Selected work

Representative Papers

Xmodel-2.5: 1.3B Data-Efficient Reasoning SLM

Nov 23, 2025

To address the challenges of deploying large language models (LLMs) on resource-constrained edge devices and cost-sensitive settings, this paper proposes an efficient small language model (SLM) with 1.3 billion parameters. The method integrates three key innovations: (1) micro-parameterization (μP) to enable hyperparameter transfer across model scales; (2) a three-phase training paradigm—Warmup–Stable–Decay—with a switch to the Muon optimizer during the Decay phase; and (3) architectural and systems optimizations, including tied word embeddings, FP8 mixed-precision training, and large-scale pretraining. Evaluated on 13 reasoning benchmarks, the model achieves an average improvement of +4.58% over baseline SLMs, while remaining deployable on edge hardware. All training code and checkpoints are publicly released.

0 citationsRead paper

MemOrb: A Plug-and-Play Verbal-Reinforcement Memory Layer for E-Commerce Customer Service

Sep 23, 2025

In e-commerce customer service scenarios, large language model (LLM) agents suffer from memory decay, repetitive errors, and a lack of mechanisms for continuous self-improvement. Method: This paper proposes a lightweight, plug-and-play memory layer centered on a shared, structured strategy-reflection memory bank. It enables cross-session experience reuse without model fine-tuning; reflection is distilled through multi-turn interaction, and historical strategies are dynamically retrieved via vector search to guide real-time decision-making. Contribution/Results: Experiments demonstrate a 63-percentage-point improvement in task success rate. The method significantly enhances decision stability and consistency across repeated trials, offering an efficient and practical solution for ensuring the long-term reliability of LLM agents in dynamic service environments.

0 citationsRead paper

Survey of Specialized Large Language Model

Aug 27, 2025

This work addresses the suboptimal performance and inefficient deployment of general-purpose large language models (LLMs) in specialized domains—including healthcare, finance, law, and technical engineering—by proposing a “domain-native architecture” paradigm that transcends conventional fine-tuning approaches. Methodologically, it integrates sparse computation, quantization-aware parameter-efficient adaptation, domain-specific fine-tuning, and multimodal fusion to enable deep structural-task alignment. Experiments demonstrate consistent superiority over both general-purpose LLMs and state-of-the-art fine-tuned variants across multiple domain-specific benchmarks, with notable gains in accuracy, robustness, and inference efficiency in professional services and e-commerce applications. The core contribution is a paradigm shift from *adapting general-purpose models* to *natively designing domain-optimized architectures*, establishing a systematic, efficient, and trustworthy technical pathway for deploying specialized large models in practice.

0 citationsRead paper

MindFlow+: A Self-Evolving Agent for E-Commerce Customer Service

Jul 24, 2025

Traditional intent recognition systems in e-commerce customer service exhibit weak contextual understanding and poor adaptability in multi-turn dynamic dialogues. Method: This paper proposes a continuously evolving dialogue agent framework that integrates large language models (LLMs), imitation learning, and offline reinforcement learning. It introduces tool-augmented demonstration construction, reward-conditioned data modeling, and an AI contribution quantification mechanism, combined with ReAct-style tool invocation and domain knowledge enhancement to enable task-driven response generation and policy optimization. Contribution/Results: Evaluated on real-world e-commerce dialogue data, the framework achieves significant improvements in context relevance (+12.3%), adaptability (+15.7%), and task accuracy (+18.1%), demonstrating its effectiveness and scalability in complex service scenarios.

0 citationsRead paper

ECom-Bench: Can LLM Agent Resolve Real-World E-commerce Customer Support Issues?

Jul 07, 2025

Existing large language model (LLM)-based agents lack systematic, multimodal evaluation frameworks tailored to real-world e-commerce customer service scenarios. Method: We introduce EC-Bench—the first multimodal agent benchmark specifically designed for e-commerce customer service—constructed from millions of real-world dialogues. It features a user-profile-driven dynamic simulation mechanism and a high-difficulty composite task suite encompassing cross-modal understanding, multi-turn reasoning, and real-time decision-making. Contribution/Results: EC-Bench significantly enhances evaluation authenticity and challenge. Experiments reveal that state-of-the-art multimodal models (e.g., GPT-4o) achieve only 10–20% pass@3 accuracy, exposing critical gaps in operational capability. This work shifts e-commerce AI agent evaluation from isolated skill assessment toward end-to-end problem-solving performance. The benchmark code and data will be publicly released.

0 citationsRead paper

Recent publications

Latest Papers

Xmodel-2.5: 1.3B Data-Efficient Reasoning SLM

Nov 23, 2025

To address the challenges of deploying large language models (LLMs) on resource-constrained edge devices and cost-sensitive settings, this paper proposes an efficient small language model (SLM) with 1.3 billion parameters. The method integrates three key innovations: (1) micro-parameterization (μP) to enable hyperparameter transfer across model scales; (2) a three-phase training paradigm—Warmup–Stable–Decay—with a switch to the Muon optimizer during the Decay phase; and (3) architectural and systems optimizations, including tied word embeddings, FP8 mixed-precision training, and large-scale pretraining. Evaluated on 13 reasoning benchmarks, the model achieves an average improvement of +4.58% over baseline SLMs, while remaining deployable on edge hardware. All training code and checkpoints are publicly released.

0 citationsRead paper

MemOrb: A Plug-and-Play Verbal-Reinforcement Memory Layer for E-Commerce Customer Service

Sep 23, 2025

In e-commerce customer service scenarios, large language model (LLM) agents suffer from memory decay, repetitive errors, and a lack of mechanisms for continuous self-improvement. Method: This paper proposes a lightweight, plug-and-play memory layer centered on a shared, structured strategy-reflection memory bank. It enables cross-session experience reuse without model fine-tuning; reflection is distilled through multi-turn interaction, and historical strategies are dynamically retrieved via vector search to guide real-time decision-making. Contribution/Results: Experiments demonstrate a 63-percentage-point improvement in task success rate. The method significantly enhances decision stability and consistency across repeated trials, offering an efficient and practical solution for ensuring the long-term reliability of LLM agents in dynamic service environments.

0 citationsRead paper

Survey of Specialized Large Language Model

Aug 27, 2025

This work addresses the suboptimal performance and inefficient deployment of general-purpose large language models (LLMs) in specialized domains—including healthcare, finance, law, and technical engineering—by proposing a “domain-native architecture” paradigm that transcends conventional fine-tuning approaches. Methodologically, it integrates sparse computation, quantization-aware parameter-efficient adaptation, domain-specific fine-tuning, and multimodal fusion to enable deep structural-task alignment. Experiments demonstrate consistent superiority over both general-purpose LLMs and state-of-the-art fine-tuned variants across multiple domain-specific benchmarks, with notable gains in accuracy, robustness, and inference efficiency in professional services and e-commerce applications. The core contribution is a paradigm shift from *adapting general-purpose models* to *natively designing domain-optimized architectures*, establishing a systematic, efficient, and trustworthy technical pathway for deploying specialized large models in practice.

0 citationsRead paper

MindFlow+: A Self-Evolving Agent for E-Commerce Customer Service

Jul 24, 2025

Traditional intent recognition systems in e-commerce customer service exhibit weak contextual understanding and poor adaptability in multi-turn dynamic dialogues. Method: This paper proposes a continuously evolving dialogue agent framework that integrates large language models (LLMs), imitation learning, and offline reinforcement learning. It introduces tool-augmented demonstration construction, reward-conditioned data modeling, and an AI contribution quantification mechanism, combined with ReAct-style tool invocation and domain knowledge enhancement to enable task-driven response generation and policy optimization. Contribution/Results: Evaluated on real-world e-commerce dialogue data, the framework achieves significant improvements in context relevance (+12.3%), adaptability (+15.7%), and task accuracy (+18.1%), demonstrating its effectiveness and scalability in complex service scenarios.

0 citationsRead paper

ECom-Bench: Can LLM Agent Resolve Real-World E-commerce Customer Support Issues?

Jul 07, 2025

Existing large language model (LLM)-based agents lack systematic, multimodal evaluation frameworks tailored to real-world e-commerce customer service scenarios. Method: We introduce EC-Bench—the first multimodal agent benchmark specifically designed for e-commerce customer service—constructed from millions of real-world dialogues. It features a user-profile-driven dynamic simulation mechanism and a high-difficulty composite task suite encompassing cross-modal understanding, multi-turn reasoning, and real-time decision-making. Contribution/Results: EC-Bench significantly enhances evaluation authenticity and challenge. Experiments reveal that state-of-the-art multimodal models (e.g., GPT-4o) achieve only 10–20% pass@3 accuracy, exposing critical gaps in operational capability. This work shifts e-commerce AI agent evaluation from isolated skill assessment toward end-to-end problem-solving performance. The benchmark code and data will be publicly released.

0 citationsRead paper