Papers by Kai Gao
Consistent Representation Learning for Continual Relation Extraction (2022.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to train relation extraction models overfit memory samples and perform poorly on imbalanced datasets. |
| Approach: | They propose a method which uses contrastive learning and knowledge distillation to train a model on data with new relations while avoiding forgetting old ones. |
| Outcome: | The proposed method significantly outperforms state-of-the-art baselines and yields strong robustness on the imbalanced datasets. |
Capability Salience Vector: Fine-grained Alignment of Loss and Capabilities for Downstream Task Scaling Law (2025.acl-long)
Copied to clipboard
Qiming Ge, Shuhao Xing, Songyang Gao, Yunhua Zhou, Yicheng Zou, Songyang Zhang, Zhi Chen, Hang Yan, Qi Zhang, Qipeng Guo, Kai Chen
| Challenge: | Large language models have demonstrated impressive performance across a wide range of tasks, but this achievement comes with the trade-off of significant computational demands. |
| Approach: | They propose a scaling law that decomposes the overall validation loss and assigns different importance weights to tokens to assess a specific meta-capability. |
| Outcome: | The proposed model can predict the loss trending of models across different levels of computation without a gap between validation loss and model's downstream capabilities. |
OpenVNA: A Framework for Analyzing the Behavior of Multimodal Language Understanding System under Noisy Scenarios (2024.acl-demos)
Copied to clipboard
| Challenge: | OpenVNA is an open-source framework for analyzing the behavior of multimodal language understanding systems under noisy conditions. |
| Approach: | They propose to use OpenVNA to analyze behavior of multimodal language understanding systems under noisy conditions. |
| Outcome: | The proposed framework provides high flexibility and extensibility, enabling customization with user-defined noise types and models. |
Layer-wise Importance Matters: Less Memory for Better Performance in Parameter-efficient Fine-tuning of Large Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Parameter-Efficient Fine-Tuning (PEFT) methods have gained popularity for adapting pre-trained Large Language Models (LLMs) to downstream tasks. |
| Approach: | They propose a method to optimize the importance of full layers with layer-wise importance scoring by leveraging the estimated importance scores. |
| Outcome: | The proposed method is compatible with PEFT methods that operate on a per-layer basis and achieves better performance. |
ReviewRL: Towards Automated Scientific Review with RL (2025.emnlp-main)
Copied to clipboard
Sihang Zeng, Kai Tian, Kaiyan Zhang, Yuru Wang, Junqi Gao, Runze Liu, Sa Yang, Jingxuan Li, Xinwei Long, Jiaheng Ma, Biqing Qi, Bowen Zhou
| Challenge: | Existing automated review systems struggle with factual accuracy, rating consistency, and analytical depth. |
| Approach: | They propose a framework for generating comprehensive and factually grounded scientific paper reviews using supervised fine-tuning and reinforcement learning. |
| Outcome: | The proposed framework outperforms existing methods on ICLR 2025 papers. |
SDBench: A Survey-based Domain-specific LLM Benchmarking and Optimization Framework (2025.acl-long)
Copied to clipboard
| Challenge: | acquiring domain-specific knowledge often requires professional expert manpower. |
| Approach: | They propose a generic framework for generating evaluation datasets for domain-specific LLMs. |
| Outcome: | The proposed framework reduces the reliance on expert manpower while ensuring that the collected data is uniformly distributed. |
CityEQA: A Hierarchical LLM Agent on Embodied Question Answering Benchmark in City Space (2025.emnlp-main)
Copied to clipboard
Yong Zhao, Kai Xu, Zhengqiu Zhu, Yue Hu, Zhiheng Zheng, Yingfeng Chen, Yatai Ji, Chen Gao, Yong Li, Jincai Huang
| Challenge: | Embodied Question Answering (EQA) tasks are primarily focused on indoor environments, leaving the complexities of urban settings unexplored. |
| Approach: | They propose a task where an embodied agent answers open-vocabulary questions in dynamic city spaces. |
| Outcome: | The proposed agent achieves 60.7% of human-level answering accuracy compared to baselines . the proposed agent outperforms existing agents in open-ended city spaces . |
Consolidation or Adaptation? PRISM: Disentangling SFT and RL Data via Gradient Concentration (2026.acl-long)
Copied to clipboard
Yang Zhao, Yangou Ouyang, Xiao Ding, Hepeng Wang, Bibo Cai, Kai Xiong, Jinglong Gao, Zhouhao Sun, Li Du, Bing Qin, Ting Liu
| Challenge: | Existing data arbitration strategies for large language model training rely on surface-level heuristics that fail to diagnose intrinsic learning needs. |
| Approach: | They propose a framework that arbitrates data based on its degree of cognitive conflict with the model's existing knowledge. |
| Outcome: | Extensive experiments on WebShop and ALFWorld show that PRISM outperforms state-of-the-art hybrid methods while reducing computational costs by up to 3.22 . |
Beyond Similarity: A Gradient-based Graph Method for Instruction Tuning Data Selection (2025.acl-long)
Copied to clipboard
Yang Zhao, Li Du, Xiao Ding, Yangou Ouyang, Hepeng Wang, Kai Xiong, Jinglong Gao, Zhouhao Sun, Dongliang Xu, Qing Yang, Dongchen Li, Bing Qin, Ting Liu
| Challenge: | Existing methods for selecting training data from general datasets fail to account for the joint distribution of instructions, resulting in inefficient learning and suboptimal knowledge transfer. |
| Approach: | They propose a method that constructs a mixed gradient-based instruction graph to capture the joint distribution and interdependencies among instructions. |
| Outcome: | The proposed method outperforms existing methods on domain adaptation tasks and in complex, data-scarce scenarios. |
M-SENA: An Integrated Platform for Multimodal Sentiment Analysis (2022.acl-demo)
Copied to clipboard
| Challenge: | M-SENA is an open-source platform for multimodal sentiment analysis. |
| Approach: | They propose to use a platform for multimodal sentiment analysis to facilitate advanced research by providing flexible toolkits, reliable benchmarks, and intuitive demonstrations. |
| Outcome: | The proposed framework provides reliable benchmarks and baseline results of different modality features and MSA benchmarks. |
BadWindtunnel: Defending Backdoor in High-noise Simulated Training with Confidence Variance (2025.findings-acl)
Copied to clipboard
| Challenge: | Current backdoor attack defenders in NLP typically involve data reduction or model pruning, risking losing crucial information. |
| Approach: | They propose a backdoor defender that allows precise control over training conditions to model backdoor learning behavior without affecting the final model. |
| Outcome: | The proposed model reduces the backdoor learning behavior without affecting the final model. |
Firewall Routing: Blocking Leads to Better Hybrid Inference for LLMs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models have significantly enhanced performance across various NLP tasks . high computational costs and latency associated with deploying such models pose bottlenecks . |
| Approach: | They propose a dynamic hybrid inference framework that efficiently selects between a strong and a weak LLM based on the complexity of the query. |
| Outcome: | The proposed method outperforms existing routing strategies by up to 5.29% in APGR . large models often introduce higher latency, making them less suitable for real-time or resource-constrained applications. |
CompassVerifier: A Unified and Robust Verifier for LLMs Evaluation and Outcome Reward (2025.emnlp-main)
Copied to clipboard
Shudong Liu, Hongwei Liu, Junnan Liu, Linchen Xiao, Songyang Gao, Chengqi Lyu, Yuzhe Gu, Wenwei Zhang, Derek F. Wong, Songyang Zhang, Kai Chen
| Challenge: | Existing approaches lack robustness to handle complex edge cases and generalizability across different domains. |
| Approach: | They develop an accurate and lightweight verifier model for evaluation and outcome reward that matches unstructured outputs against standard answers. |
| Outcome: | The proposed model can process multiple answer types including multi-subproblems, formulas, and sequence answers while identifying abnormal/invalid responses. |
SWE-Fixer: Training Open-Source LLMs for Effective and Efficient GitHub Issue Resolution (2025.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models excel in code generation benchmarks, but these benchmarks focus on single-file scenarios with constrained context scope. |
| Approach: | They propose an open-source framework to effectively resolve GitHub issues using a code file retrieval module and a model-based code editing module. |
| Outcome: | The proposed approach achieves state-of-the-art performance on two GitHub benchmarks. |
ARM: An Alignment-and-Replacement Module for Chinese Spelling Check Based on LLMs (2024.emnlp-main)
Copied to clipboard
| Challenge: | Chinese Spelling Check (CSC) aims to identify and correct spelling errors in Chinese texts, where enhanced semantic understanding of a sentence can significantly improve correction accuracy. |
| Approach: | They propose a plug-and-play Alignment-and -Replacement module that enhances existing Chinese CSC models without retraining or fine-tuning. |
| Outcome: | The proposed module improves existing models while reducing retraining and fine-tuning. |
GradOT: Training-free Gradient-preserving Offsite-tuning for Large Language Models (2025.acl-long)
Copied to clipboard
Kai Yao, Zhaorui Tan, Penglei Gao, Lichun Li, Kaixin Wu, Yinggui Wang, Yuan Zhao, Yixin Ji, Jianke Zhu, Wei Wang
| Challenge: | Existing methods for offsite-tuning of large language models require high computational costs and lack theoretical analysis. |
| Approach: | They propose an offsite-tuning approach that selectively applies compression techniques such as rank compression and channel pruning to preserve the gradients of fine-tuned adapters while ensuring privacy. |
| Outcome: | The proposed method surpasses existing OT methods in privacy protection and model performance. |
Reinforcement Learning on Pre-Training Data (2026.acl-long)
Copied to clipboard
Siheng Li, Kejiao Li, Zenan Xu, Guanhua Huang, Kun Li, Haoyuan Wu, null Wujiajia, Zihao Zheng, Chenchen Zhang, Kun Shi, Xue Gong, Qi Yi, Ruibin Xiong, Tingqiang Xu, Yuhao Jiang, Jianfeng Yan, Yuyuan Zeng, Guanghui Xu, Jinbao Xue, Zhijiang xu, Zheng Fang, Shuai LI, Qibin Liu, Xiaoxue Li, Zhuoyu Li, Yangyu Tao, Fei Gao, Cheng Jiang, Bochao Wang, Kai Liu, Jianchen Zhu, Wai Lam, Bo Zhou, Di Wang
| Challenge: | Recent progress in large language models is driven by scaling of training compute through pre-training with nexttoken prediction (NTP) or post-training (RL) Pre-training using NTP enables models to acquire extensive knowledge and skills from general data, but it suffers from data inefficiency and catastrophic forgetting in continual learning settings. |
| Approach: | They propose to scale training compute through pre-training with next-token prediction (NTP) or post-training by scaling reinforcement learning (RL) to improve learning from general data. |
| Outcome: | Experiments on multiple benchmarks and models show that the proposed approach improves continual pre-training and provides a strong foundation for post-training on Qwen3-8B-Base. |
MAESTRO: Meta-learning Adaptive Estimation of Scalarization Trade-offs for Reward Optimization (2026.acl-long)
Copied to clipboard
Yang Zhao, Hepeng Wang, Xiao Ding, Yangou Ouyang, Bibo Cai, Kai Xiong, Jinglong Gao, Zhouhao Sun, Li Du, Bing Qin, Ting Liu
| Challenge: | Group-Relative Policy Optimization (GRPO) has emerged as an efficient paradigm for aligning Large Language Models (LLMs), but its efficacy is confined to domains with verifiable ground truths. |
| Approach: | They propose a meta-cognitive orchestration layer that treats reward scalarization as a dynamic latent policy, leveraging the model’s terminal hidden states as 'a semantic bottleneck' . Across seven benchmarks, MAESTRO consistently outperforms single-reward and static multi-objective baselines while preserving the efficiency advantages of GRPO. |
| Outcome: | The proposed model outperforms single-reward and static multi-objective baselines while preserving efficiency advantages. |
Scaling up the State Size of RNN LLMs for Long-Context Scenarios (2025.acl-long)
Copied to clipboard
| Challenge: | Existing RNN-based LLMs struggle with long-context scenarios due to their quadratic computational complexity and linear memory requirements. |
| Approach: | They propose an efficient scaling method to scale RNN models to match the 2k context length of Transformers with small parameters overhead. |
| Outcome: | The proposed method improves long-context understanding and improves performance on FDA recall-intensive tasks. |
Text2Tree: Aligning Text Representation to the Label Tree Hierarchy for Imbalanced Medical Classification (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches to medical text classification are struggling with imbalanced data distribution and rare labels. |
| Approach: | They propose a framework-agnostic algorithm that only utilizes internal label hierarchy in training deep learning models. |
| Outcome: | The proposed approach performs better on public datasets and real-world medical records than existing methods. |
Are Your LLMs Capable of Stable Reasoning? (2025.findings-acl)
Copied to clipboard
Junnan Liu, Hongwei Liu, Linchen Xiao, Ziyi Wang, Kuikun Liu, Songyang Gao, Wenwei Zhang, Songyang Zhang, Kai Chen
| Challenge: | Existing evaluation protocols and metrics do not capture the full spectrum of LLM capabilities, especially in complex reasoning tasks. |
| Approach: | They propose a new evaluation metric that continuously assesses model performance across multiple sampling attempts, quantifying both the model’s potential capabilities and operational consistency. |
| Outcome: | The proposed evaluation metric measures model performance across multiple sampling attempts and provides comprehensive insights into their potential capabilities and operational consistency. |
MDTeamGPT: Mitigating Context Collapse and Enabling Self-Evolution in Medical Multi-Agent Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Long, multi-round, multirole interaction trajectories lead to severe information dilution and context window overload, triggering context collapse which destabilizes reasoning. |
| Approach: | They propose a multi-agent framework that compresses and reorganizes multi-round consensus. |
| Outcome: | The proposed framework outperforms baselines across text-based and multimodal tasks while demonstrating superior diagnostic performance and stability in complex clinical scenarios. |
ProReason: Multi-Modal Proactive Reasoning with Decoupled Eyesight and Wisdom (2025.emnlp-main)
Copied to clipboard
Jingqi Zhou, Sheng Wang, Jingwei Dong, Kai Liu, Lei Li, Jiahui Gao, Jiyue Jiang, Lingpeng Kong, Chuan Wu
| Challenge: | Large vision-language models often prioritize language knowledge over image information on visual reasoning tasks, incurring performance degradation. |
| Approach: | They propose a visual reasoning framework that decouples vision-reasoning capabilities and multi-run proactive perception. |
| Outcome: | The proposed framework outperforms existing models on benchmarks for open-source and closed-source models with 13.2% performance gain. |
Multi-Defendant Legal Judgment Prediction via Hierarchical Reasoning (2023.findings-emnlp)
Copied to clipboard
Yougang Lyu, Jitai Hao, Zihan Wang, Kai Zhao, Shen Gao, Pengjie Ren, Zhumin Chen, Fang Wang, Zhaochun Ren
| Challenge: | Existing methods for predicting judgment results for multiple defendants are ineffective. |
| Approach: | They propose a method to predict the judgment results for each defendant in multi-defendant cases . they formalize the multi-diffendant judgment process as hierarchical reasoning chains . |
| Outcome: | The proposed method can predict the judgment results for multiple defendants in multi-defendant cases. |
GAST: Gradient-aligned Sparse Tuning of Large Language Models with Data-layer Selection (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing methods focus on layer-selective and data-selectory fine-tuning, but ignore the fact that different data points contribute varying degrees to distinct model layers. |
| Approach: | They propose a method that performs selective fine-tuning at both data and layer dimensions as integral components of a unified optimization strategy. |
| Outcome: | Experiments show that the proposed method outperforms baseline methods in terms of performance and performance. |
BrowseComp-Plus: A Fair and Disentangled Evaluation Benchmark for Deep Search Agents (2026.acl-long)
Copied to clipboard
Zijian Chen, Xueguang Ma, Shengyao Zhuang, Ping Nie, Kai Zou, Sahel Sharifymoghaddam, Andrew Liu, Joshua Green, Kshama Patel, Ruoxi Meng, Mingyi Su, Yanxi Li, Haoran Hong, Xinyu Shi, Xuye Liu, Hosna Oyarhoseini, Nandan Thakur, Crystina Zhang, Luyu Gao, Wenhu Chen, Jimmy Lin
| Challenge: | Existing benchmarks for deep search agents rely on blackbox web search APIs . dynamic and opaque web APIs hinder reproducibility and fair comparisons - authors . |
| Approach: | They propose a benchmark that employs a fixed corpus for controlled retrieval for deep search agents. |
| Outcome: | The new benchmark shows that agents that combine large language models with retrieval tools excel at complex, reasoning-intensive queries. |
Com2 : A Causal-Guided Benchmark for Exploring Complex Commonsense Reasoning in Large Language Models (2025.acl-long)
Copied to clipboard
Kai Xiong, Xiao Ding, Yixin Cao, Yuxiong Yan, Li Du, Yufei Zhang, Jinglong Gao, Jiaqian Liu, Bing Qin, Ting Liu
| Challenge: | Existing works focus on complex tasks like math and code, while complex commonsense reasoning remains underexplored due to its uncertainty and lack of structure. |
| Approach: | They propose to build a benchmark for large language models based on complex commonsense reasoning based upon causal event graphs and causal theory. |
| Outcome: | The proposed benchmark combines a complex commonsense reasoning benchmark with a detective story to achieve a more challenging subset. |
ReAL: How Can LLMs Simulate the Real Teacher? Retrieval-enhanced Agent for Adaptive Learning (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Prior methods model learner-item interactions based only on ID sequences, leading to insufficient use of both learner and item information. |
| Approach: | They propose a Retrieval-enhanced Agent for Adaptive Learning powered by large language models to simulate teacher decision-making with extensive prior knowledge and teaching experience. |
| Outcome: | The proposed model outperforms existing models on three real-world datasets in both internal and external perspectives. |
Unsupervised Multimodal Clustering for Semantics Discovery in Multimodal Utterances (2024.acl-long)
Copied to clipboard
| Challenge: | Existing methods for semantics discovery focus on text, video, and audio, failing to leverage the rich multimodal information in the real world. |
| Approach: | They propose a method to construct augmentation views for multimodal data and use them to perform pre-training to establish well-initialized representations for subsequent clustering. |
| Outcome: | The proposed method improves on benchmark multimodal intent and dialogue act datasets by 2-6% over state-of-the-art methods. |
Mixture of insighTful Experts (MoTE): The Synergy of Reasoning Chains and Expert Mixtures in Self-Alignment (2025.acl-long)
Copied to clipboard
Zhili Liu, Yunhao Gou, Kai Chen, Lanqing Hong, Jiahui Gao, Fei Mi, Yu Zhang, Zhenguo Li, Xin Jiang, Qun Liu, James Kwok
| Challenge: | Recent studies show that reasoning abilities contribute significantly to model safety, while integrating Mixture-of-Experts (MoE) architectures can further enhance alignment. |
| Approach: | They propose a framework that synergistically combines reasoning chains and expert mixtures to improve self-alignment. |
| Outcome: | The proposed framework improves model safety, jailbreak resistance, and over-refusal capabilities, achieving performance comparable to OpenAI’s state-of-the-art o1 model. |
TEXTOIR: An Integrated and Visualized Platform for Text Open Intent Recognition (2021.acl-demo)
Copied to clipboard
| Challenge: | TEXTOIR is the first integrated platform for text open intent recognition . currently, many dialogue systems are limited to handle the uncertain open intents . |
| Approach: | TEXTOIR is the first integrated platform for text open intent recognition . it is composed of two main modules: open intent detection and open intent discovery . authors propose a framework to implement a complete process to identify known intents and discover open intents . |
| Outcome: | TEXTOIR is the first integrated and visualized platform for text open intent recognition . it integrates state-of-the-art algorithms and benchmark intent datasets . however, there are still some issues, which bring difficulties for future research . |
Continual Machine Reading Comprehension via Uncertainty-aware Fixed Memory and Adversarial Domain Adaptation (2022.findings-naacl)
Copied to clipboard
| Challenge: | Existing stationary-trained MRC systems are usually trained with in-domain data but are applied to new domain data. |
| Approach: | They propose a continual machine reading comprehension model with uncertainty-aware fixed memory and adversarial domain adaptation that keeps a stable understanding by learning both memory and new domain data. |
| Outcome: | The proposed model is superior to strong baselines and has a substantial incremental learning ability without catastrophically forgetting under two different continual MRC settings. |