arXiv subjects

Yupeng Qi

Publications and source records attributed to Yupeng Qi.

5 recordsLinked to original sources

arXiv2026

Please refuse to answer me! Mitigating Over-Refusal in Large Language Models via Adaptive Contrastive Decoding

Safety-aligned large language models (LLMs) often generate refusal responses to harmless queries due to the over-refusal problem. However, existing methods for mitigating over-refusal cannot maintain a low refusal ratio for harmless queries while keeping a high refusal ratio for malicious ones. In this paper, we analyze how system prompts with varying safety levels affect LLM refusal behaviors when facing over-refusal queries. A key observation is that, when LLMs suffer from the over-refusal issue, non-refusal tokens remain present in the next-token candidate list, but the model systematically fails to select them, despite the generation of refusal tokens. Based on this observation, we propose a training-free and model-agnostic approach, Adaptive Contrastive Decoding (AdaCD), to mitigate over-refusal while maintaining LLM safety. First, AdaCD compares the output distributions of the LLM with or without an extreme safety system prompt to refine the refusal token distribution. Second, we introduce an adaptive contrastive decoding strategy that dynamically incorporates or removes the refusal token distribution, adaptively boosting the probability of selecting refusal or non-refusal tokens. Experimental results on five benchmark datasets show that, on average, AdaCD reduces the refusal ratio for over-refusal queries by 10.35%, yet still increases the refusal ratio for malicious queries by 0.13%. Code is available at https://github.com/OutdoorManofML/AdaCD.

cs.CL↗

arXiv2026

RAIE: Region-Aware Incremental Preference Editing with LoRA for LLM-based Recommendation

Large language models (LLMs) are increasingly adopted as the backbone of recommender systems. However, user-item interactions in real-world scenarios are non-stationary, making preference drift over time inevitable. Existing model update strategies mainly rely on global fine-tuning or pointwise editing, but they face two fundamental challenges: (i) imbalanced update granularity, where global updates perturb behaviors unrelated to the target while pointwise edits fail to capture broader preference shifts; (ii) unstable incremental updates, where repeated edits interfere with prior adaptations, leading to catastrophic forgetting and inconsistent recommendations. To address these issues, we propose Region-Aware Incremental Editing (RAIE), a plug-in framework that freezes the backbone model and performs region-level updates. RAIE first constructs semantically coherent preference regions via spherical k-means in the representation space. It then assigns incoming sequences to regions via confidence-aware gating and performs three localized edit operations - Update, Expand, and Add - to dynamically revise the affected region. Each region is equipped with a dedicated Low-Rank Adaptation (LoRA) module, which is trained only on the region's updated data. During inference, RAIE routes each user sequence to its corresponding region and activates the region-specific adapter for prediction. Experiments on two benchmark datasets under a time-sliced protocol that segments data into Set-up (S), Finetune (F), and Test (T) show that RAIE significantly outperforms state-of-the-art baselines while effectively mitigating forgetting. These results demonstrate that region-aware editing offers an accurate and scalable mechanism for continual adaptation in dynamic recommendation scenarios. Our code is available at https://github.com/fengaogao/RAIE.

cs.IR↗

arXiv2025

Engineering.ai: A Platform for Teams of AI Engineers in Computational Design

In modern engineering practice, human engineers collaborate in specialized teams to design complex products, with each expert completing their respective tasks while communicating and exchanging results and data with one another. While this division of expertise is essential for managing multidisciplinary complexity, it demands substantial development time and cost. Recently, we introduced OpenFOAMGPT (1.0, 2.0), which functions as an autonomous AI engineer for computational fluid dynamics, and turbulence.ai, which can conduct end-to-end research in fluid mechanics draft publications and PhD theses. Building upon these foundations, we present Engineering.ai, a platform for teams of AI engineers in computational design. The framework employs a hierarchical multi-agent architecture where a Chief Engineer coordinates specialized agents consisting of Aerodynamics, Structural, Acoustic, and Optimization Engineers, each powered by LLM with domain-specific knowledge. Agent-agent collaboration is achieved through file-mediated communication for data provenance and reproducibility, while a comprehensive memory system maintains project context, execution history, and retrieval-augmented domain knowledge to ensure reliable decision-making across the workflow. The system integrates FreeCAD, Gmsh, OpenFOAM, CalculiX, and BPM acoustic analysis, enabling parallel multidisciplinary simulations while maintaining computational accuracy. The framework is validated through UAV wing optimization. This work demonstrates that agentic-AI-enabled AI engineers has the potential to perform complex engineering tasks autonomously. Remarkably, the automated workflow achieved a 100% success rate across over 400 parametric configurations, with zero mesh generation failures, solver convergence issues, or manual interventions required, validating that the framework is trustworthy.

cs.AI↗

arXiv2025

FeaGPT: an End-to-End agentic-AI for Finite Element Analysis

Large language models (LLMs) are establishing new paradigms for engineering applications by enabling natural language control of complex computational workflows. This paper introduces FeaGPT, the first framework to achieve complete geometry-mesh-simulation workflows through conversational interfaces. Unlike existing tools that automate individual FEA components, FeaGPT implements a fully integrated Geometry-Mesh-Simulation-Analysis (GMSA) pipeline that transforms engineering specifications into validated computational results without manual intervention. The system interprets engineering intent, automatically generates physics-aware adaptive meshes, configures complete FEA simulations with proper boundary condition inference, and performs multi-objective analysis through closed-loop iteration. Experimental validation confirms complete end-to-end automation capability. Industrial turbocharger cases (7-blade compressor and 12-blade turbine at \SI{110000}{rpm}) demonstrate the system successfully transforms natural language specifications into validated CalculiX simulations, producing physically realistic results for rotating machinery analysis. Additional validation through 432 NACA airfoil configurations confirms scalability for parametric design exploration. These results demonstrate that natural language interfaces can effectively democratize access to advanced computational engineering tools while preserving analytical rigor.

cs.SE↗

arXiv2025

MidPO: Dual Preference Optimization for Safety and Helpfulness in Large Language Models via a Mixture of Experts Framework

As large language models (LLMs) are increasingly applied across various domains, enhancing safety while maintaining the helpfulness of LLMs has become a critical challenge. Recent studies solve this problem through safety-constrained online preference optimization or safety-constrained offline preference optimization. However, the safety-constrained online methods often suffer from excessive safety, which might reduce helpfulness, while the safety-constrained offline methods perform poorly in adaptively balancing safety and helpfulness. To address these limitations, we propose MidPO, a \textbf{\underline{Mi}}xture of Experts (MoE) framework for safety-helpfulness \textbf{\underline{d}}ual \textbf{\underline{P}}reference \textbf{\underline{O}}ptimization. Firstly, MidPO devises single-preference enhanced direct preference optimization approach to transform the base model into two independent experts, termed safety and helpfulness experts, and fine-tunes the two independent experts for optimal safety or helpfulness performance. Secondly, to achieve an effective balance between safety and helpfulness, MidPO incorporates the two experts into the MoE framework and designs a dynamic routing mechanism to allocate contributions from each expert adaptively. We conduct quantitative and qualitative experiments on three popular datasets to demonstrate the proposed MidPO significantly outperforms state-of-the-art approaches in both safety and helpfulness. The code and models will be released.

cs.CL↗