Papers by Zhe Sun

28 papers
We-Math: Does Your Large Multimodal Model Achieve Human-like Mathematical Reasoning? (2025.acl-long)

Copied to clipboard

Challenge: Existing benchmarks focus more on end-to-end performance, but neglect the underlying principles of knowledge acquisition and generalization.
Approach: They propose a benchmark specifically designed to explore the problem-solving principles by decomposing 6.5K visual math problems into 10.9K step-level questions for evaluation.
Outcome: The proposed benchmark covers 6.5K visual math problems and 10.9K step-level questions spanning 5 layers of knowledge granularity and 67 hierarchical knowledge concepts.
Unleashing the Unseen: Harnessing Benign Datasets for Jailbreaking Large Language Models (2026.findings-eacl)

Copied to clipboard

Challenge: Despite significant efforts in safety alignment, large language models (LLMs) such as GPT-4 and LLaMA 3 remain vulnerable to jailbreak attacks that can induce harmful behaviors.
Approach: They propose a feature extraction method to extract sample-agnostic features from benign datasets in the form of adversarial suffixes and propose 'suffix maybe features' they show that adversarials generated from jailbreak attacks may contain meaningful features, i.e. appending the same suffix to different prompts results in responses exhibiting specific characteristics.
Outcome: The proposed method extracts sample-agnostic features from benign datasets and shows that they may contain meaningful features.
Improve Vision Language Model Chain-of-thought Reasoning (2025.acl-long)

Copied to clipboard

Challenge: Current training recipes often rely on datasets dominated by short annotations with limited rationales, hindering the models' ability to generalize to tasks requiring comprehensive reasoning.
Approach: They propose a two-stage post-training strategy that augments short answers with CoT reasoning generated by GPT-4o, enhancing the VLM's CoT capabilities through fine-tuning.
Outcome: The proposed strategy enhances the model's CoT capabilities through fine-tuning and reinforcement learning.
SLARD: A Chinese Superior Legal Article Retrieval Dataset (2025.coling-main)

Copied to clipboard

Challenge: Existing retrieval methods struggle to achieve ideal results, a study finds . existing large language models lack prior knowledge of the content of superior legal articles .
Approach: They propose to use a Chinese superior legal article retrieval dataset to find relevant articles with higher legal effectiveness.
Outcome: The proposed dataset shows that existing retrieval methods struggle to achieve ideal results.
Cluster-Former: Clustering-based Sparse Transformer for Question Answering (2021.findings-acl)

Copied to clipboard

Challenge: Existing models for encoding long sequences in deep learning suffer from high latency and memory demands.
Approach: They propose a clustering-based sparse Transformer framework to perform attention across chunked sequences.
Outcome: The proposed framework achieves state-of-the-art on several major QA benchmarks.
Detect Camouflaged Spam Content via StoneSkipping: Graph and Text Joint Embedding for Chinese Character Variation Representation (D19-1)

Copied to clipboard

Challenge: Currently, Chinese characters share glyph and phonetic variations to escape detection algorithms due to their complexity and complexity.
Approach: They propose a Chinese variation-enhanced Graph Embedding algorithm that can learn Chinese character embeddings and latent variation families.
Outcome: The proposed model outperforms state-of-the-art models on Chinese spam detection datasets and review datasets.
Contrastive Distillation on Intermediate Representations for Language Model Compression (2020.emnlp-main)

Copied to clipboard

Challenge: Existing methods to compress language models use a simple L_2 loss to distill knowledge in the intermediate representations of a large BERT model to a smaller one.
Approach: They propose a method that uses knowledge distillation to distill knowledge through intermediate layers of the teacher via a contrastive objective.
Outcome: The proposed method outperforms state-of-the-art methods on the GLUE benchmark.
TencentPretrain: A Scalable and Flexible Toolkit for Pre-training Models of Different Modalities (2023.acl-demo)

Copied to clipboard

Challenge: Several pre-training models of different modalities are showing a rising trend of homogeneity in their model structures.
Approach: They propose a toolkit that supports pre-training models of different modalities.
Outcome: The proposed toolkit can match the performance of the original implementations on text, vision, and audio benchmarks.
Privacy Risks of Intermediate Representations: Attribute Inference in Distributed LLM Inference (2026.findings-acl)

Copied to clipboard

Challenge: Distributed LLMs avoid raw inputs by transmitting intermediate hidden states, a practice widely assumed to preserve privacy.
Approach: They propose a distributed inference framework that transmits intermediate hidden states to avoid sending raw inputs by exposing sensitive user attributes.
Outcome: The proposed approach achieves Top-1 accuracy of 0.997 on CMS, 0.980 on Skytrax, and 0.986 on ECHR.
CLUE: A Chinese Language Understanding Evaluation Benchmark (2020.coling-main)

Copied to clipboard

Challenge: Existing language evaluation benchmarks for English are limited to English . lack of such benchmarks makes it difficult to replicate success in other languages .
Approach: They introduce a large-scale Chinese language understanding evaluation benchmark . the benchmark uses a set of current state-of-the-art pre-trained Chinese models .
Outcome: The first large-scale Chinese Language Understanding Evaluation (CLUE) benchmark is released . the benchmark evaluates models across a wide range of tasks on original Chinese text . existing language evaluation benchmarks are mostly limited to English .
UICOMPASS: UI Map Guided Mobile Task Automation via Adaptive Action Generation (2025.emnlp-main)

Copied to clipboard

Challenge: Mobile task automation is an emerging technology that leverages AI to automatically execute routine tasks by users’ commands on mobile devices like Android.
Approach: They propose a UI Map-guided LLM-based approach to automate mobile tasks using static analysis and LLMs.
Outcome: The proposed approach achieves a 15.87% higher task execution success rate than SOTA approaches even when only APK is available.
Camouflaged Chinese Spam Content Detection with Semi-supervised Generative Active Learning (2020.acl-main)

Copied to clipboard

Challenge: Existing active learning models for text spam detection tasks are based on pool-based active learning, but the annotating process is laborious and time consuming for humans.
Approach: They propose a semi-supervised active learning model to address spam imbalances . they propose masked attention learning approach and character variation graph-enhanced augmentation procedure .
Outcome: The proposed model can improve the performance of existing models for Chinese spam detection task.
CuMA: Aligning LLMs with Sparse Cultural Values via Demographic-Aware Mixture of Adapters (2026.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) have a global audience, so alignment must extend to cultural resonance.
Approach: They propose a framework that frames alignment as a conditional capacity separation problem.
Outcome: The proposed framework outperforms both dense baselines and semantic-only MoEs on three large language models.
On the Helpfulness of Document Context to Sentence Simplification (2020.coling-main)

Copied to clipboard

Challenge: Text simplification is a hot issue in the field of natural language generation (NLG).
Approach: They propose to use Wikipedia context to improve sentence simplification by using neural networks to learn the effects of preceding and following sentences on current sentences.
Outcome: The proposed model outperforms the best performing model on the baseline dataset by 2.46 (7.22%).
Do Influence Functions Work on Large Language Models? (2025.findings-emnlp)

Copied to clipboard

Challenge: Influence functions are important for quantifying the impact of individual training data points on a model’s predictions.
Approach: They conduct a systematic study to address a key question: do influence functions work on large language models?
Outcome: The influence functions perform poorly across multiple tasks and are therefore unsuitable for large language models.
Cross-Thought for Sentence Encoder Pre-training (2020.emnlp-main)

Copied to clipboard

Challenge: Existing models to pretrain sentence encoders with large unlabeled corpus are lacking in linguistic information retrieval.
Approach: They propose a novel approach to pre-training sequence encoder using transformers . they propose to train a Transformer-based sequence encoded over a large set of short sequences based on a set of masked words .
Outcome: The proposed approach outperforms state-of-the-art encoders on hotpotQA by improving intermediate information retrieval performance.
Guided by the Plan: Enhancing Faithful Autoregressive Text-to-Audio Generation with Guided Decoding (2026.eacl-long)

Copied to clipboard

Challenge: Autoregressive (AR) models excel at generating temporally coherent audio by producing tokens sequentially, yet they often falter in faithfully following complex textual prompts.
Approach: They propose a lightweight auxiliary model trained with a GAE-inspired objective to predict final instruction-following quality from partial generations.
Outcome: The proposed model achieves 10 points improvement in CLAP score over baseline AR models while maintaining computational parity with best-of-N decoding.
Multi-stage Distillation Framework for Cross-Lingual Semantic Similarity Matching (2022.findings-naacl)

Copied to clipboard

Challenge: Existing studies have shown that cross-lingual knowledge distillation can improve the performance of pre-trained models for cross-linguistic similarity matching tasks.
Approach: They propose a multi-stage distillation framework for constructing a small-size but high-performance cross-lingual model using contrastive learning, bottleneck, and parameter recurrent strategies.
Outcome: The proposed model can compress the size of XLM-R and MiniLM by more than 50% while the performance is only reduced by about 1%.
M3AV: A Multimodal, Multigenre, and Multipurpose Audio-Visual Academic Lecture Dataset (2024.acl-long)

Copied to clipboard

Challenge: Publishing open-source academic video recordings is an emerging approach to sharing knowledge online.
Approach: They propose a multimodal, multigenre, and multipurpose audio-visual academic lecture dataset with human annotations for multimodal content recognition and understanding tasks.
Outcome: The proposed dataset can be used for multiple audio-visual recognition and understanding tasks.
Defending Large Language Models Against Jailbreak Attacks via Layer-specific Editing (2024.findings-emnlp)

Copied to clipboard

Challenge: Existing defense methods focus on detecting harmful prompts or reducing the likelihood of harmful responses.
Approach: They propose a layer-specific editing method to align LLMs to harmful prompts by supervised fine-tuning and reinforcement learning.
Outcome: The proposed method improves the performance of large language models against jailbreak attacks while maintaining performance on benign prompts.
JTAV: Jointly Learning Social Media Content Representation by Fusing Textual, Acoustic, and Visual Features (C18-1)

Copied to clipboard

Challenge: Existing studies on learning social media content focus on single modal or bi-modal learning, but this approach is non-trivial and challenging because content is multi-modal and involves several types of data, including text, audio, and image.
Approach: They propose to combine textual, acoustic, and visual information to learn social media content by fusing them jointly.
Outcome: The proposed model outperforms the state-of-the-art approaches on real-world datasets by a large margin.
Hierarchical Graph Network for Multi-hop Question Answering (2020.emnlp-main)

Copied to clipboard

Challenge: Existing multi-hop question answering models focus on multi-level reasoning across multiple documents or paragraphs.
Approach: They propose a hierarchical graph network that aggregates clues from scattered texts . they use a set of contextual encoders to initialize nodes on different levels of granularity .
Outcome: The proposed model outperforms existing multi-hop QA approaches on the HotpotQA benchmark.
Domain Adaptive Text Style Transfer (D19-1)

Copied to clipboard

Challenge: Text style transfer without parallel data is a promising method for learning, but in the scenario where less data is available, it may yield poor performance.
Approach: They propose to leverage available data to learn domain-adaptive text style transfer models . they evaluate two style transfer tasks where only limited non-parallel data is available .
Outcome: The proposed models learn from the source domain to: (i) distinguish stylized information and generic content information; (ii) maximally preserve content information and (iv) adaptively transfer the styles in a domain-aware manner.
Patient Knowledge Distillation for BERT Model Compression (D19-1)

Copied to clipboard

Challenge: Pre-trained language models such as BERT have proven to be highly effective for natural language processing tasks, but the high demand for computing resources hinders their application in practice.
Approach: They propose to compress an original large model (teacher) into an equally-effective lightweight shallow network (student) Empirically, this translates into improved results on multiple NLP tasks with a significant gain in training efficiency, without sacrificing model accuracy.
Outcome: The proposed model reduces the computational cost of training models using the teacher model into a lightweight shallow network.
Beyond Superficial Tests: Adversarial Refinement for Reliable Property-Based Testing (2026.findings-acl)

Copied to clipboard

Challenge: Large Language Models (LLMs) have demonstrated remarkable proficiency in code generation, yet their application to Property-Based Testing (PBT) remains fraught with a superficiality gap.
Approach: They propose an agentic framework that hardens software properties through Adversarial Refinement.
Outcome: a new framework hardens software properties through Adversarial Refinement that detects and fixes bugs in top-tier libraries.
KnowLA: Enhancing Parameter-efficient Finetuning with Knowledgeable Adaptation (2024.naacl-long)

Copied to clipboard

Challenge: Existing methods for parameter-efficient finetuning (PEFT) are limited and only finetune a small number of parameters using limited instruction data.
Approach: They propose a method that inserts an adaptation layer into an LLM to integrate embeddings of entities appearing in the input text.
Outcome: The proposed method can activate parameterized knowledge in an LLM without changing its parameters or input prompts.
Zero-Shot Defense Against Toxic Images via Inherent Multimodal Alignment in LVLMs (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing safeguards relying on pre-filtering or fine-tuning are costly and diminish overall utility.
Approach: They propose a lightweight method that leverages LVLMs’ inherent multimodal alignment for zero-shot toxic image detection.
Outcome: The proposed method achieves a 66.9% defense success rate with only 3.2% false positive rate and 7.2% overhead.
Confusion is the Final Barrier: Rethinking Jailbreak Evaluation and Investigating the Real Misuse Threat of LLMs (2025.findings-emnlp)

Copied to clipboard

Challenge: Large Language Models have been developed to deal with real-world crimes, but it remains unclear whether they internalize authentic knowledge or are forced to simulate toxic language patterns.
Approach: They construct knowledge-intensive Q&A to investigate misuse threats of Large Language Models in terms of dangerous knowledge possession, harmful task planning utility, and harmfulness judgment robustness.
Outcome: The findings raise concerns that jailbreak success is often attributable to a hallucination loop between jailbroken LLM and judger LLM .

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