Chen Baiyu  陈柏聿 陈柏聿  Chen Baiyu

Shuangyang Zheng, Yangchen Zhong, Xiaobao Que, Jingming Shen, Baiyu Chen, Weiqin Shen, Shibao Zheng, Bo Liu, Jin Gao

ICCSMT '25 · ACM · pp. 1747–1755 (9 pages) · Conf. dates:会议时间： 26–28 Dec 2025, Xiamen, China · Published:发布日期： 01 April 2026

Shuangyang Zheng, Yangchen Zhong, Xiaobao Que, Jingming Shen, Baiyu Chen, Weiqin Shen, Shibao Zheng, Hengxin Deng, Jin Gao

BDEIM '25 · ACM · pp. 1477–1487 (11 pages) · Conf. dates:会议时间： 19–21 Dec 2025, Shanghai, China · Published:发布日期： 15 May 2026

Defines hallucination evaluation requirements for generative AI models (LLMs, multimodal models, AI agents) across 6 core dimensions: factual accuracy · source fidelity · internal consistency · instruction following · logical reasoning · honest uncertainty expression. 规定生成式 AI 模型（大语言模型、多模态模型、AI 智能体）幻觉评估的 6 大核心维度：事实准确性·来源忠实性·内部一致性·指令遵循能力·逻辑推理能力·不确定性诚实表达。

Key contribution: First proposed the concept of Execution Hallucination (执行幻觉) — a new hallucination type specific to tool-calling AI agents, where the model fabricates execution states, falsely reports task completion, or claims unauthorised capabilities. This fills a critical gap in AI safety evaluation for agentic systems. 核心贡献：首次提出执行幻觉（Execution Hallucination）概念——专门针对具备工具调用能力的 AI 智能体，指模型虚构执行状态、虚报任务完成状态或声称具备未被授权执行能力的幻觉现象，填补了 AI 智能体安全评估领域的重要空白。

Project Researcher / Principal Researcher · Contributed to RAG-enhanced LLM, multi-agent systems, RL applied to financial regulation, futures anomaly monitoring, arbitrage strategy, and A-share manipulation detection research. 项目研究员 / 首席研究员 · 参与 RAG 增强 LLM、多智能体系统、强化学习在金融监管中的应用，以及期货异常监测、跨市场套利策略、A 股市场操纵识别等研究。

Independently led the end-to-end development of two core AI systems. Assessed as "Excellent" by on-site supervisor. 个人主导完成两大核心系统的全流程研发，获实习单位指导教师评定"实习成绩优秀"。

Independently refactored a semi-automatic order-execution strategy into a fully automated system with volatility filtering and dynamic long/short signal logic. Conducted multi-contract backtesting and live simulated trading with systematic post-trade analysis. Gained deep insight into the gap between live-trading constraints and backtest results, and into the practical logic of human–AI collaboration within a quantitative team. 独立将半自动下单策略改造为融合波动率筛选与动态多空信号逻辑的全自动系统；进行多品种回测与模拟实盘交易，建立系统化复盘机制。深入理解实盘约束与策略回测结果之间的差异，以及量化团队中人机协同的实践逻辑。

Designed and deployed a full-stack personalised stock-investment advisory Agent combining Computer Science × Finance × Psychology. Quantifies behavioural biases (overconfidence, loss aversion) as real-time emotional signals. Data sources: Sina Finance API · Eastmoney · Weibo sentiment · Baidu AI Search MCP · CSMAR database. Deployed on Baidu Qianfan AppBuilder as a production AI-native app. 设计并上线全栈个性化股票投资策略智能体，融合计算机科学 × 金融工程 × 心理学三学科视角，将过度自信、损失厌恶等行为偏差量化为实时情绪信号。数据源：新浪财经 API · 东方财富 · 微博舆情 · 百度 AI 搜索 MCP · 国泰安数据库。已在百度智能云千帆 AppBuilder 上线投入使用。

Constructed a comprehensive quantitative analysis framework on B2B software-sales data, covering sales funnel efficiency, opportunity prioritisation, customer lifetime value, and win-rate prediction. 基于 B2B 软件服务公司销售数据，构建涵盖销售漏斗效率、商机优先级评估、客户生命周期价值与赢单概率预测的综合定量分析框架。

Designed and implemented a competitive LLM hallucination detection & mitigation system through 7 major architectural iterations. Core innovation: replaced 5 domain experts with 12 task-type experts (translation · code generation · knowledge QA · reasoning · problem solving, etc.) plus a meta-expert router — enabling finer-grained, task-aware hallucination control. Dual API failover: DeepSeek (primary) + Qwen3-480B (backup). 经过 7 次主要架构迭代，设计并实现竞赛级 LLM 幻觉检测与缓解系统。核心创新：将 5 个领域专家替换为 12 个任务型专家（翻译·代码生成·知识问答·逻辑推理·问题求解等）+ 元专家路由器，实现更细粒度的任务感知幻觉控制。双 API 容错：DeepSeek（主）+ Qwen3-480B（备）。

Built a complete "prediction → feature fusion → optimisation → Agent" pipeline on ~100,000 road segment records. Stage 1 Prediction: TabNet (sequential attention, 99.99%) + custom DRGANDRNELM / LTCNBLS / TCAETCNResNet hybrid (99.85%) + DNN+SHAP. Stage 2 Feature Fusion: RLHF-based dual-weight fusion reconciles conflicting TabNet and SHAP importance rankings by simulating human preference — producing a more robust feature priority list. Stage 3 Optimisation: quantum annealing → HCAOA (budget-aware initialisation + adaptive threshold + multi-strategy local search). Stage 4 Agent: RAG knowledge base (road history data) + full-stack LLM chatbot for real-time maintenance advisory. 在约 10 万条路段记录上构建"预测 → 特征融合 → 优化 → 智能体"完整流水线。预测层：TabNet（顺序注意力机制，99.99%）+ 定制 DRGANDRNELM/LTCNBLS/TCAETCNResNet 混合架构（99.85%）+ DNN+SHAP。特征融合层：RLHF 双权重融合机制——模拟人类偏好，调和 TabNet 与 SHAP 特征重要性冲突，输出更鲁棒的特征优先级排名。优化层：量子退火 → HCAOA（预算感知初始化 + 自适应阈值 + 多策略局部搜索）。智能体层：RAG 路面历史知识库 + 全栈 LLM 实时维护顾问 Agent。

Constructed a multi-layer ensemble prediction framework on physical-examination biochemical data. Identified key urinary biomarkers (specific gravity, pH, osmolality, etc.) via modified Z-score outlier detection, Spearman rank correlation, and SHAP analysis. Compared four algorithms — automated deep forest, Bayesian network, Gaussian process classifier, and TabNet — with cross-validation optimisation. 基于体检尿液生化指标数据，构建多层次集成预测框架。通过修正 Z-Score 异常值处理、Spearman 秩相关与 SHAP 值分析，识别与肾结石高度相关的关键生理指标（尿比重、pH、渗透压等）。比较自动化深度森林、贝叶斯网络、高斯过程分类器与 TabNet 四种算法，交叉验证优化模型参数。

Designed and led implementation of a DeepSeek LLM-driven futures trading strategy. Market state is formatted as structured prompts integrating multi-dimensional technical indicators (RSI, MACD, Bollinger Bands + novel academic indicators GVR, GDMA, DoMA) across three timeframes (30 min / 1 h / 4 h); DeepSeek generates trade direction, confidence, target price, and stop-loss. Strict 1:2 risk-reward ratio with multi-layer risk controls (max position 25%, min holding time, frequency limits). 主导设计并实现 DeepSeek 大语言模型驱动的期货交易策略。将多维度技术指标（RSI、MACD、布林带 + 前沿学术指标 GVR、GDMA、DoMA）整合为结构化提示词，跨三个时间周期（30 分钟 / 1 小时 / 4 小时）验证后输入 DeepSeek，由模型生成含方向、置信度、目标价与止损价的决策信号。严格执行 1:2 风险收益比，配备多层次风控机制（最大仓位 25%、最小持仓时间、交易频率限制）。