Papers by Xunliang Cai
VANE: Guiding High-Value Exploration in RLVR via Outcome-Process Novelty Shaping (2026.findings-acl)
Copied to clipboard
| Challenge: | Extensive experiments on large-scale mathematical reasoning and out-of-distribution tasks demonstrate the effectiveness and generalization of the proposed method. |
| Approach: | They propose a method that quantifies novelty across the outcome space and semantic process space by using reward or solution divergence. |
| Outcome: | Experiments on Qwen2.5-Math-7B demonstrate the proposed method is general and efficient. |
Don’t Half-listen: Capturing Key-part Information in Continual Instruction Tuning (2025.acl-long)
Copied to clipboard
Yongquan He, Wenyuan Zhang, Xuancheng Huang, Peng Zhang, Lingxun Meng, Xiang Zhou, Ke Zeng, Xunliang Cai
| Challenge: | Existing methods to improve instruction tuning for large language models may cause catastrophic forgetting (CF) CF is a problem where previously learned abilities are degraded . |
| Approach: | They propose a continual instruction tuning method that uses key-part information gain to replay data and refine training objective. |
| Outcome: | The proposed method achieves superior performance on both seen and held-out tasks. |
Learning or Self-aligning? Rethinking Instruction Fine-tuning (2024.acl-long)
Copied to clipboard
| Challenge: | Instruction fine-tuning (IFT) is a crucial phase in building large language models (LLMs). |
| Approach: | They propose a knowledge intervention framework to decouple the potential underlying factors of IFT and enable individual analysis of different factors. |
| Outcome: | The proposed framework decouples the potential underlying factors of IFT, enabling individual analysis of different factors. |
Graph-Structured Speculative Decoding (2024.findings-acl)
Copied to clipboard
Zhuocheng Gong, Jiahao Liu, Ziyue Wang, Pengfei Wu, Jingang Wang, Xunliang Cai, Dongyan Zhao, Rui Yan
| Challenge: | Speculative decoding is a promising technique to accelerate the inference of Large Language Models. |
| Approach: | They propose a method that uses a token graph to record multiple sequence hypotheses within a single draft stage. |
| Outcome: | The proposed method significantly accelerates the inference of Large Language Models (LLMs) it allows the LLM to choose from and select the longest sequence that meets its standards. |
Leveraging Dual Process Theory in Language Agent Framework for Real-time Simultaneous Human-AI Collaboration (2025.acl-long)
Copied to clipboard
Shao Zhang, Xihuai Wang, Wenhao Zhang, Chaoran Li, Junru Song, Tingyu Li, Lin Qiu, Xuezhi Cao, Xunliang Cai, Wen Yao, Weinan Zhang, Xinbing Wang, Ying Wen
| Challenge: | Large language models (LLMs) excel in turn-by-turn human-AI collaboration but struggle with simultaneous tasks requiring real-time interaction. |
| Approach: | They propose a language agent framework that integrates *System 1* and *System 2* for efficient real-time simultaneous human-AI collaboration. |
| Outcome: | The proposed framework improves on existing LLM-based agents and human collaborators by integrating Theory of Mind and asynchronous reflection to infer human intentions and perform reasoning-based autonomous decisions. |
AMO-Bench: Large Language Models Still Struggle in High School Math Competitions (2026.findings-acl)
Copied to clipboard
Junlin Liu, Shengnan An, Shuang Zhou, Dan Ma, Yehao Lin, Xinxuan Lv, Xuanlin Wang, Xiaoyu Li, Ziwen Wang, Xuezhi Cao, Xunliang Cai
| Challenge: | Existing benchmarks for mathematical reasoning are becoming less effective due to performance saturation. |
| Approach: | They propose to use a mathematical reasoning benchmark with Olympiad difficulty to evaluate top-tier LLMs. |
| Outcome: | The proposed benchmarks are cross-validated by experts to meet IMO difficulty standards and entirely original problems to prevent performance leakages from data memorization. |
Turning Failures into Value: Negative Experience Replay for RLVR via Confidence Gating and Boundary Failure Sampling (2026.acl-long)
Copied to clipboard
| Challenge: | Existing experience replay methods for RLVR ignore sample inefficiency . expensive reasoning trajectories are discarded immediately after a single gradient update . |
| Approach: | They propose a method to replay failure trajectories to improve model refinement . they propose 'nexGRPO' which employs mid-confidence gating to filter invalid noise and saturated errors. |
| Outcome: | The proposed model outperforms strong baaselines and achieves improved out-of-distribution generalization. |
Improving Input-label Mapping with Demonstration Replay for In-context Learning (2023.findings-emnlp)
Copied to clipboard
| Challenge: | In-context learning (ICL) is an emerging capability of large autoregressive language models where a few demonstrations are appended to the input to enhance the model’s understanding of downstream NLP tasks without directly adjusting the model parameters. |
| Approach: | They propose a method where a few demonstrations are appended to the input to enhance the model's understanding of downstream NLP tasks without directly adjusting the model parameters. |
| Outcome: | The proposed method significantly improves the input-label mapping in ICL demonstrations. |
From Paraphrasing to Semantic Parsing: Unsupervised Semantic Parsing via Synchronous Semantic Decoding (2021.acl-long)
Copied to clipboard
Shan Wu, Bo Chen, Chunlei Xin, Xianpei Han, Le Sun, Weipeng Zhang, Jiansong Chen, Fan Yang, Xunliang Cai
| Challenge: | Experimental results show that Synchronous Semantic Decoding (SSD) can achieve state-of-the-art unsupervised semantic parsing performance on multiple datasets. |
| Approach: | They propose an unsupervised method which solves the semantic gap and the structure gap by leveraging paraphrasing and grammar-constrained decoding. |
| Outcome: | The proposed method can solve the semantic gap and structure gap on multiple datasets. |
Large-Scale Diverse Synthesis for Mid-Training (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing data synthesis methods generate simplistic and homogeneous QA pairs with limited scale and diversity. |
| Approach: | They propose a framework to synthesize large-scale, diverse, and high-quality QA data for mid-training. |
| Outcome: | The proposed framework improves on 500B-token BoostQA data over pre-training benchmarks. |
Instance-level Randomization: Toward More Stable LLM Evaluations (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Evaluations of large language models suffer from instability, where small changes of random factors can lead to drastic fluctuations of scores and even model rankings. |
| Approach: | They propose an instance-level randomization method to reduce variance and improve fairness in evaluations by randomizing all factors that affect evaluation scores for every single instance. |
| Outcome: | The proposed method reduces variance and improves fairness in model comparisons by using instance-level randomization. |
Counteracting the Matthew Effect in Self-Improvement of LVLMs through Head-Tail Re-balancing (2026.acl-long)
Copied to clipboard
Xin Guo, Zhiheng Xi, Yiwen Ding, Yitao Zhai, Xiaowei Shi, Xunliang Cai, Tao Gui, Qi Zhang, Xuanjing Huang
| Challenge: | Large vision language models have impressive reasoning capabilities across complex multimodal tasks. |
| Approach: | They propose to use distribution-reshaping and trajectory-rebalancing to improve visual reasoning capabilities. |
| Outcome: | Experiments on Qwen2-VL-7B-Instruct and InternVL2.5-4B models show that their methods outperform baselines by 3.86 points. |
Mitigating Tail Narrowing in LLM Self-Improvement via Socratic-Guided Sampling (2025.naacl-long)
Copied to clipboard
Yiwen Ding, Zhiheng Xi, Wei He, Lizhuoyuan Lizhuoyuan, Yitao Zhai, Shi Xiaowei, Xunliang Cai, Tao Gui, Qi Zhang, Xuanjing Huang
| Challenge: | Large language models (LLMs) generate solutions themselves and iteratively train on filtered, high-quality rationales, but performance reaches a ceiling after a few iterations. |
| Approach: | They propose a strategy to improve the efficiency of sampling heavy-tailed data by using Socratic-style guidance signals to help LLMs reasoning with complex queries. |
| Outcome: | The proposed approach is effective on difficult queries and on held-out tasks, while requiring human supervision. |
Multi-Programming Language Sandbox for LLMs (2025.acl-demo)
Copied to clipboard
Shihan Dou, Jiazheng Zhang, Jianxiang Zang, Yunbo Tao, Weikang Zhou, Haoxiang Jia, Shichun Liu, Yuming Yang, Shenxi Wu, Zhiheng Xi, Muling Wu, Rui Zheng, Changze Lv, Limao Xiong, Shaoqing Zhang, Lin Zhang, Wenyu Zhan, Rongxiang Weng, Jingang Wang, Xunliang Cai, Yueming Wu, Ming Wen, Yixin Cao, Tao Gui, Xipeng Qiu, Qi Zhang, Xuanjing Huang
| Challenge: | MPLSandbox is an out-of-the-box multi-programming language sandbox designed to provide unified and comprehensive feedback from compiler and analysis tools for Large Language Models (LLMs). |
| Approach: | They propose a multi-programming language sandbox that provides unified feedback from compilers and analysis tools for Large Language Models. |
| Outcome: | The proposed multi-language sandbox can provide comprehensive feedback from compilers and analysis tools for large language models (LLMs). |
LinkQA: Synthesizing Diverse QA from Multiple Seeds Strongly Linked by Knowledge Points (2026.acl-long)
Copied to clipboard
| Challenge: | Existing training data is limited in high-quality training data, limiting the ability to produce high-performance LLMs. |
| Approach: | They propose a KP-graph-based synthesis framework that extracts KPs from QA seed data and constructs a graph of KP data from multiple seeds strongly linked by KP. |
| Outcome: | The proposed framework enables flexible control over discipline and difficulty distributions while balancing KP coverage and popularity. |
FRAME: Boosting LLMs with A Four-Quadrant Multi-Stage Pretraining Strategy (2025.findings-acl)
Copied to clipboard
Xuemiao Zhang, Feiyu Duan, Xu Liangyu, Yongwei Zhou, Sirui Wang, Rongxiang Weng, Jingang Wang, Xunliang Cai
| Challenge: | Multi-stage pretraining methods lack quantitative criteria for data partitioning and instead rely on intuitive heuristics. |
| Approach: | They propose a Four-quadRAnt Multi-stage prEtraining strategy that partitions data into four quadrants to achieve significant loss reductions four times. |
| Outcome: | The proposed strategy achieves 16.8% improvement over random across MMLU and CMMLU for the 3B model. |
The Role of Visual Modality in Multimodal Mathematical Reasoning: Challenges and Insights (2025.acl-long)
Copied to clipboard
Yufang Liu, Yao Du, Tao Ji, Jianing Wang, Yang Liu, Yuanbin Wu, Aimin Zhou, Mengdi Zhang, Xunliang Cai
| Challenge: | Existing models that leverage visual information do not improve math reasoning performance . authors suggest that visual information is important for multimodal reasoning . |
| Approach: | They propose a dataset to require image reliance for problem-solving and challenge models with similar, yet distinct, images that change the correct answer. |
| Outcome: | The proposed model performance is unaffected by changes to or removal of images in the dataset. |
DIFFA-2: A Practical Diffusion Large Language Model for General Audio Understanding (2026.findings-acl)
Copied to clipboard
| Challenge: | Autoregressive (AR) large audio language models are expensive in data and computation . prior work shows diffusion-based LALMs can improve audio understanding under matched settings . |
| Approach: | They propose a diffusion-based LALM that upgrades the speech encoder and employs dual semantic and acoustic adapters. |
| Outcome: | a new model improves over existing autoregressive large language models and is competitive to strong AR models . the proposed model can make use of limited training data and improve inference efficiency . a recent study shows that diffusion-based models can improve audio understanding . |
Leveraging Unpaired Feedback for Long-Term LLM-based Recommendation Tuning (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a recent study highlights unpaired feedback as a key challenge for long-term LLM-based recommenders . unpaired user feedback is crucial for improving LLMs in dynamic user environments, authors say . |
| Approach: | They propose a framework that incorporates unpaired feedback into LLMs to improve long-term recommendation performance. |
| Outcome: | The proposed framework improves long-term recommendation performance by incorporating unpaired feedback without requiring paired supervision. |
A Reasoner for Real-World Event Detection: Scaling Reinforcement Learning via Adaptive Perplexity-Aware Sampling Strategy (2025.emnlp-industry)
Copied to clipboard
| Challenge: | Existing methods for abnormal event detection face two predominant limitations . existing methods rely on specialized small models and are limited by performance bottlenecks . |
| Approach: | They propose a framework that leverages the advanced reasoning capabilities of large language models for abnormal event detection. |
| Outcome: | The proposed framework achieves the highest F1 score and an average improvement of 9.59% in OOD transfer tests. |
Not All Contexts Are Equal: Teaching LLMs Credibility-aware Generation (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing RAG paradigms suffer from the impact of flawed information introduced during the retrieval phrase, thereby diminishing the reliability and correctness of the generated output. |
| Approach: | They propose a framework that empowers models to discern and process information based on its credibility. |
| Outcome: | The proposed framework outperforms existing models with retrieval augmentation and exhibits robustness despite increasing noise in the context. |
LogicPro: Improving Complex Logical Reasoning via Program-Guided Learning (2025.acl-long)
Copied to clipboard
Jin Jiang, Yuchen Yan, Yang Liu, Jianing Wang, Shuai Peng, Xunliang Cai, Yixin Cao, Mengdi Zhang, Liangcai Gao
| Challenge: | LogicPro is a data synthesis method that uses LeetCode-style algorithm problems and their corresponding Program solutions to generate complex logic data. |
| Approach: | They propose a new method which leverages LeetCode-style algorithm Problems and their corresponding Program solutions to synthesize complex logic data in text format. |
| Outcome: | The proposed method outperforms existing models for BBH27, LogicBench, DROP, AR-LSAT, and GSM8K, and a wide range of reasoning datasets. |
How Do Your Code LLMs perform? Empowering Code Instruction Tuning with Really Good Data (2024.emnlp-main)
Copied to clipboard
Yejie Wang, Keqing He, Dayuan Fu, Zhuoma GongQue, Heyang Xu, Yanxu Chen, Zhexu Wang, Yujia Fu, Guanting Dong, Muxi Diao, Jingang Wang, Mengdi Zhang, Xunliang Cai, Weiran Xu
| Challenge: | Recent research has shown that code pre-trained models improve coding capabilities. |
| Approach: | They propose a code data pruning strategy to identify which datasets are high-quality code instruction data. |
| Outcome: | The proposed model achieves state-of-the-art performance using fewer training data. |
ReMamba: Equip Mamba with Effective Long-Sequence Modeling (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Mamba models demonstrate superior inference efficiency and competitive performance on short-context tasks, but their capacity to comprehend long contexts is limited compared to transformer-based models. |
| Approach: | They propose a model which incorporates selective compression and adaptation techniques within a two-stage re-forward process, incurring minimal additional inference costs overhead. |
| Outcome: | The proposed model improves on the LongBench and L-Eval benchmarks by 3.2 and 1.6 points and attains performance almost on par with same-size transformer models. |
Attribution-Based Analysis and Optimization of Modular Agentic Workflows (2026.findings-acl)
Copied to clipboard
Yingxuan Yang, Bo Huang, Siyuan Qi, Chao Feng, Haoyi Hu, Yuxuan Zhu, Jinbo Hu, Haoran Zhao, Ziyi He, Xiao Liu, ZongYu Wang, Muning Wen, Lin Qiu, Xuezhi Cao, Xunliang Cai, Yong Yu, Weinan Zhang
| Challenge: | Large Language Models (LLMs) have driven the rise of agentic workflows . yet, how can we attribute performance gains to individual upgrades and their interactions? |
| Approach: | They propose a game-theoretic framework that models component upgrades as players and evaluates component coalitions to compute Shapley values. |
| Outcome: | The proposed framework provides interaction-aware attribution and recommendation for model allocation under a fixed workflow structure. |
PaTaRM: Bridging Pairwise and Pointwise Signals via Preference-Aware Task-Adaptive Reward Modeling (2026.acl-long)
Copied to clipboard
| Challenge: | Existing reward models lack generative and reasoning capabilities, resulting in poor performance. |
| Approach: | They propose a reward-aware task-adaptive reward model that enables pointwise training using readily available pairwise data via a novel Preference-Aware Reward mechanism. |
| Outcome: | The proposed reward model achieves an average relative improvement of 8.7% over the base models on RewardBench and RMBench. |
DolphCoder: Echo-Locating Code Large Language Models with Diverse and Multi-Objective Instruction Tuning (2024.acl-long)
Copied to clipboard
Yejie Wang, Keqing He, Guanting Dong, Pei Wang, Weihao Zeng, Muxi Diao, Weiran Xu, Jingang Wang, Mengdi Zhang, Xunliang Cai
| Challenge: | Numerous code large language models (LLMs) have been proposed to enhance code generation performance. |
| Approach: | They propose a diverse instruction model DolphCoder with self-evaluating for code generation that learns diverse instruction targets and combines a code evaluation objective to enhance its code generation ability. |
| Outcome: | The proposed model achieves superior performance on the HumanEval and MBPP benchmarks, demonstrating new insights for future code instruction tuning work. |
Revisiting Scaling Laws for Language Models: The Role of Data Quality and Training Strategies (2025.acl-long)
Copied to clipboard
| Challenge: | Existing scaling laws suggest augmenting model size and training data results in enhanced performance, but recent studies reveal deviations, particularly in large language models, where performance improvements decelerate—a phenomenon known as sub-scaling. |
| Approach: | They propose a sub-optimal scaling law that better predicts performance in sub-scaling regimes by examining data quality and training strategies. |
| Outcome: | The proposed scaling law better predicts performance in sub-scaling regimes, highlighting the importance of data quality and diversity. |
Why Not Act on What You Know? Unleashing Safety Potential of LLMs via Self-Aware Guard Enhancement (2025.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated impressive capabilities across various tasks but are vulnerable to meticulously crafted jailbreak attacks. |
| Approach: | They propose a training-free defense strategy to align LLMs’ strong safety discrimination performance with their relatively weaker safety generation ability. |
| Outcome: | The proposed strategy achieves an average 99% success rate against numerous complex and covert jailbreak methods while maintaining helpfulness on general benchmarks. |
MASPO: Unifying Gradient Utilization, Probability Mass, and Signal Reliability for Robust and Sample-Efficient LLM Reasoning (2026.acl-long)
Copied to clipboard
Xiaoliang Fu, Jiaye Lin, Yangyi Fang, Binbin Zheng, Chaowen Hu, Zekai Shao, Cong Qin, Lu Pan, Ke Zeng, Xunliang Cai
| Challenge: | Existing RLVR algorithms rely on rigid, uniform, and symmetric trust region mechanisms . current algorithms lack robustness, asymmetric signal reliability and inefficient gradient utilization . |
| Approach: | They propose a framework to harmonize three dimensions of RLVR algorithms, a paper argues . a binary cutoff is used to discard valuable reinforcement signals, they argue . |
| Outcome: | The proposed framework outperforms baselines in evaluating a robust RLVR solution. |
AMoPO: Adaptive Multi-objective Preference Optimization without Reward Models and Reference Models (2025.findings-acl)
Copied to clipboard
Qi Liu, Jingqing Ruan, Hao Li, Haodong Zhao, Desheng Wang, Jiansong Chen, Wan Guanglu, Xunliang Cai, Zhi Zheng, Tong Xu
| Challenge: | Existing multi-objective preference alignment methods for large language models face limitations such as auxiliary reward/reference models and computational complexity. |
| Approach: | They propose a framework that achieves dynamic balance across preference dimensions by using dimension-aware generation metrics as implicit rewards. |
| Outcome: | Empirical results show that AMoPO outperforms state-of-the-art methods by 28.5% . |
MTR-Suite: A Framework for Evaluating and Synthesizing Conversational Retrieval Benchmarks (2026.acl-long)
Copied to clipboard
Junhao Ruan, Abudukeyumu Abudula, Bei Li, Yongjing Yin, Xinyu Liu, Kechen Jiao, Xin Chen, Jingang Wang, Xunliang Cai, Tong Xiao, JingBo Zhu
| Challenge: | Existing conversational retrieval benchmarks suffer from costly, sparse human annotation or rigid, unnatural automated heuristics. |
| Approach: | They propose a framework for auditing, synthesizing, and benchmarking conversational retrieval. |
| Outcome: | The proposed framework is based on three LLM-based auditors and a multi-agent system . it mimics production-style challenges (hard topic switching, verbosity) and offers superior discriminative power. |
Large Language Models Meet Open-World Intent Discovery and Recognition: An Evaluation of ChatGPT (2023.emnlp-main)
Copied to clipboard
Xiaoshuai Song, Keqing He, Pei Wang, Guanting Dong, Yutao Mou, Jingang Wang, Yunsen Xian, Xunliang Cai, Weiran Xu
| Challenge: | Existing methods to fine-tune discriminative models address these challenges by focusing on in-domain intents. |
| Approach: | They evaluate ChatGPT on OOD intent discovery and generalized intent discovery tasks . they outline the strengths and weaknesses of ChatGPt and outline their results . |
| Outcome: | The proposed task aims to extend a closed intent classifier to open-world intent sets. |
LLMs Know What They Need: Leveraging a Missing Information Guided Framework to Empower Retrieval-Augmented Generation (2025.coling-main)
Copied to clipboard
| Challenge: | Existing solutions to improve the accuracy of RAG are based on retrieval-augmented generation . however, RAG still faces several challenges in tackling complex multi-hop queries . |
| Approach: | They propose a Missing Information Guided Retrieve-Extraction-Solving paradigm that leverages the identification of missing information to generate a targeted query. |
| Outcome: | The proposed method can extract information from retrieved knowledge and know what is still missing. |
Domain-Lifelong Learning for Dialogue State Tracking via Knowledge Preservation Networks (2021.emnlp-main)
Copied to clipboard
Qingbin Liu, Pengfei Cao, Cao Liu, Jiansong Chen, Xunliang Cai, Fan Yang, Shizhu He, Kang Liu, Jun Zhao
| Challenge: | Existing offline DST models require a fixed dataset to train . Existing domain-lifelong learning methods are impractical in real-world applications . |
| Approach: | They propose a domain-lifelong learning method to continuously train a DST model on new data to learn incessantly emerging new domains while avoiding catastrophically forgetting old learned domains. |
| Outcome: | The proposed method outperforms state-of-the-art lifelong learning methods by 4.25% and 8.27% on the MultiWOZ and the SGD benchmarks. |
Revisit Self-Debugging with Self-Generated Tests for Code Generation (2025.acl-long)
Copied to clipboard
Xiancai Chen, Zhengwei Tao, Kechi Zhang, Changzhi Zhou, Xinyu Zhang, Wanli Gu, Yuanpeng He, Mengdi Zhang, Xunliang Cai, Haiyan Zhao, Zhi Jin
| Challenge: | Large language models (LLMs) have made significant advances in code generation, but they still face challenges when tackling complex programming tasks beyond their basic capabilities. |
| Approach: | They propose to integrate self-generated tests into the code generation process . they propose to use post-execution and in-exection self-debugging to mitigate test bias . |
| Outcome: | The proposed method improves the performance of large language models in code generation tasks by leveraging execution feedback from tests. |
BAPO: Boundary-Aware Policy Optimization for Reliable Agentic Search (2026.findings-acl)
Copied to clipboard
Shiyu Liu, Yongjing Yin, Jianhao Yan, Yunbo Tang, Qinggang Zhang, Bei Li, Xin Chen, Jingang Wang, Xunliang Cai, Jinsong Su
| Challenge: | Existing RL-based agentic search models fail to recognize reasoning boundaries and rarely admit "I DON'T KNOW" lack of reliability leads to plausible but unreliable answers, introducing significant risks . |
| Approach: | They propose a framework to cultivate reliable boundary awareness without compromising accuracy. |
| Outcome: | Experiments show that the proposed framework improves the reliability of agentic search models. |
Too Consistent to Detect: A Study of Self-Consistent Errors in LLMs (2025.emnlp-main)
Copied to clipboard
Hexiang Tan, Fei Sun, Sha Liu, Du Su, Qi Cao, Xin Chen, Jingang Wang, Xunliang Cai, Yuanzhuo Wang, Huawei Shen, Xueqi Cheng
| Challenge: | Existing detection methods fail to account for **self-consistent error** . study identifies self-consistency errors and evaluates them . |
| Approach: | They propose a method that fuses hidden state evidence from an external verifier LLM to detect self-consistent errors. |
| Outcome: | The proposed method significantly enhances performance on self-consistent errors across three LLM families. |
FIRE: Flexible Integration of Data Quality Ratings for Effective Pretraining (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to evaluate data quality rely on heuristic techniques or single quality signals. |
| Approach: | They propose a framework for integrating multiple data quality raters that integrates multiple quality signals into a unified space and provides a comprehensive quality signal for each data point. |
| Outcome: | The proposed framework outperforms existing methods and boosts model performance across a wide range of downstream tasks while requiring less than 37.5% tokens to reach the target performance. |
AJ-Bench: Benchmarking Agent-as-a-Judge for Environment-Aware Evaluation (2026.findings-acl)
Copied to clipboard
Wentao Shi, Yu Wang, Yuyang Zhao, Yuxin Chen, Fuli Feng, Xueyuan Hao, Xi Su, Qi GU, Hui Su, Xunliang Cai, Xiangnan He
| Challenge: | Existing approaches to verify agent behaviors in complex environments rely on rule-based verifiers or LLM-as-a-Judge models. |
| Approach: | They propose a benchmark to evaluate Agent-as-a-Judge across three domains . the benchmark covers search, data systems, and graphical user interfaces - with 155 tasks and 516 trajectories . |
| Outcome: | The proposed benchmark outperforms existing benchmarks in search, data systems, and GUI domains while revealing open challenges in agent-based verification. |
Beyond the Known: Investigating LLMs Performance on Out-of-Domain Intent Detection (2024.lrec-main)
Copied to clipboard
Pei Wang, Keqing He, Yejie Wang, Xiaoshuai Song, Yutao Mou, Jingang Wang, Yunsen Xian, Xunliang Cai, Weiran Xu
| Challenge: | Out-of-domain (OOD) intent detection is crucial for task-oriented dialogue systems. |
| Approach: | They conduct a comprehensive evaluation of large language models (LLMs) under various experimental settings and outline their strengths and weaknesses. |
| Outcome: | The proposed models exhibit strong zero-shot and few-shot capabilities, but is still at a disadvantage compared to models fine-tuned with full resource. |
When to Continue Thinking: Adaptive Thinking Mode Switching for Efficient Reasoning (2025.findings-emnlp)
Copied to clipboard
Xiaoyun Zhang, Jingqing Ruan, Xing Ma, Yawen Zhu, Haodong Zhao, Hao Li, Jiansong Chen, Ke Zeng, Xunliang Cai
| Challenge: | Large reasoning models (LRMs) incur excessive computational overhead due to redundant reasoning, especially on simple tasks. |
| Approach: | They propose an Adaptive Self-Recovery Reasoning framework that suppresses unnecessary reasoning and enables implicit recovery. |
| Outcome: | The proposed framework suppresses unnecessary reasoning and enables implicit recovery. |
Retrieval-based Knowledge Transfer: An Effective Approach for Extreme Large Language Model Compression (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Large-scale pre-trained language models (LLMs) have demonstrated exceptional performance in various natural language processing (NLP) tasks. |
| Approach: | They propose a new compression paradigm that extracts knowledge from pre-trained language models to construct a knowledge store from which the model can leverage it for effective inference. |
| Outcome: | The proposed model extracts knowledge from LLMs to construct a knowledge store, which the model can leverage for effective inference. |
APP: Adaptive Prototypical Pseudo-Labeling for Few-shot OOD Detection (2023.findings-emnlp)
Copied to clipboard
Pei Wang, Keqing He, Yutao Mou, Xiaoshuai Song, Yanan Wu, Jingang Wang, Yunsen Xian, Xunliang Cai, Weiran Xu
| Challenge: | Existing methods for detecting out-of-domain (OOD) intents are hard to label . previous studies use labeled in-domain data to learn intent representations . |
| Approach: | They propose a prototypical pseudo-labeling method for few-shot OOD detection . they propose 'protoOOD' framework and adaptive pseudo-labeled method . |
| Outcome: | The proposed method is able to detect out-of-domain (OOD) intents from user queries. |
Prejudge-Before-Think: Enhancing Large Language Models at Test-Time by Process Prejudge Reasoning (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Experimental results from competition-level complex reasoning demonstrate that bootstrapping with process prejudge can significantly enhance the reasoning ability of LLMs. |
| Approach: | They propose a new process prejudge strategy for LLM reasoning that bootstraps with process prejudgment . |
| Outcome: | The proposed method can be bootstrapped with process prejudge in LLM reasoning . it allows the model to anticipate errors rather than relying on trial and error. |
Rethinking the Reversal Curse of LLMs: a Prescription from Human Knowledge Reversal (2024.emnlp-main)
Copied to clipboard
Zhicong Lu, Li Jin, Peiguang Li, Yu Tian, Linhao Zhang, Sirui Wang, Guangluan Xu, Changyuan Tian, Xunliang Cai
| Challenge: | Existing methods for large language models (LLMs) are limited by their aggressive sample permutation and lack a detailed understanding of the underlying reasons for the reversal curse. |
| Approach: | They propose a method which enhances bidirectional entity correlation modeling and pairwise relationship reasoning to overcome the reversal curse. |
| Outcome: | The proposed method overcomes the reversal curse by augmenting the samples with entity order-reversals and semantically preserved question-answer pairs. |
How to Allocate, How to Learn? Dynamic Rollout Allocation and Advantage Modulation for Policy Optimization (2026.findings-acl)
Copied to clipboard
Yangyi Fang, Jiaye Lin, Xiaoliang Fu, Cong Qin, Haolin Shi, Chaowen Hu, Lu Pan, Ke Zeng, Xunliang Cai
| Challenge: | Existing methods for reinforcement learning with verifiable rewards are limited by the complexity of the problem and the complexity. |
| Approach: | They propose a theoretically-grounded dual-pronged optimization framework for reinforcement learning with verifiable rewards that compensates for gradient attenuation of high-confidence correct actions while utilizing entropy changes as computable indicators to stabilize excessive update magnitudes. |
| Outcome: | The proposed framework compensates for gradient attenuation of high-confidence correct actions while utilizing entropy changes as computable indicators to stabilize excessive update magnitudes. |
Optimizing Native Sparse Attention with Latent Attention and Local Global Alternating Strategies (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing research has proposed a variety of training-free and post-training methods for selecting critical key-value pairs at each generation step. |
| Approach: | They propose to use local (sliding-window) and global (compression/selective) attention across layers to enlarge long-context modeling. |
| Outcome: | Experiments on models from 340M to 1.3B parameters show that the proposed method matches or exceeds full attention and native sparse attention in both common-sense reasoning and long-context understanding tasks. |
Dynamic Fisher-weighted Model Merging via Bayesian Optimization (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing merging approaches involve scaling the parameters model-wise or integrating parameter importance parameter-wise. |
| Approach: | They propose a method for merging model-based models at the parameter level without training data or joint training. |
| Outcome: | The proposed model merging framework outperforms baseline models on validation sets. |
WISCA: A Lightweight Model Transition Method to Improve LLM Training via Weight Scaling (2026.findings-acl)
Copied to clipboard
Jiacheng Li, Jianchao Tan, Zhidong Yang, Pingwei Sun, Feiye Huo, Jiayu Qin, Xiangyu Zhang, Maoxin He, Guangming Tan, Weile Jia, Xunliang Cai, Tong Zhao
| Challenge: | Recent advances in training optimization for Transformer-based large language models lack systematic optimization of weight patterns during training. |
| Approach: | They propose a Weight Scaling method that rescales weights while preserving model outputs to improve model training efficiency and model quality. |
| Outcome: | The proposed method significantly improves convergence quality and loss reduction in LLMs with Grouped Query Attention architectures and LoRA fine-tuning tasks. |
Unlocking Implicit Experience: Synthesizing Tool-Use Trajectories from Text (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) can be used to effectively utilize tools in multi-turn interactions, but acquiring diverse and realistic multi-step tool-use data remains a challenge. |
| Approach: | They propose a text-based data synthesis pipeline that generates multi-turn tool-use trajectories from text corpora using relevance filtering, workflow tool extraction, trajectory grounding, and complexity refinement. |
| Outcome: | The proposed model achieves 14.9% improvement on the BFCL V3 Multi-turn benchmark while significantly reducing inference latency and costs. |
Speculative Decoding via Early-exiting for Faster LLM Inference with Thompson Sampling Control Mechanism (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to generate draft tokens in large language models are expensive and resource-intensive. |
| Approach: | They propose an approach to generate draft tokens using a segment of the LLM and a self-distillation method to enhance the quality of draft token. |
| Outcome: | The proposed approach generates draft tokens using a segment of the LLM and a self-distillation method to improve quality and speed up generation. |
MUSE: MCTS-Driven Red Teaming Framework for Enhanced Multi-Turn Dialogue Safety in Large Language Models (2025.emnlp-main)
Copied to clipboard
Siyu Yan, Long Zeng, Xuecheng Wu, Chengcheng Han, Kongcheng Zhang, Chong Peng, Xuezhi Cao, Xunliang Cai, Chenjuan Guo
| Challenge: | Existing defenses target single-turn attacks, but real-world usage involves multi-turn dialogues, exposing models to attacks that exploit conversational context to bypass safety measures. |
| Approach: | They propose a framework that tackles multi-turn jailbreaks from both attack and defense angles. |
| Outcome: | Experiments on large language models show that MUSE effectively mitigates multi-turn jailbreaks. |
Preference Curriculum: LLMs Should Always Be Pretrained on Their Preferred Data (2025.findings-acl)
Copied to clipboard
Xuemiao Zhang, Xu Liangyu, Feiyu Duan, Yongwei Zhou, Sirui Wang, Rongxiang Weng, Jingang Wang, Xunliang Cai
| Challenge: | Existing methods of uniformly sampling data throughout the pretraining process are suboptimal because they overlook the model's evolving data preferences. |
| Approach: | They propose a Perplexity Difference (PD) based Preference Curriculum learning framework which perceives and uses the data preferred by LLMs as their capabilities improve . they propose PDPC to complete the arrangement of the dataset offline and ensure continuous training without interruption. |
| Outcome: | The proposed framework surpasses baselines on 1.3B and 3B models and achieves an increased average accuracy of over 8.1% across MMLU and CMMLU. |
CoreCodeBench: Decoupling Code Intelligence via Fine-Grained Repository-Level Tasks (2026.acl-long)
Copied to clipboard
Lingyue Fu, Hao Guan, Bolun Zhang, Haowei Yuan, Yaoming Zhu, Lin Qiu, ZongYu Wang, Xuezhi Cao, Xunliang Cai, Weiwen Liu, Weinan Zhang, Yong Yu
| Challenge: | Existing large language models for software engineering rely on coarse-grained pass rates obscuring specific cognitive bottlenecks. |
| Approach: | They propose a repository-level benchmark that dissects coding capabilities through atomized tasks. |
| Outcome: | The proposed framework achieves a 78.55% validity yield, surpassing the 31.7% retention rate of SWE-bench-Verified. |
SampleMix: A Sample-wise Pre-training Data Mixing Strategy by Coordinating Data Quality and Diversity (2025.findings-emnlp)
Copied to clipboard
Xiangyu Xi, Deyang Kong, Jian Yang, Jiawei Yang, Zhengyu Chen, Wei Wang, Jingang Wang, Xunliang Cai, Shikun Zhang, Wei Ye
| Challenge: | Existing methods for pretraining data mixing for large language models neglect significant inter-domain overlaps and commonalities, failing to control the global diversity of the constructed training dataset. |
| Approach: | They propose a sample-wise data mixture approach that performs global cross-domain sampling by systematically evaluating the quality and diversity of each sample. |
| Outcome: | The proposed method exceeds existing domain-based methods in multiple downstream tasks and perplexity assessments. |
Harmonizing Dense and Sparse Signals in Multi-turn RL: Dual-Horizon Credit Assignment for Industrial Sales Agents (2026.acl-industry)
Copied to clipboard
| Challenge: | Large language models for industrial sales require balancing long-term commercial objectives with immediate linguistic constraints such as fluency and compliance. |
| Approach: | They propose a framework that disentangles optimization across time scales by normalizing advantages from turn-level and session-level rewards before fusion. |
| Outcome: | The proposed framework outperforms the state-of-the-art GRPO model in conversion rate and identity detection rate. |
LANG: Reinforcement Learning for Multilingual Reasoning with Language-Adaptive Hint Guidance (2026.acl-long)
Copied to clipboard
Yuchun Fan, Bei Li, Peiguang Li, Yilin Wang, Yongyu Mu, Jian Yang, Xin Chen, Rongxiang Weng, Jingang Wang, Xunliang Cai, JingBo Zhu, Tong Xiao
| Challenge: | Existing methods for enhancing multi-step reasoning have not fully translated to multilingual contexts. |
| Approach: | They propose a framework that leverages language-conditioned hints to guide exploration in non-English reasoning tasks. |
| Outcome: | Empirical results show that the proposed framework improves reasoning performance without compromising language consistency. |