Is this you? Claim this profile to correct it, add a bio and choose the work people see first.
Claim this profileAcademic lineage
Possible advisorsa guess from early papers, not confirmed
- Ke XuSuggested from co-authorship
- Ming ZhouSuggested from co-authorship
Is this you? Claim this profile to confirm or dismiss it.
Works111 from public data
- CommonGen: A Constrained Text Generation Challenge for Generative Commonsense Reasoning426
A constrained text generation task, CommonGen associated with a benchmark dataset, to explicitly test machines for the ability of generative commonsense reasoning, and demonstrates that the learned generative Commonsense reasoning capability can be transferred to improve downstream tasks such as CommonsenseQA by generating additional context.
- Memory in the Age of AI Agents288
This work aims to provide an up-to-date landscape of current agent memory research, clearly delineating the scope of agent memory and distinguishing it from related concepts such as LLM memory, retrieval augmented generation (RAG), and context engineering, and articulate a forward-looking perspective on emerging research frontiers.
- SuperGPQA: Scaling LLM Evaluation across 285 Graduate Disciplines266
This work presents SuperGPQA, a comprehensive benchmark that evaluates graduate-level knowledge and reasoning capabilities across 285 disciplines and management of a large-scale annotation process, involving over 80 expert annotators and an interactive Human-LLM collaborative system, offering valuable methodological guidance for future research initiatives of comparable scope.
- OpenMoE: An Early Effort on Open Mixture-of-Experts Language Models232
This investigation confirms that MoE-based LLMs can offer a more favorable cost-effectiveness trade-off than dense LLMs, highlighting the potential effectiveness for future LLM development and proposes potential strategies for mitigating the issues found and further improving off-the-shelf MoE LLM designs.
- BERT-of-Theseus: Compressing BERT by Progressive Module Replacing230
This paper proposes a novel model compression approach to effectively compress BERT by progressive module replacing, which outperforms existing knowledge distillation approaches on GLUE benchmark, showing a new perspective of model compression.
- Agents: An Open-source Framework for Autonomous Language Agents151
Agents is an open-source library that enables non-specialists to build, customize, test, tune, and deploy state-of-the-art autonomous language agents without much coding.
- To Repeat or Not To Repeat: Insights from Scaling LLM under Token-Crisis150
This study empirically investigates the consequences of repeating pre-training data, revealing that the model is susceptible to overfitting, leading to multi-epoch degradation, and discovers that leveraging mixture-of-experts (MoE) enables cost-effective and efficient hyper-parameter tuning for computationally intensive dense LLMs with comparable trainable parameters.
- A Comprehensive Survey on Long Context Language Modeling144
A comprehensive survey on recent advances in long-context modeling for large language models, including evaluation paradigms for long-context comprehension and long-form generation, as well as behavioral analysis and mechanism interpretability of LCLMs.
- Controlled Text Generation with Natural Language Instructions139
This work presents InstructCTG, a controlled text generation framework that incorporates different constraints by conditioning on natural language descriptions and demonstrations of the constraints, and is more flexible to different constraint types than existing search-based or score-based methods.
- BERT Learns to Teach: Knowledge Distillation with Meta Learning122
Experiments show that MetaDistil can yield significant improvements compared with traditional KD algorithms and is less sensitive to the choice of different student capacity and hyperparameters, facilitating the use of KD on different tasks and models.
- AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems?119
This work proposes AgenTracer, the first automated framework for annotating failed multi-agent trajectories via counterfactual replay and programmed fault injection, and develops AgenTracer-8B, a lightweight failure tracer trained with multi-granular reinforcement learning, capable of efficiently diagnosing errors in verbose multi-agent interactions.
- How Many Unicorns Are in This Image? A Safety Evaluation Benchmark for Vision LLMs116
This work introduces a comprehensive safety evaluation suite, covering both out-of-distribution (OOD) generalization and adversarial robustness, and proposes a straightforward attack strategy for misleading VLLMs to produce visual-unrelated responses.
- 115
- BERT-based Lexical Substitution114
This work proposes an end-to-end BERT-based lexical substitution approach which can propose and validate substitute candidates without using any annotated data or manually curated resources and achieves the state-of-the-art results in both LS07 and LS14 benchmarks.
- RecurrentGPT: Interactive Generation of (Arbitrarily) Long Text107
RecurrentGPT is a language-based simulacrum of the recurrence mechanism in RNNs that uses natural language to simulate the Long Short-Term Memory mechanism in an LSTM and demonstrates the utility of borrowing ideas from popular model designs in cognitive science and deep learning for prompting LLMs.
- AutoAct: Automatic Agent Learning from Scratch for QA via Self-Planning102
AutoAct is introduced, an automatic agent learning framework for QA that does not rely on large-scale annotated data and synthetic planning trajectories from closed-source models and leverages a division-of-labor strategy to automatically differentiate based on the target task information and synthesized trajectories, producing a sub-agent group to complete the task.
- Risks of AI scientists: prioritizing safeguarding over autonomy98
The authors examine the vulnerabilities of AI scientists, propose a risk taxonomy based on user intent and impact domains, and develop a triadic safeguarding framework emphasizing human regulation, agent alignment, and environmental feedback understanding.
- Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving92
AGENT KB is introduced, a universal memory infrastructure enabling seamless experience sharing across heterogeneous agent frameworks without retraining and establishes the foundation for collective agent intelligence through shared memory infrastructures.
- MAP-Neo: Highly Capable and Transparent Bilingual Large Language Model Series89
The MAP-Neo is the first fully open-sourced bilingual LLM with comparable performance compared to existing state-of-the-art LLMs, and is hoped to enhance and strengthen the open research community and inspire more innovations and creativities to facilitate the further improvements of LLMs.
- A Comparative Study on Reasoning Patterns of OpenAI's o1 Model88
This work compares o1 with existing Test-time Compute methods by using OpenAI's GPT-4o as a backbone on general reasoning benchmarks in three domains, and shows that the o1 model has achieved the best performance on most datasets.
- MemEvolve: Meta-Evolution of Agent Memory Systems85
This work proposes MemEvolve, a meta-evolutionary framework that jointly evolves agents'experiential knowledge and their memory architecture, allowing agent systems not only to accumulate experience but also to progressively refine how they learn from it.
- AutoKaggle: A Multi-Agent Framework for Autonomous Data Science Competitions85
This work proposes AutoKaggle, a powerful and user-centric framework that assists data scientists in completing daily data pipelines through a collaborative multi-agent system that implements an iterative development process that combines code execution, debugging, and comprehensive unit testing to ensure code correctness and logic consistency.
- X2-VLM: All-in-One Pre-Trained Model for Vision-Language Tasks75
X-VLM is presented, an all-in-one model with a flexible modular architecture, in which it is shown that the modular design of X-VLM outperforms state-of-the-art multilingual multi-modal pre-trained models without any multilingual pre-training.
- Scaling Test-time Compute for LLM Agents61
This is the first systematic exploration of applying test-time scaling methods to language agents and investigates the extent to which it improves their effectiveness, and finds that among different verification and result merging approaches, the list-wise method performs best.
- 55
- A Survey on Latent Reasoning53
This survey provides a comprehensive overview of the emerging field of latent reasoning, examining the foundational role of neural network layers as the computational substrate for reasoning and exploring diverse latent reasoning methodologies, including activation-based recurrence, hidden state propagation, and fine-tuning strategies that compress or internalize explicit reasoning traces.
- M+: Extending MemoryLLM with Scalable Long-Term Memory51
M+ is introduced, a memory-augmented model based on MemoryLLM that significantly enhances long-term information retention and integrates a long-term memory mechanism with a co-trained retriever, dynamically retrieving relevant information during text generation.
- Interactive Natural Language Processing51
This paper provides a systematic classification of iNLP, dissecting its various components, including interactive objects, interaction interfaces, and interaction methods, and delves into the evaluation methodologies used in the field, exploring its diverse applications, and scrutinize its ethical and safety issues.
- Learning to Compare for Better Training and Evaluation of Open Domain Natural Language Generation Models51
Experimental results show that the proposed model correlates better with human preference compared with previous automated evaluation approaches, and is proposed to apply as a performance indicator during training for better hyperparameter tuning and early-stopping.
- TaskCraft: Automated Generation of Agentic Tasks50
An automated workflow for generating difficulty-scalable, multi-tool, and verifiable agentic tasks with execution trajectories, and empirical results show that these tasks improve prompt optimization in the generation workflow and enhance supervised fine-tuning of agentic foundation models.
- Beyond Preserved Accuracy: Evaluating Loyalty and Robustness of BERT Compression49
Two new metrics, label loyalty and probability loyalty, are proposed that measure how closely a compressed model mimics the original model and the effect of compression with regard to robustness under adversarial attacks is explored.
- HelloBench: Evaluating Long Text Generation Capabilities of Large Language Models48
The Hierarchical Long Text Generation Benchmark (HelloBench), a comprehensive, in-the-wild, and open-ended benchmark to evaluate LLMs' performance in generating long text and the proposed HelloEval, a human-aligned evaluation method that significantly reduces the time and effort required for human evaluation while maintaining a high correlation with human evaluation.
- Scheduled DropHead: A Regularization Method for Transformer Models47
This work introduces DropHead, a structured dropout method specifically designed for regularizing the multi-head attention mechanism which is a key component of transformer, and proposes a novel dropout rate scheduler to adjust the drop out rate of DropHead throughout training, which results in a better regularization effect.
- 46
- Improving Grammatical Error Correction with Machine Translation Pairs44
A novel data synthesis method to generate diverse error-corrected sentence pairs for improving grammatical error correction based on a pair of machine translation models of different qualities, which can effectively help a GEC model to improve the performance and achieve the state-of-the-art single-model performance in BEA-19 and CoNLL-14 benchmark datasets.
- 44
- Struc-Bench: Are Large Language Models Good at Generating Complex Structured Tabular Data?41
This study assesses LLMs’ proficiency in structuring tables and introduces a novel fine-tuning method, cognizant of data structures, to bolster their performance, and unveils Struc-Bench, a comprehensive benchmark featuring prominent LLMs, which spans text tables, HTML, and LaTeX formats.
- Findings of the WMT 2023 Shared Task on Machine Translation with Terminologies41
The WMT 2023 Terminology Shared Task investigates progress in machine translation of texts with specialized vocabulary and evaluates 21 submissions from 7 teams on two main criteria: general translation quality and the effectiveness of translating specialized terminology.
- Reverse-Engineered Reasoning for Open-Ended Generation38
This work introduces REverse-Engineered Reasoning (REER), a new paradigm that fundamentally shifts the approach of deep reasoning, and works ``backwards'' from known-good solutions to computationally discover the latent, step-by-step deep reasoning process that could have produced them.
- AI PERSONA: Towards Life-long Personalization of LLMs38
This work provides a clear task formulation and introduces a simple, general, effective, and scalable framework for life-long personalization of LLM systems and language agents and introduces methods to synthesize realistic benchmarks and robust evaluation metrics.
- VideoScore2: Think before You Score in Generative Video Evaluation36
VideoScore2 is presented, a multi-dimensional, interpretable, and human-aligned framework that explicitly evaluates visual quality, text-to-video alignment, and physical/common-sense consistency while producing detailed chain-of-thought rationales.
- 34
- 34
- Struc-Bench: Are Large Language Models Really Good at Generating Complex Structured Data?33
This study assesses the capability of Current LLMs in generating complex structured data and proposes a structure-aware fine-tuning approach as a solution to improve this ability, and presents an ability map of model capabilities from six dimensions.
- Connecting the Dots Between Fact Verification and Fake News Detection33
The proposed approach makes use of the recent success of fact verification models and enables zero-shot fake news detection, alleviating the need of large scale training data to trainfake news detection models.
- O-Mem: Omni Memory System for Personalized, Long Horizon, Self-Evolving Agents32
The initial design of O-Mem is proposed, a novel memory framework based on active user profiling that dynamically extracts and updates user characteristics and event records from their proactive interactions with agents, enabling more adaptive and coherent personalized responses.
- Weaver: Foundation Models for Creative Writing29
This work introduces Weaver, the first family of large language models (LLMs) dedicated to content creation that natively supports retrieval-augmented generation (RAG) and function calling (tool usage) and presents various use cases of these abilities for improving AI-assisted writing systems.
- Efficient Prompting via Dynamic In-Context Learning29
DynaICL is proposed, a recipe for efficient prompting with black-box generalist models that dynamically allocate in-context examples according to the input complexity and the computational budget and it is found that a meta controller trained on a certain backbone model and tasks can successfully generalize to unseen models and tasks.
- 27
- 21
- OThink-R1: Intrinsic Fast/Slow Thinking Mode Switching for Over-Reasoning Mitigation19
OThink-R1 is proposed, a hybrid reasoning framework that integrates both modes within a single LRM and enables automatic mode switching based on problem characteristics and shows that OThink-R1 reduces reasoning token usage significantly while maintaining competitive accuracy.
- Learning from Perturbations: Diverse and Informative Dialogue Generation with Inverse Adversarial Training19
Experimental results on two benchmark datasets show that the proposed Inverse Adversarial Training (IAT) algorithm can better model dialogue history and generate more diverse and consistent responses.
- 18
- Blow the Dog Whistle: A Chinese Dataset for Cant Understanding with Common Sense and World Knowledge18
Experiments suggest that a task for cant understanding requires deep language understanding, common sense, and world knowledge and thus can be a good testbed for pretrained language models and help models perform better on other tasks.
- 17
- Improving Sequence-to-Sequence Pre-training via Sequence Span Rewriting17
This paper conducts extensive experiments by using SSR to improve the typical Seq2Seq pre-trained model T5 in a continual pre-training setting and show substantial improvements over T5 on various natural language generation tasks.
- From Code Foundation Models to Agents and Applications: A Comprehensive Survey and Practical Guide to Code Intelligence15
This work provides a comprehensive synthesis and practical guide about code LLMs, systematically examining the complete model life cycle from data curation to post-training through advanced prompting paradigms, code pre-training, supervised fine-tuning, reinforcement learning, and autonomous coding agents.
- PersonaFeedback: A Large-scale Human-annotated Benchmark For Personalization15
PersonaFeedback is introduced, a new benchmark that directly evaluates LLMs' ability to provide personalized responses given pre-defined user personas and queries and demonstrates that the current retrieval-augmented framework should not be seen as a de facto solution for personalization tasks.
- 15
- Learning to Predict Persona Information for Dialogue Personalization without Explicit Persona Description15
A novel approach that learns to predict persona information based on the dialogue history to personalize the dialogue agent without relying on any explicit persona descriptions during inference is proposed and found that a trained persona prediction model can be successfully transferred to other datasets and help generate more relevant responses.
- 15
- VLUE: A Multi-Task Benchmark for Evaluating Vision-Language Models14
The VLUE benchmark is introduced, a multi-task multi-dimension benchmark for evaluating the generalization capabilities and the efficiency-performance trade-off (``Pareto SOTA'') of VLP models, and it is demonstrated that there is a sizable generalization gap for all V LP models when testing on out-of-distribution test sets annotated on images from a more diverse distribution that spreads across cultures.
- Towards Personalized Deep Research: Benchmarks and Evaluations13
This work introduces Personalized Deep Research Bench (PDR-Bench), the first benchmark for evaluating personalization in DRAs, and proposes the PQR Evaluation Framework, which jointly measures Personalization Alignment, Content Quality, and Factual Reliability.
- How Far Are We from Genuinely Useful Deep Research Agents?13
FINDER is presented, an enhanced benchmark consisting of 100 human-curated research tasks with 419 structured checklist items that standardize report structure, analytical depth, and factual grounding and proposes Deep rEsearch Failure Taxonomy (DEFT), the first failure taxonomy for deep research agents.
- EdgeBench: Unveiling Scaling Laws of Learning from Real-World Environments12
This work finds the first evidence that overall performance during environment learning follows a log-sigmoid scaling law with remarkably high precision, reaching R^2 = 0.998, and finds that agent learning speed roughly doubles every three months.
- Towards LifeSpan Cognitive Systems12
This paper proposes a potential instantiation for LSCS that can integrate all four classes of technologies based on a conceptual metric called Storage Complexity, which measures the relative space required to store past experiences.
- 12
- CodeCriticBench: A Holistic Code Critique Benchmark for Large Language Models11
A holistic code critique benchmark for LLMs called CodeCriticBench is introduced, which includes two mainstream code tasks (i.e., code generation and code QA) with different difficulties and the evaluation protocols include basic critique evaluation and advanced critique evaluation for different characteristics.
- 11
- 11
- O-Researcher: An Open Ended Deep Research Model via Multi-Agent Distillation and Agentic RL9
This work introduces a novel framework for the automated synthesis of sophisticated, research-grade instructional data based on a multi-agent workflow where collaborative AI agents simulate complex tool-integrated reasoning to generate diverse and high-fidelity data end-to-end.
- MiCoTA: Bridging the Learnability Gap with Intermediate CoT and Teacher Assistants9
This work introduces MiCoTAl, a framework for improving long CoT distillation for SLMs that employs intermediate-sized models as teacher assistants and utilizes intermediate-length CoT sequences to bridge both the capacity and reasoning length gaps.
- 9
- ScaleLong: A Multi-Timescale Benchmark for Long Video Understanding7
ScaleLong offers a fine-grained, multi-timescale benchmark for advancing MLLM capabilities in long-video understanding by embedding questions targeting four hierarchical timescales -- clip (seconds), shot (tens of seconds), event (minutes), and story (hours), all within the same video content.
- 6
- Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization5
This work proposes \emph{Search More, Think Less} (SMTL), a framework for long-horizon agentic search that targets both efficiency and generalization, and introduces a unified data synthesis pipeline that constructs search tasks spanning both deterministic question answering and open-ended research scenarios with task appropriate evaluation metrics.
- SupReMix: Supervised contrastive learning for medical imaging regression with mixup5
This paper argues that the potential of contrastive learning for medical image regression has been overshadowed due to the neglect of two crucial aspects: ordinality-awareness and hardness, and proposes Supervised Contrastive Learning for Medical Imaging Regression with Mixup (SupReMix).
- MIMIR: A Customizable Agent Tuning Platform for Enhanced Scientific Applications5
MIMIR is a streamlined platform that leverages large LLMs to generate agent-tuning data for fine-tuning smaller, specialized models, and enables larger models to simulate various roles and create interaction data, which can be used to fine-tune open-source models like LLaMA2.
- Pseudo-Bidirectional Decoding for Local Sequence Transduction5
PBD provides right-side context information for the decoder, reduces the number of parameters by half, and provides good regularization effects and Experimental results on several benchmark datasets show that this approach consistently improves the performance of standard seq2seq models on LST tasks.
- DirectorBench: Diagnosing Long-Form Video Generation with Personalized Multi-Agent Evaluation4
The results show that DirectorBench captures human-perceptible quality differences and reveals workflow- and profile-dependent failure modes that are hidden by aggregate scoring, highlighting the importance of diagnostic and profile-aware benchmarking for long-form video generation.
- Workflow-GYM: Towards Long-Horizon Evaluation of Computer-use Agentic tasks in Real-World Professional Fields4
Workflow-GYM is introduced, a benchmark for long-horizon GUI tasks centered on professional domains and specialized software environments and provides important insights into the limitations of current agent systems and suggest key directions for the next generation of GUI-agent research.
- 4
- Towards Faithful and Controllable Personalization via Critique-Post-Edit Reinforcement Learning3
Critique-Post-Edit is proposed, a robust reinforcement learning framework that enables more faithful and controllable personalization in large language models and substantially outperforms standard PPO on personalization benchmarks.
- 3
- A$^2$FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning3
This work presents Adaptive Agent Foundation Model (A$^2$FM), a unified framework that follows a route-then-align principle: the model first learns task-aware routing and then aligns mode-specific trajectories under a shared backbone, performing competitively with frontier LLMs across agentic, reasoning, and general benchmarks.
- 3
- 3
- 3
- 3
- EcoGym: Evaluating LLMs for Long-Horizon Plan-and-Execute in Interactive Economies2
EcoGym is introduced, a generalizable benchmark for continuous plan-and-execute decision making in interactive economies and released as an open, extensible testbed for transparent long-horizon agent evaluation and for studying controllability utility trade-offs in economic settings.
- Training-free Periodic Interest Augmentation in Incremental Recommendation2
TPIA, a Training-free approach for Periodic Interest Augmentation in incremental recommendation is proposed, which first calculates the importance score of each model in the previous period and develops a method for trimming redundant and abnormal parameters during model merging.
- 2
- MIMIR: A Streamlined Platform for Personalized Agent Tuning in Domain Expertise2
A streamlined platform offering a customizable pipeline that enables users to leverage both private knowledge and publicly available, legally compliant datasets at scale for personalized agent tuning, and dual capability ensures that language agents developed through the platform possess both specific agent abilities and general competencies.
- 2
- Doolittle: Benchmarks and Corpora for Academic Writing Formalization2
This work forms this language refinement task as a formal text style transfer task which transfers informal-academic text to formal-academic and contributes a large-scale non-parallel dataset for this purpose.
- OPD-Evolver: Cultivating Holistic Agent Evolver via On-Policy Distillation1
OPD-Evolver internalizes high-value experience and memory management, enabling OPD-Evolver-9B to challenge giant counterparts such as Qwen3.5-397B-A17B and Step-3.5-Flash, pointing beyond memory-augmented agents toward genuinely qualified agent evolvers.
- HiNS: Hierarchical Negative Sampling for More Comprehensive Memory Retrieval Embedding Model1
This work proposes a principled data construction framework HiNS that explicitly models negative sample difficulty tiers and incorporates empirically grounded negative ratios derived from conversational data, enabling the training of embedding models with substantially improved retrieval fidelity and generalization in memory-intensive tasks.
- Reliable and Responsible Foundation Models: A Comprehensive Survey1
Critical issues, including bias and fairness, security and privacy, uncertainty, explainability, and distribution shift are explored, as well as methods like alignment and Artificial Intelligence-Generated Content (AIGC) detection.
- 1
- 1
- 1
- SmartTrim: Adaptive Tokens and Attention Pruning for Efficient Vision-Language Models1
This work proposes an adaptive acceleration method S MART T RIM for VLMs, which adjusts the inference overhead based on the complexity of instances, and incorporates lightweight trimming modules into the backbone to perform task-specific pruning on redundant inputs and parameters, without the need for additional pre-training or data augmentation.
- 1
- –
- OThink-SRR1: Search, Refine and Reasoning with Reinforced Learning for Large Language Models–
This work proposes OThink-SRR1, a framework that enhances large models with an iterative Search-Refine-Reason process trained via reinforcement learning, and introduces GRPO-IR, an end-to-end reinforcement learning algorithm that rewards accurate evidence identification while penalizing excessive retrievals, thus training the model to be both focused and efficient.
- –
- COIG-P: A High-Quality and Large-Scale Chinese Preference Dataset for Alignment with Human Values–
Evaluation results show that that COIG-P significantly outperforms other Chinese preference datasets, and it brings significant performance improvements ranging from 2% to 12% for the Qwen2/2.5 and Infinity-Instruct-3M-0625 model series, respectively.
- –
- –
- –
- –
Publication data from OpenAlex, with missing venues and authors filled in from Crossref; citation counts are the higher of OpenAlex and Semantic Scholar, last synced 2026-10-12. One-sentence summaries under some papers are written by Semantic Scholar’s model. Citation counts may be lower than on Google Scholar, which indexes more sources.
Report an error
Wrong papers, two people merged into one, or a profile that should not be here? Tell us and we will fix or hide it.