Jiaju Han

China University of Petroleum (Beijing)

B.Eng. in Data Science and Big Data Technology (Second Bachelor's Degree), Sep 2025 - Jun 2027

China University of Petroleum (Beijing)

B.A. in English, Sep 2021 - Jul 2025

CFGPatch: GEOMETRIC ADVERSARIAL FRAMEWORK FOR VISIBLE-INFRARED VLMs

May 2026 - Completed · Co-author · Submitted to NeurIPS

• Contributed to a unified geometric adversarial framework for visible-infrared vision-language models.
• Studied cross-task transferability across visible and infrared settings with shared fractal geometry construction.
• Participated in experiment organization, result analysis, and manuscript preparation.

Atmospheric Retrieval Hijacking in Remote Sensing Vision-Language RAG

May 2026 - Completed · First author · Submitted to NeurIPS

• Led a study on atmospheric-triggered retrieval hijacking in remote sensing vision-language RAG systems.
• Designed attack scenarios, retrieval evaluation protocols, and analysis for hallucination-inducing evidence retrieval.
• Coordinated manuscript writing and experimental validation across multimodal retrieval and generation settings.

XSPA: SPARSE ADVERSARIAL PERTURBATIONS FOR TRANSFERABLE ATTACKS ON VLMs

Oct 2025 - Completed · Second author · Submitted to ACM MM

• Contributed to implementation, experiments, evaluation pipeline, and manuscript preparation.
• Implemented a fixed X-shaped sparse perturbation pipeline with joint optimization over classification attack objectives, semantic attraction, semantic suppression, and smoothness regularization.
• Built batch evaluation workflows across zero-shot classification, image captioning, and VQA settings on multiple CLIP-style encoders and downstream VLMs.
• Achieved substantial degradation while perturbing about 1.76% of pixels at 224x224 resolution.

LLM-Based Two-Stage Credit Rating Explanation Generation System

Nov 2025 - Present

• Led a two-stage pipeline for rating prediction, explanation generation, automatic judging, and preference optimization.
• Improved rating classification to 89.54% accuracy and 88.80% Macro-F1 through data auditing and feature enhancement.
• Built a Llama-3.1-8B-Instruct + LoRA explanation generator covering SFT, candidate generation, LLM-as-judge evaluation, and DPO signal construction.
• Reached a 4.8675 / 5 average judge score and 98.8% pass rate on the test set.

ScholarRAG-ZH: Citation-Grounded QA Benchmark and RAG Evaluation Toolkit

Apr 2026 - Present

• Designed and implemented a benchmark and evaluation toolkit for academic RAG over TeX, Markdown, and PDF sources.
• Built an evidence-annotated benchmark with manually verified queries and metrics covering hit rate, citation precision/recall, answer accuracy, and faithfulness.
• Compared lexical retrieval, BM25, latent semantic retrieval, sentence-transformers embeddings, cross-encoder reranking, and structure-aware retrieval.
• Achieved 0.92 answer accuracy with structure-aware retrieval on the draft-derived split.

EXPOSING VULNERABILITIES IN VISIBLE-INFRARED VLMs: A UNIFIED GEOMETRIC ADVERSARIAL FRAMEWORK WITH CROSS-TASK TRANSFERABILITY

arXiv: https://arxiv.org/abs/2605.22273

Submission: NeurIPS · Author line includes Jiaju Han as fifth author.

FROM CLOUDS TO HALLUCINATIONS: ATMOSPHERIC RETRIEVAL HIJACKING IN REMOTE SENSING VISION-LANGUAGE RAG

arXiv: https://services.arxiv.org/html/submission/7564477/view

Submission: NeurIPS · Author position: First author

XSPA: CRAFTING IMPERCEPTIBLE X-SHAPED SPARSE ADVERSARIAL PERTURBATIONS FOR TRANSFERABLE ATTACKS ON VLMs

arXiv: https://arxiv.org/abs/2603.28568

Submission: ACM MM · Author position: Second author

WHEN SURFACES LIE: EXPLOITING WRINKLE-INDUCED ATTENTION SHIFT TO ATTACK VISION-LANGUAGE MODELS

arXiv: https://arxiv.org/abs/2603.27759

Submission: ACM MM · Author line includes Jiaju Han as third author.