Papers by Jingyu Liu
Beyond Reasoning Gains: Mitigating General-Capability Forgetting in Large Reasoning Models (2026.findings-acl)
Copied to clipboard
Hoang Phan, Xianjun Yang, Yuanshun Yao, Jingyu Zhang, Shengjie Bi, Xiaocheng Tang, Madian Khabsa, Lijuan Liu, Deren Lei
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning . however, the recipe introduces a significant risk of capability regression, where models forget foundational skills after prolonged training without employing regularization strategies. |
| Approach: | They propose a replay strategy with dynamic objective reweighting for general knowledge preservation using short-horizon signals of convergence and instability. |
| Outcome: | The proposed method preserves general capabilities and improves reasoning . it can be applied to existing RLVR pipelines without training additional models or tuning . |
Core: Robust Factual Precision with Informative Sub-Claim Identification (2025.findings-acl)
Copied to clipboard
Zhengping Jiang, Jingyu Zhang, Nathaniel Weir, Seth Ebner, Miriam Wanner, Kate Sanders, Daniel Khashabi, Anqi Liu, Benjamin Van Durme
| Challenge: | Using the Decompose-Then-Verify framework, such as FActScore, can be manipulated by adding obvious or repetitive subclaims to artificially inflate scores. |
| Approach: | They propose a decomposition-based tool called Core to filter subclaims based on their uniqueness and informativeness. |
| Outcome: | The proposed evaluation framework supports easy and modular use of Core and various decomposition strategies. |
RAFFLES: Reasoning-based Attribution of Faults for LLM Systems (2026.eacl-long)
Copied to clipboard
Chenyang Zhu, Spencer Hong, Jingyu Wu, Kushal Chawla, Yuhui Tang, Youbing Yin, Nathan Wolfe, Erin Babinsky, Daben Liu
| Challenge: | Existing evaluation frameworks focus on simple metrics and end-to-end outcomes, but they struggle with longer contexts. |
| Approach: | They propose an offline evaluation architecture that incorporates iterative reasoning to evaluate the quality of the candidate faults and rationales of the Judge. |
| Outcome: | The proposed architecture outperforms baseline evaluation frameworks with two datasets to identify step-level faults in multi-agent systems and ReasonEval datasets. |
Harmonizing Diverse Models: A Layer-wise Merging Strategy for Consistent Generation (2025.emnlp-industry)
Copied to clipboard
| Challenge: | RAG systems often generate inconsistent outputs for semantically equivalent inputs . this unpredictability undermines the reliability of RAG and poses challenges for adoption in high-stakes or knowledge-sensitive domains such as finance, healthcare, and scientific research. |
| Approach: | They propose a method that integrates knowledge from specialized models into a single model to improve output consistency. |
| Outcome: | The proposed model significantly improves output consistency, achieving approximately 47.5% improvement in response similarity over baseline. |
Exploring Response Uncertainty in MLLMs: An Empirical Evaluation under Misleading Scenarios (2025.emnlp-main)
Copied to clipboard
Yunkai Dang, Mengxi Gao, Yibo Yan, Xin Zou, Yanggan Gu, Jungang Li, Jingyu Wang, Peijie Jiang, Aiwei Liu, Jia Liu, Xuming Hu
| Challenge: | Existing studies have focused mainly on visual–textual misalignment, leaving largely unexplored the MLLMs’ ability to preserve an original correct answer when confronted with misleading information. |
| Approach: | They propose a two-stage evaluation pipeline to quantify the response uncertainty phenomenon by eliciting each model’s original response on unperturbed inputs and injecting explicit (false-answer hints) and implicit (contextual contradictions) misleading instructions. |
| Outcome: | The proposed model overturns a correct answer in 65% of cases after receiving a single deceptive cue. |
Large Vision-Language Model Alignment and Misalignment: A Survey Through the Lens of Explainability (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large Vision-Language Models have demonstrated remarkable capabilities in processing both visual and textual information. |
| Approach: | They examine the challenge of alignment and misalignment in LVLMs through an explainability lens. |
| Outcome: | The findings highlight the need for standardized evaluation protocols and in-depth explainability studies. |
Strategic Demonstration Selection for Improved Fairness in LLM In-Context Learning (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies highlight the effectiveness of using in-context learning (ICL) to steer large language models in processing tabular data. |
| Approach: | They propose a method that uses clustering and evolutionary strategies to curate a representative sample set from training data. |
| Outcome: | The proposed method significantly improves fairness across various metrics, showing its efficacy in real-world scenarios. |
Optimizing Multi-Hop Document Retrieval Through Intermediate Representations (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to addressing multi-hop queries are computationally expensive . despite their success, large language models often generate factually incorrect answers . |
| Approach: | They propose a layer-by-layer reasoning approach that leverages intermediate representations from the middle layers to retrieve external knowledge. |
| Outcome: | The proposed method outperforms existing RAG methods on open-domain multi-hop question-answering datasets while maintaining inference overhead similar to that of standard RAG. |
RATIONALYST: Pre-training Process-Supervision for Improving Reasoning (2025.acl-long)
Copied to clipboard
Dongwei Jiang, Guoxuan Wang, Yining Lu, Andrew Wang, Jingyu Zhang, Chuyu Liu, Benjamin Van Durme, Daniel Khashabi
| Challenge: | RATIONALYST is a model for process-supervision of reasoning based on pretraining on rationale annotations extracted from unlabeled data. |
| Approach: | They propose a model for process-supervision of reasoning based on pre-training on rationale annotations extracted from unlabeled data. |
| Outcome: | RATIONALYST improves reasoning accuracy by 3.9% on representative reasoning benchmarks. |
Effective Long-Context Scaling of Foundation Models (2024.naacl-long)
Copied to clipboard
Wenhan Xiong, Jingyu Liu, Igor Molybog, Hejia Zhang, Prajjwal Bhargava, Rui Hou, Louis Martin, Rashi Rungta, Karthik Abinav Sankararaman, Barlas Oguz, Madian Khabsa, Han Fang, Yashar Mehdad, Sharan Narang, Kshitiz Malik, Angela Fan, Shruti Bhosale, Sergey Edunov, Mike Lewis, Sinong Wang, Hao Ma
| Challenge: | Large language models (LLMs) are rapidly deployed and continue to evolve through scaling. |
| Approach: | They propose a method to train strong long-context LLMs that are capable of utilizing massive context windows of up to 32,000 tokens. |
| Outcome: | The proposed model can surpass gpt-3.5-turbo-16k's overall performance on long-context benchmarks with a cost-effective instruction tuning procedure that is free of expensive annotations. |
Do not Abstain! Identify and Solve the Uncertainty (2025.acl-long)
Copied to clipboard
| Challenge: | Existing solutions rely on evasive responses when confronting uncertain scenarios. |
| Approach: | They propose a benchmark to assess LLMs' ability to recognize and address uncertainty . they generate context-aware inquiries that highlight the confusing aspect of the original query . |
| Outcome: | Experiments with ConfuseBench show that LLMs struggle to identify root cause of uncertainty and solve it. |
Text-guided 3D Human Generation from 2D Collections (2023.findings-emnlp)
Copied to clipboard
| Challenge: | 3D human modeling is used for engaging interaction in gaming, film, and animation. however, the customization of characters is crucial for creativity and scalability. |
| Approach: | They propose a 3D human generation using fashion descriptions to enhance 3D geometry transformation and fine-grained consistency. |
| Outcome: | The proposed model can generate a 3D human, guided by a fashion description, with high efficiency. |
SlideCoder: Layout-aware RAG-enhanced Hierarchical Slide Generation from Design (2025.emnlp-main)
Copied to clipboard
Wenxin Tang, Jingyu Xiao, Wenxuan Jiang, Xi Xiao, Yuhang Wang, Xuxin Tang, Qing Li, Yuehe Ma, Junliang Liu, Shisong Tang, Michael R. Lyu
| Challenge: | Existing natural language-based LLM generation methods struggle to capture visual and structural nuances of slide designs. |
| Approach: | They propose a layout-aware framework for generating editable slides from reference images . they propose python code that translates NL instructions into Python code to construct each slide . |
| Outcome: | The proposed framework outperforms state-of-the-art models by up to 40.5 points . it also outperformed open-source models with improved reverse-engineered data. |
Stepwise Reasoning Disruption Attack of LLMs (2025.acl-long)
Copied to clipboard
Jingyu Peng, Maolin Wang, Xiangyu Zhao, Kai Zhang, Wanyu Wang, Pengyue Jia, Qidong Liu, Ruocheng Guo, Qi Liu
| Challenge: | Existing attacks on LLM reasoning are constrained by specific settings or lack of imperceptibility, limiting their feasibility and generalizability. |
| Approach: | They propose a stepwise rEasoning error disruption attack that subtly injects errors into prior reasoning steps to mislead the model into producing incorrect subsequent reasoning and final answers. |
| Outcome: | The proposed attack is compatible with zero-shot and few-shot settings, maintains the natural reasoning flow, and ensures covert execution without modification of the instruction. |
JI2S: Joint Influence‐Aware Instruction Data Selection for Efficient Fine‐Tuning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Prior selection strategies score samples using generalpurpose LLMs, leveraging their strong language understanding but introducing inherent biases that misalign with the target model’s behavior and yield unstable downstream performance. |
| Approach: | They propose a framework that jointly models marginal and combinatorial influences within sample groups and evaluate them on Open LLM Benchmarks, MTBench, and GPT4–judged pairwise comparisons. |
| Outcome: | The proposed framework outperforms fulldataset training and strong baselines on Open LLM Benchmarks, MTBench, and GPT4–judged pairwise comparisons. |
DF-RAG: Query-Aware Diversity for Retrieval-Augmented Generation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Retrieval-augmented generation (RAG) is a common technique for grounding language models in domain-specific information. |
| Approach: | They propose a new retrieval technique that incorporates diversity into the retrieval step to improve performance on reasoning-intensive QA benchmarks. |
| Outcome: | The proposed method outperforms baselines on reasoning-intensive QA benchmarks by 4–10%. |
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
Jiaxing Liu, Qi Qi, Haifeng Sun, Dunjun Li, Zirui Zhuang, Bo He, Xiang Yang, Cong Liu, Jianxin Liao, Jingyu Wang
| 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. |