CHÚC MỪNG NHÓM SINH VIÊN LỚP TMĐT2024.1 NGÀNH THƯƠNG MẠI ĐIỆN TỬ CÓ BÀI BÁO ĐƯỢC CHẤP NHẬN TẠI HỘI NGHỊ KHOA HỌC QUỐC TẾ ICEBA 2026
Thông tin chi tiết về hội nghị có thể tham khảo tại: ICEBA 2026 — 7th International Conference on Engineering, Physics, MEMS-Biosensors & Applications
Tên bài báo: “Beyond Blind Autonomy: A Provenance-Sensitive Delegation Framework for Large Language Model Trading Controllers”
Sinh viên thực hiện:
- Hà Nam Khang – MSSV: 24520733 – Lớp TMĐT2024.1
Lê Hiếu Huy – MSSV: 24520666 – Lớp TMĐT2024.1
Giảng viên hướng dẫn: TS. Liêu Nguyễn Ngọc Quế
Abstract: Evaluations of large language model (LLM) trading agents can wrongly attribute end-to-end performance to the model when legacy rules, imported state, or execution assumptions also determine outcomes. Such provenance failure can convert composite-system performance into unjustified automation authority. We introduce the Agentic-Control Delegation Gate (ACDG), a provenance-sensitive framework requiring five conjunctive claims: attributable agency, information separation, execution validity, simultaneous benefit-risk sufficiency, and operational governance. In an exploratory archival audit of eight AAPL- and GOLD-labeled asset-period cases using meta-llama/Llama-3.3-70B-Instruct traces, 338 of 744 closes (45.4%) are automatic legacy exits.
Because the retained table records entry timing but not direction, 744 entry sides are recovered from same-day executed entries; nine unsupported entries fail closed. Under common next-open execution, unit exposure, and turnover costs, 5-bp annualized mean returns are 11.45% for the LLM decision trace, 13.30% for the legacy-exit overlay, and 4.57% for the full hybrid, versus 14.24% for SVM, 13.03% for XGBoost, and 21.55% for buy-and-hold. The decision trace is positive versus Cash but fails the material-benefit and relative-risk gates; circular-shift placebos provide no consistent timing evidence. ACDG therefore supports explanation, exception triage, and audit assistance, but not irreversible execution control.

.










