interactive model refinement

Traditional neural network training relies on fixed optimization pipelines, rendering it inflexible in dynamically addressing training instability and anomalies. To address this limitation, we propose the first interactive training framework enabling real-time human–AI collaborative intervention. Our method employs a lightweight control server that integrates expert human directives with AI agent feedback to dynamically adjust hyperparameters, data sampling strategies, and model checkpoints during training. This framework introduces, for the first time, a closed-loop interactive paradigm into neural network training, establishing a scalable human–machine collaboration interface coupled with automated response mechanisms. Experimental results demonstrate significant improvements in training stability, reduced sensitivity to initial hyperparameter configurations, and enhanced real-time responsiveness to user-specified customization requirements. The effectiveness is validated across multiple benchmark tasks.

Large language models (LLMs) suffer from static knowledge and cumbersome updates, leading to factual hallucinations; existing knowledge editing methods are largely confined to single-model settings, exhibiting limited generalizability and efficiency. This paper proposes OnceEdit—the first plug-and-play framework enabling parameterized knowledge editing across diverse LLMs. Its core innovations are: (1) a dynamic weight token mechanism that injects learnable, lightweight editing signals into model parameters; and (2) an integrated enhancement mechanism that jointly optimizes editing stability and cross-model transferability via model ensembling, selective parameter freezing, and gradient reweighting. Evaluated on multiple LLM benchmarks, OnceEdit achieves a 12.6% improvement in editing accuracy and a 3.2× speedup in inference latency. Crucially, it enables zero-shot transfer of edits to unseen models—marking the first demonstration of cross-architecture knowledge editing and substantially advancing beyond the constraints of single-model editing paradigms.

This work addresses the lack of a general, auditable dynamic control mechanism in existing training systems, which typically rely on framework-specific code. The authors propose the first cross-framework, open-source control plane that exposes training interfaces through a unified protocol, integrating declarative configuration, request validation, and secure control-point scheduling within the Aim workspace to enable metric monitoring, real-time intervention, and operational traceability. The system supports safe human and automated controller interventions during training while fully logging all operational trajectories. Experiments across five NLP and reinforcement learning tasks demonstrate its effectiveness, and the open-source implementation provides a foundation for reproducible human-in-the-loop training.

Existing LLM-driven automated visual modeling approaches rely on global, one-shot optimization, resulting in poor attribution, slow convergence, low stability, and limited accessibility for non-experts. Method: We propose an “iterative single-component fine-tuning” strategy, inspired by expert human practice, wherein only one module in the pipeline is optimized per iteration. This is integrated with training-feedback-guided modular updates, zero-shot prompt engineering, and a multi-domain evaluation protocol to construct an end-to-end LLM agent framework. Contribution/Results: Our approach significantly enhances interpretability, stability, and convergence efficiency of optimization. Evaluated across multiple standard benchmarks and Kaggle datasets, it consistently outperforms state-of-the-art zero-shot LLM methods, achieving superior classification accuracy and generalization capability.

Efficiently reusing multiple domain- or task-specific fine-tuned expert models while achieving high performance and strong generalization remains challenging. Method: We propose MoErging—a unified methodology for model merging, Mixture of Experts (MoE), and multi-task learning—featuring input-aware dynamic routing, parameter-space fusion, learnable router design, and collaborative multi-expert inference. Contribution/Results: We introduce the first taxonomy of MoErging methods; develop an open-source toolchain and standardized evaluation benchmark; and construct the first multidimensional MoErging knowledge graph. Our analysis rigorously characterizes applicability boundaries and performance trade-offs across paradigms, establishing a theoretical framework and practical guidelines for collaborative model reuse.

This study addresses the limitations of open-source large language models in automatically generating high-quality code review feedback, particularly in comparison to closed-source counterparts. Focusing on Java error feedback generation in programming education, it presents the first systematic comparison between parameter-efficient fine-tuning (PEFT) and prompt engineering. Leveraging the Code Llama model and a high-quality feedback dataset, the authors conduct a multidimensional evaluation using BLEU, ROUGE, BERTScore, and human assessment. Results demonstrate that PEFT substantially outperforms prompt engineering, with student evaluations indicating that the generated feedback achieves quality comparable to that of ChatGPT. These findings validate the feasibility of deploying open-source models at scale in educational programming contexts.

This study investigates how human-in-the-loop (HITL) feedback influences users’ perceptions of system accuracy and trust, highlighting the critical moderating role of task subjectivity. Through three controlled user experiments that systematically differentiate between objective and subjective task contexts, the research analyzes behavioral measures to assess the effects of feedback interaction. Findings reveal that in objective tasks, providing feedback significantly diminishes users’ trust in and perceived accuracy of the system, whereas this negative effect vanishes in subjective tasks. These results underscore task type as a pivotal factor shaping human–AI trust dynamics and offer important theoretical grounding and practical guidance for the design of HITL systems.

研究通过引入固定预算修订协议和确定性验证器解决大型语言模型中闭环修订失败问题，评估了精确反馈下的模型修正行为。

Existing single-prompt agents in 3D modeling lack planning and reflective capabilities, struggling to simultaneously achieve high geometric accuracy, aesthetic quality, and task completion rates. This work proposes a Planner-Actor-Critic multi-agent framework that introduces, for the first time, a structured self-reflection mechanism. By integrating real-time human guidance with the Blender MCP toolchain, the framework establishes a collaborative closed loop among planning, execution, and critique. This approach enables synchronous human-AI co-modeling and significantly outperforms baseline methods across diverse scenarios, demonstrating marked improvements in geometric precision, aesthetic quality, and task success rate, along with a substantial reduction in error frequency.

This study addresses the problem of "model collapse"—a degradation in performance arising when multiple models interactively learn from synthetic data generated by one another. By formalizing inter-model interactions as a directed graph, the work establishes, for the first time, necessary and sufficient conditions for model collapse in multi-model settings, thereby extending beyond prior analyses limited to single-model self-training. The theoretical framework integrates directed graph topology, finite-sample analysis of linear regression, and asymptotic theory of M-estimators to rigorously characterize the collapse mechanism. Extensive numerical experiments validate the theoretical findings and uncover an intrinsic relationship between the structure of the interaction graph and the extent of performance degradation across models.