Yufeng Liang

B.Sc. Graduate · 2026

Bio

I recently graduated from Shandong University with a B.Sc. in Statistics (Data Science & AI) and am currently seeking PhD opportunities in AI. My previous work spans both research prototypes and real-world AI systems, including long-horizon agents, evidence-grounded RAG, and knowledge-grounded production planning. Through these projects, I have worked across the full process from research questions and system design to implementation and evaluation. I am particularly interested in building AI systems that can learn and adapt reliably from long-term interaction and experience. My future research will focus on long-horizon and self-evolving AI agents, especially memory, continual learning, and personalized adaptation.

Research Interests
Agent Memory Long-Running Agents Self-Evolving Agents Retrieval-Augmented Generation Knowledge Graphs Knowledge-Grounded AI Reliable AI Systems AI for Financial Compliance
Reliable Adaptationlong-running agents
Statetrue now
Memorypersist
Actionchange next
Experiencelearn

Primary Research Project

State, Memory, and Reliable Adaptation in Long-Horizon Agents

Research & Industry Experience

Mar. 2026 — Aug. 2026 · Independent Researcher

RegSuitQA: Evidence-Grounded RAG for Chinese Financial Suitability Compliance

Background

Financial suitability decisions are scenario-based and often require several complementary provisions, yet existing benchmarks rarely test whether a RAG system can recover the complete evidence set needed for a defensible conclusion.

Method

Built RegSuitQA from six CSRC / AMAC regulations using Law → Article → Atomic Segment evidence units, then constructed scenario, regular, no-evidence, and malicious cases with minimal sufficient gold evidence.

Finding

Found that multi-provision evidence recovery is the main bottleneck: Hybrid + Rerank is the strongest baseline, while coarse routing and naive decomposition do not consistently improve complete evidence acquisition.

240 QA 453 Atomic Evidence Units Long Paper · Under Review at NLLP 2026
6 Regulatory Sources CSRC + AMAC Legal-Structure-Aware Segmentation Law → Article → Atomic Segment 453 Atomic Evidence Units Answerable QA · 192 84 regular · 108 scenario 58 single · 38 double · 96 multi-provision Boundary Cases · 48 24 no-evidence · 24 malicious abstain · safety refuse
Question Malicious-Request Gate Hybrid Retrieval Rerank Top-10 Evidence → Answer / Abstain Routing Variants Soft · Hard Decomposition Retrieval Variants BM25 · Dense
Jun. 2026 — Present · Independent Researcher & Developer

Knowledge-Grounded Agent for Garment Production Planning

Problem

Historical garment production records are highly heterogeneous: feature descriptions are free-form, process codes and names are inconsistent, and fine-grained feature–process mappings are often missing.

Method

I developed a feature-anchored knowledge construction pipeline that grounds noisy descriptions into canonical feature entities, locates one or more process anchors in full historical sequences, traces semantically related operations under object constraints, and merges them into feature-conditioned local process packages. These packages are then normalized across historical instances through code–name consolidation and semantic-equivalence matching to form reusable canonical process packages.

System

These canonical feature–process mappings form the core knowledge graph for downstream planning. Around this layer, I built a complete system combining confidence-based routing with a trained language-model backbone for feature decomposition, normalization, and grounding, enabling the system to inherit, adapt, or reconstruct production processes for new garment designs.

LOCAL PACKAGE EXTRACTIONCROSS-INSTANCE CONSOLIDATIONREUSABLE KNOWLEDGEHistoricalFeaturefree-form descriptionnoisy / non-canonicalFeatureGroundingpart / objectoperationcontextcanonical feature entityFull Historical Process Sequencerelated operations may be separated in the full sequenceAnchor-based Tracing1. Locate candidate anchor(s)term · object · semantic cues2. Trace related operationsobject constraint +process semanticsnon-contiguous operations allowedLocal ProcessPackagefeature-conditioned operationsfrom one historical itemAcross Historical ItemsPackage APackage BPackage CProcessNormalizationcode name consolidationsemantic disambiguationPackage Consolidationsemantic equivalence+ historical supportselect representative packageCanonical Featurenormalized feature entityCanonical Process Packagereusable local process knowledgeFeature–ProcessKnowledge Graphfeature package
OFFLINE KNOWLEDGECONSTRUCTIONONLINE SEMANTICUNDERSTANDINGCONFIDENCE-BASED ROUTINGPRODUCTION PLANNING ENGINEOUTPUT & HUMAN REVIEWHistorical ProductionRecordsFeature PackageBuilderextract · normalize · consolidateNew Garment DesignTrained ModelBackbonefeature decomposition · normalizationsemantic grounding / matchingKNOWLEDGE CORECanonical Feature Entitiespart · operation · contextCanonical Process Packagesreusable feature-conditioned proceduresFeature–Process Knowledge Graphrelations · order · context · provenanceHighroute by match confidenceDirect inheritanceMediumAdaptive editingLowKG reconstructionRetrievereuse historical planEditapply feature deltaConstructassemble from KGGeneratedProcess PlanUI Review& confirmation

Publication

2026

Y. Liang. RegSuitQA: Benchmarking Evidence-Grounded RAG for Chinese Financial Suitability Compliance.

Long Paper · Under Review at NLLP 2026

Under Review

Education & Professional Experience

Education

Shandong University

B.Sc. in Statistics · Data Science & Artificial Intelligence
Sep. 2022 — Jun. 2026 Average Score 86.53/100
Competitions
China International College Student Innovation Competition
Gold Award · Campus Round · 2025
ASC24 Student Supercomputer Challenge
Second Prize · 2024

Professional & Community Experience

SDIC Securities

Rotational Intern
Jan. 2026 — Feb. 2026

Conducted macro, industry, and company research and supported investment-strategy and asset-allocation analysis across new-consumption and AI sectors.

Guangdong South China Tiger Technology Co., Ltd.

AI Research Intern
Jun. 2025 — Dec. 2025

Supported RAG/LLM prototyping for an IPO document-intelligence system under mentor guidance, contributing to retrieval design, prompt engineering, and workflow testing.

Coding Girls

Volunteer Educator & Core Project Member
Jul. 2024 — Jun. 2025

Supported participant mentoring and coding-platform development, and designed a 10-lesson data-literacy curriculum for students in Xinjiang.

LLaMA-3 Architecture Reproduction

Reimplemented core LLaMA-3 modules in PyTorch to study transformer architecture and lightweight inference.

LLM Architecture PyTorch RoPE RMSNorm
GitHub

Literature Search & Topic Analysis

Built a literature retrieval platform with BERTopic topic modeling and research-trend exploration.

Information Retrieval BERTopic NLP
GitHub

Intelligent Consultation Agent

Developed a Chinese medical dialogue agent for symptom elicitation, preliminary assessment, and doctor recommendation.

AI Agent Dialogue System Chinese NLP
GitHub

Vision-Assisted Automatic Cat Feeder

Built an ESP32-CAM prototype linking lightweight vision recognition with hardware control for automatic feeding.

Computer Vision Embedded AI ESP32-CAM
GitHub