Safety Mirage: How Spurious Correlations Undermine VLM Safety Fine-Tuning and Can Be Mitigated by Machine Unlearning
Yiwei Chen*, Yuguang Yao*, Yihua Zhang, Bingquan Shen, Gaowen Liu, Sijia Liu
ICLR, 2026
Code / Paper

Conventional safety fine-tuning of MLLMs suffers from a ``safety mirage'' caused by training bias, leading to spurious correlations and over-rejections under one-word attacks. Employing unlearning algorithms effectively remove harmful content and mitigate these issues.

Unlearning Isn't Invisible: Detecting Unlearning Traces in LLMs from Model Outputs
Yiwei Chen*, Soumyadeep Pal*, Yimeng Zhang, Qing Qu, Sijia Liu
ICLR, 2026. Short version accepted for Oral at MUGen @ ICML'25
Code / Paper

Large language models retain persistent fingerprints in their outputs and hidden activations even after unlearning. These traces enable detection of whether a model has been “unlearned,” exposing a new vulnerability in reverse-engineering forgotten information.

Who Built This Model? Tracing LLM Lineage via Spectral Fingerprints in Weight Space
Yiwei Chen*, Bingqi Shang*, Sijia Liu
In Submission, 2026
Paper

A geometric fingerprinting framework to trace LLM lineage in weight space, combining spectral energy for coarse-grained discrimination and subspace alignment for fine-grained analysis, enabling data-free model provenance identification.

Data-Centric Benchmarking of Exploit Generation in LLMs: Understanding the Impact of Fine-Tuning
Yiwei Chen*, Lichi Li*, Kai Cheung, Vinny Parla, Ganesh Sundaram
Manuscript, 2026
Paper

The first data-centric benchmark for CVE-conditioned exploit generation with a 6-level context hierarchy and 8-criterion evaluation framework, benchmarking 17 LLMs. Fine-tuned Qwen3-8B with reasoning-aware fine-tuning achieves +42.5% improvement, competitive with frontier LLMs.

Forgetting to Forget: Attention Sink as A Gateway for Backdooring LLM Unlearning
Bingqi Shang*, Yiwei Chen*, Yihua Zhang, Bingquan Shen, Sijia Liu
Under Review, 2025
Code / Paper

Demonstrated that LLM unlearning can be compromised through attention-sink-guided backdoor unlearning, where triggers placed at attention sinks enable models to recover forgotten knowledge while maintaining normal behavior without triggers.

Tile Classification Based Viewport Prediction with Multi-modal Fusion Transformer
Zhihao Zhang*, Yiwei Chen*, Weizhan Zhang, Caixia Yan, Qinghua Zheng, Qi Wang, Wangdu Chen
ACM MM, 2023
Code / Paper

Propose a tile classification based viewport prediction method with Multi-modal Fusion Transformer to improve the robustness of viewport prediction.