Papers by Jiajun Sun

17 papers
Muse: Towards Reproducible Long-Form Song Generation with Fine-Grained Style Control (2026.findings-acl)

Copied to clipboard

Challenge: Recent commercial systems such as Suno demonstrate strong capabilities in long-form song generation, but academic research remains non-reproducible due to the lack of publicly available training data.
Approach: They propose a system for long-form song generation with fine-grained style conditioning that includes a licensed synthetic dataset and a song generation model, Muse.
Outcome: The proposed system achieves competitive performance on phoneme error rate, text–music style similarity, and audio aesthetic quality while enabling controllable segment-level generation across different musical structures.
AgentGym2: Benchmarking Large Language Model Agents in De-Idealized Real-World Environments (2026.acl-long)

Copied to clipboard

Challenge: Existing benchmarks evaluate agents in simplified, idealized settings, relying on pre-packaged tool interfaces, overlooking critical steps, and assume inputs are clean and fully specified.
Approach: They propose a framework that evaluates language agents in simplified, idealized settings . they show that even SOTA systems like Gemini and GPT-5 struggle on AgentGym2 .
Outcome: Experiments on 15 proprietary and open-source models show that even SOTA systems like Gemini and GPT-5 struggle on AgentGym2 .
The Mirage of Model Editing: Revisiting Evaluation in the Wild (2025.acl-long)

Copied to clipboard

Challenge: despite near-perfect results, effectiveness of model editing in real-world applications remains unclear.
Approach: They propose QAEdit and WILD to better reflect real-world use of model editing . they propose a benchmark aligned with widely used question answering datasets and a task-agnostic evaluation framework .
Outcome: The proposed QAEdit benchmark and WILD evaluation framework show that current models perform worse than previously reported.
An Efficient and Precise Training Data Construction Framework for Process-supervised Reward Model in Mathematical Reasoning (2025.acl-long)

Copied to clipboard

Challenge: Existing methods for constructing process supervision training data are costly or suffer from poor quality.
Approach: They propose a framework called EpicPRM which annotates each intermediate reasoning step based on its quantified contribution and uses an adaptive binary search algorithm to enhance annotation precision and efficiency.
Outcome: The proposed framework improves annotation precision and efficiency and can be used to train a high-quality training dataset with 50k annotated intermediate steps.
Alleviating the Inequality of Attention Heads for Neural Machine Translation (2022.coling-1)

Copied to clipboard

Challenge: Recent studies show that the attention heads in Transformer are not equal.
Approach: They propose a masking method to mask attention heads in Transformer . they empirically validate the inequality and propose 'head mask' method to avoid bottleneck .
Outcome: The proposed masking method improves translation performance on multiple languages . it can be used to remove a small subset of heads without affecting performance .
Towards Database-Free Text-to-SQL Evaluation: A Graph-Based Metric for Functional Correctness (2025.coling-main)

Copied to clipboard

Challenge: Existing metrics for evaluating functional correctness of SQL queries are prone to false positives due to inadequately prepared test databases.
Approach: They propose a graph-based metric that uses a relational operator tree to extract rich semantic information from the logical execution plan of SQL queries and embed it into a diagram.
Outcome: The proposed method eliminates the need for extensive test database preparation and performs graph matching on unseen SQL queries.
Rethinking Document-level Neural Machine Translation (2022.findings-acl)

Copied to clipboard

Challenge: Neural machine translation models are weak enough for document-level translation . current models only translate sentences individually, resulting in poor document coherence .
Approach: They propose to use the original Transformer model to test document-level neural machine translation . they find that the original transformer models can achieve strong results for document translation if trained properly .
Outcome: The proposed model outperforms sentence-level models on nine datasets and two sentence- level datasets across six languages.
Distill and Replay for Continual Language Learning (2020.coling-main)

Copied to clipboard

Challenge: Existing models fail to isolate acquired knowledge and forget previously learned tasks when learning in a stream where data distribution may shift.
Approach: They propose a framework that distills knowledge and replays experience from previous tasks when fitting on a new task.
Outcome: The proposed framework outperforms state-of-the-art models in continuously learning tasks of the same type but from different domains, as well as tasks of different types.
MdEval: Massively Multilingual Code Debugging (2026.findings-acl)

Copied to clipboard

Challenge: Existing benchmarks primarily focus on Python and are limited in terms of language diversity.
Approach: They propose a multilingual debugging benchmark that includes 3.9K test samples of 20 programming languages and introduces the debug instruction corpora MdEval-Instruct by injecting bugs into the correct multilingual queries and solutions.
Outcome: The proposed benchmark includes 3.9K test samples of 20 programming languages and covers the automated program repair task, bug localization task, and bug identification task.
CapArena: Benchmarking and Analyzing Detailed Image Captioning in the LLM Era (2025.findings-acl)

Copied to clipboard

Challenge: Image captioning has been a challenge for vision-language researchers for decades . current VLMs focus on tasks like visual question answering (YA) but image captioning is not as advanced as expected.
Approach: They evaluate VLMs' performance on image captioning using human annotations . they find that some metrics show high caption-level agreement with humans .
Outcome: The proposed model outperforms open-source models on image captioning . it achieves 93.4% correlation with human rankings at $4 per test .
AutoSearch: Adaptive Search Depth for Efficient Agentic RAG via Reinforcement Learning (2026.findings-acl)

Copied to clipboard

Challenge: Prior work limits search depth to reduce cost, but this often leads to underexploration of complex questions.
Approach: They propose a reinforcement learning framework that evaluates each search step via self-generated intermediate answers.
Outcome: Extensive experiments on multiple benchmarks show that AutoSearch achieves a superior accuracy-efficiency trade-off, alleviating over-searching while preserving search quality.
SDGO: Self-Discrimination-Guided Optimization for Consistent Safety in Large Language Models (2025.emnlp-main)

Copied to clipboard

Challenge: Large Language Models (LLMs) excel at various tasks but are vulnerable to jailbreak attacks that induce harmful content generation.
Approach: They propose a reinforcement learning framework that leverages the model’s own discrimination capabilities as a reward signal to enhance generation safety through iterative self-improvement.
Outcome: The proposed framework improves model safety by iterative self-improvement without additional annotated data or external models during training phase.
From Scores to Preferences: Redefining Evaluation Paradigm for Speech Quality Reward Modeling (2026.findings-acl)

Copied to clipboard

Challenge: Experimental results show that the MOS-aware GRM significantly improves fine-grained speech quality discrimination.
Approach: They propose a MOS-aware reward model that incorporates MOS gap into reward function during reinforcement learning.
Outcome: The proposed model significantly improves fine-grained speech quality discrimination.
Mitigating Object Hallucinations in MLLMs via Multi-Frequency Perturbations (2025.findings-emnlp)

Copied to clipboard

Challenge: Multimodal large language models have demonstrated remarkable performance in visual-language tasks, but their authenticity is often compromised by object hallucinations.
Approach: They propose a multi-frequency perturbation method that leverages both low-frequency and high-frequency features of images to perturb visual feature representations and explicitly suppress redundant frequency-domain features during inference.
Outcome: The proposed method significantly mitigates object hallucinations across various model architectures.
The Fall of ROME: Understanding the Collapse of LLMs in Model Editing (2024.findings-emnlp)

Copied to clipboard

Challenge: Recent studies have found that model editing methods can cause large language models to collapse with just a single edit.
Approach: They propose a method that uses prefixed keys and adds prefixes during testing to prevent model collapse.
Outcome: The proposed method prevents model collapse while maintaining effectiveness, the authors show . Rank-One Model Editing (ROME) has been found to cause model collapse with just a single edit .
Beyond Scaling: Measuring and Predicting the Upper Bound of Knowledge Retention in Language Model Pre-Training (2026.acl-long)

Copied to clipboard

Challenge: Existing methods to predict performance of large language models are lacking . authors propose a size-dependent mutual information predictor for closed-book question answering accuracy .
Approach: They propose a size-dependent mutual information predictor that integrates knowledge frequency, knowledge specificity, and model size to forecast closed-book question answering accuracy.
Outcome: The proposed method outperforms baseline models and achieves R2 > 0.7 in predicting QA accuracy without additional training.
Look Again, Think Slowly: Enhancing Visual Reflection in Vision-Language Models (2025.emnlp-main)

Copied to clipboard

Challenge: Recent advances in text-only "slow thinking" reasoning have prompted efforts to transfer this capability to vision-language models (VLMs).
Approach: They propose a VRM Reflection-V which enhances visual reflection based on reasoning data for cold-start and reward design for reinforcement learning.
Outcome: The proposed model improves visual reflection for cold-start and reward design for reinforcement learning (RL) it maintains a stronger and more consistent reliance on visual information during visual reasoning, indicating effective enhancement in visual reflection capabilities.

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