perform poorly under highly concurrent scenarios.
Approach: They propose a system that optimizes stream configurations based on model characteristics and concurrency levels.
Outcome: The proposed system outperforms existing methods under high concurrency scenarios.
MoE Adapter for Large Audio Language Models: Sparsity, Disentanglement, and Gradient-Conflict-Free (2026.findings-acl)

Copied to clipboard

Challenge: Existing research on Large Language Models (LLMs) limited to textual input modality . acoustic information is intrinsically heterogeneous, entangling attributes such as speech, music, and environmental context.
Approach: They propose a sparse Mixture-of-Experts architecture to decouple acoustic information by routing audio tokens to specialized experts.
Outcome: The proposed architecture outperforms existing models on audio semantic and paralinguistic tasks while retaining shared experts for global context.
The Threat of PROMPTS in Large Language Models: A System and User Prompt Perspective (2025.findings-acl)

Copied to clipboard

Challenge: Prompts are essential for guiding model output and influencing content generation.
Approach: They propose to attack models with prompt leakage and prompt jailbreak attacks . they summarize the experimental setups of these methods and explore the relationship between prompt threats and prompt injection attacks.
Outcome: The proposed methods summarize the experimental setups and examine the relationship between prompt threats and prompt injection attacks.
Distantly Supervised Contrastive Learning for Low-Resource Scripting Language Summarization (2024.lrec-main)

Copied to clipboard

Challenge: Existing methods for code summarization are limited in resources and require atomic commands and category constraints to enhance code representations.
Approach: They propose a framework that leverages limited atomic commands and category constraints to enhance code representations.
Outcome: The proposed framework outperforms baseline methods in a number of domains and demonstrates superiority over competing frameworks.
Stop When Enough: Adaptive Early-Stopping for Chain-of-Thought Reasoning (2026.acl-long)

Copied to clipboard

Challenge: Chain-of-Thought reasoning has driven recent gains of large language models (LLMs) on reasoning-intensive tasks by externalizing intermediate steps.
Approach: They propose a training-free framework that adaptively determines when to stop reasoning to mitigate overthinking.
Outcome: The proposed framework reduces token usage by 20-55% while maintaining or improving accuracy compared to standard CoT prompting.
Correcting Chinese Spelling Errors with Phonetic Pre-training (2021.findings-acl)

Copied to clipboard

Challenge: Existing methods for Chinese spelling correction only use pre-trained language model or incorporate phonological information as external knowledge.
Approach: They propose a phonetic Chinese spelling correction model that integrates phonetic features into language model by leveraging pre-training and fine-tuning methods.
Outcome: The proposed model outperforms existing methods on SIGHAN datasets and improves on other datasets.
Uncertainty Quantification for In-Context Learning of Large Language Models (2024.naacl-long)

Copied to clipboard

Challenge: Existing studies on in-context learning have focused on quantifying the uncertainty associated with the model's response, but they neglect the complexity of the LLM and the uniqueness of in-constitut learning.
Approach: They propose a method to quantify the uncertainty associated with in-context learning and propose corresponding estimation method to quantify both types of uncertainties.
Outcome: The proposed method offers an unsupervised way to understand the prediction of in-context learning in a plug-and-play fashion.
Evaluating the Expressive Appropriateness of Speech in Rich Contexts (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for evaluating expressive speech focus on word accuracy, naturalness, signal quality, or emotional intensity at the utterance level.
Approach: They propose a framework for Evaluating Expressive Appropriateness in speech that assesses whether a speech sample aligns with the underlying communicative intent implied by its discourse-level narrative context.
Outcome: The proposed framework outperforms existing speech evaluation and analysis systems on a human-annotated test set.
Adversarial and Domain-Aware BERT for Cross-Domain Sentiment Analysis (2020.acl-main)

Copied to clipboard

Challenge: Cross-domain sentiment classification requires large amounts of labeled data.
Approach: They propose to apply a pre-training language model BERT on unsupervised domain adaptation . they propose to distill domain-specific features in a self-supervised way .
Outcome: The proposed model outperforms state-of-the-art methods on Amazon dataset . it can be applied to the unsupervised domain adaptation task without domain awareness .
Uncertainty Propagation on LLM Agent (2025.acl-long)

Copied to clipboard

Challenge: Existing methods for estimating uncertainty in large language models (LLMs) focus on final-step outputs, which fail to account for cumulative uncertainty over multi-step decision-making process and dynamic interactions between agents and their environments.
Approach: They propose a framework that propagates uncertainty through each step of an LLM-based agent’s reasoning process.
Outcome: Extensive experiments on benchmark datasets show that the proposed framework outperforms state-of-the-art methods by 20%.
DuSQL: A Large-Scale and Pragmatic Chinese Text-to-SQL Dataset (2020.emnlp-main)

Copied to clipboard

Challenge: Existing text-to-SQL parsing methods mainly focus on English, but there is no labeled data available for the language . a larges-scale and pragmatic Chinese dataset is used for cross-domain text- to-Sql task .
Approach: They propose a larges-scale Chinese dataset for a cross-domain text-to-SQL task . they analyze questions from several representative applications and use an effective data construction framework .
Outcome: The proposed dataset contains 200 databases, 813 tables, and 23,797 question/SQL pairs.
Sparse Growing Transformer: Training-Time Sparse Depth Allocation via Progressive Attention Looping (2026.findings-acl)

Copied to clipboard

Challenge: Existing approaches to increasing effective depth of LLMs rely on parameter reuse, extending computation through recursive execution.
Approach: They propose a training-time sparse depth allocation framework that progressively increases depth for a small subset of parameters as training evolves.
Outcome: The proposed model outperforms existing approaches to increasing the effective depth of language models while reducing training FLOPs overhead from approximately 16–20% to only 1–3% relative to a standard Transformer backbone.
BeamLoRA: Beam-Constraint Low-Rank Adaptation (2025.acl-long)

Copied to clipboard

Challenge: Low-Rank Adaptation (LoRA) is one of the most efficient parameter-efficient fine-tuning methods.
Approach: They propose to conceptualize each LoRA module as a beam where each rank corresponds to a potential sub-solution.
Outcome: The proposed method improves performance on three base models and 12 datasets.
Dual Activation-Weight Sparsity: A Training-Free Framework for Efficient Large Language Model Compression (2026.acl-long)

Copied to clipboard

Challenge: Large language models (LLMs) excel at natural language tasks but face deployment bottlenecks due to computational demands.
Approach: They propose a training-free framework that exploits activation and weight sparsity . they use a three-tier routing strategy that uses magnitude-based pruning .
Outcome: Experiments on Llama and Mistral models show that DAWS outperforms activation-weight sparsity pruning methods.
ERNIE-Gram: Pre-Training with Explicitly N-Gram Masked Language Modeling for Natural Language Understanding (2021.naacl-main)

Copied to clipboard

Challenge: Existing methods to model coarse-grained linguistic information do not integrate coarse-gram information into pre-training.
Approach: They propose an explicitly n-gram masking method to enhance integration of coarse-grained linguistic information into pre-training.
Outcome: The proposed method outperforms existing models on English and Chinese text corpora and fine-tunes on 19 downstream tasks.
Evaluating and Mitigating Object Hallucination in Large Vision-Language Models: Can They Still See Removed Objects? (2025.naacl-long)

Copied to clipboard

Challenge: LVLMs often mistakenly determine objects as present in images where they do not exist . authors propose a new benchmark to evaluate object hallucinations by removing objects from images and asking the model whether it can still see the removed objects.
Approach: They propose a benchmark to evaluate object hallucinations by removing objects from images . they propose oDPO, a direct preference optimization objective based on visual objects .
Outcome: The proposed benchmark reduces the likelihood of object hallucinations by removing objects from images and asking the model whether it can still see the removed objects.
Curiosity-Driven Reinforcement Learning from Human Feedback (2025.acl-long)

Copied to clipboard

Challenge: Reinforcement learning from human feedback (RLHF) has proven effective in aligning large language models with human preferences, but often at the cost of reduced output diversity.
Approach: They propose a framework that incorporates intrinsic rewards for novel states alongside traditional sparse extrinsic rewards to optimize both output diversity and alignment quality.
Outcome: The proposed framework achieves significant gains in diversity on multiple diversity-oriented metrics while maintaining alignment with human preferences comparable to standard RLHF.
ERNIE-Layout: Layout Knowledge Enhanced Pre-training for Visually-rich Document Understanding (2022.findings-emnlp)

Copied to clipboard

Challenge: Existing methods for visually rich document understanding lack layout-centered knowledge . experimental results show that ERNIE-Layout improves layout awareness .
Approach: They propose a document pre-training solution with layout knowledge enhancement in the whole workflow to learn better representations that combine the features from text, layout, and image.
Outcome: The proposed model outperforms existing models on key downstream tasks.
Example Quality Matters: Multi-Aspects Example Augmentation for Private Library Programming (2026.acl-long)

Copied to clipboard

Challenge: Existing approaches to code generation fail to consider the quality of retrieved examples.
Approach: They propose a retrieval-augmented generation method that combines existing API examples to improve complexity and readability.
Outcome: The proposed method achieves up to 22% accuracy improvement over baseline methods.
Capsule Network with Interactive Attention for Aspect-Level Sentiment Classification (D19-1)

Copied to clipboard

Challenge: Existing methods for aspect-level sentiment classification are limited for dealing with overlapped features.
Approach: They propose to use capsule network to construct vector-based feature representation and cluster features by an EM routing algorithm to model semantic relationship between aspect terms and context.
Outcome: The proposed model achieves state-of-the-art on three datasets.
ModularMoE: Fast LLM Customization with Parameter-Sharing Mixture-of-Experts for Low-Resource Settings (2026.findings-acl)

Copied to clipboard

Challenge: Large Language Models impose significant computational and storage burdens on personal devices . existing customization approaches incur excessive computational costs or lead to suboptimal performance .
Approach: They propose a training framework that converts pre-trained LLMs into parameter-sharing MoE models for lightweight deployment.
Outcome: The proposed training framework outperforms state-of-the-art training frameworks at the same sparsity level while delivering up to 2.71 inference speedup.

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