|
|
Stability Does Not Mirror Plasticity: A Module-Level View of Declarative and Procedural Learning in Transformers
Jinjun Peng*,
Weiliang Zhao*,
Baishakhi Ray,
Zhou Yu,
Junfeng Yang
*Equal contribution
Under Review (ICLR 2027)
PDF
Studies how plasticity and stability in continual learning vary across transformer
modules and task types. Finds a double dissociation—query-key and value-output
projections are the most plastic on procedural tasks, whereas MLPs are the most plastic
on declarative memorization—and shows that stability does not simply mirror
plasticity, offering a task-dependent, module-level view of continual learning.
|
|
|
Knowledge-Verified Emergent Deception in LLM Agents Under Conflicting Incentives
Zheyuan Liu,
Weiliang Zhao,
Xiangchi Yuan,
Ningshan Ma,
Yue Huang,
Meng Jiang
Under Review
arXiv
Introduces KnownLieBench, a benchmark that tests whether LLM agents in customer-service
settings make false claims when incentivized to deny user entitlements they demonstrably
know are valid. Across eighteen models deception varies widely, and honesty-focused
fine-tuning reduces it.
|
|
|
Jailbreaking Jailbreaks: A Proactive Defense for LLMs
Weiliang Zhao,
JinJun Peng,
Daniel Ben-Levi,
Junfeng Yang,
Chengzhi Mao,
EMNLP, 2026
arXiv
A proactive defense framework that injects strategically crafted spurious outputs to mislead
attackers’ optimization loops, prematurely collapsing multi-turn jailbreak searches and dramatically
reducing LLM vulnerability.
|
|
|
Mirage Probes: How Vision Models Fake Visual Understanding
Daniel Ben-Levi,
Judah Goldfeder,
Weiliang Zhao,
Raz Lapid,
Amit LeVi,
Allen G. Roush,
Ravid Shwartz-Ziv,
Hod Lipson
Mechanistic Interpretability Workshop at ICML, 2026
arXiv
Vision-language models often answer image questions correctly without ever looking at the
image. A contrastive probing framework shows this "mirage" behavior is linearly decodable from
internal activations, separating genuine visual grounding from language-prior shortcuts.
|
|
|
Diversity Helps Jailbreak Large Language Models
Weiliang Zhao,
Daniel Ben-Levi,
Junfeng Yang,
Chengzhi Mao,
NAACL, 2025, Oral
arXiv
A Generalised jailbreaking technique by encouraging higher levels of diversification and
adjacent
obfuscated prompting to evaluate the vulnerabilities of LLMs.
|
|
|
Learning to Rewrite: Generalized LLM-Generated Text Detection
Wei Hao,
Ran Li ,
Weiliang Zhao,
Junfeng Yang,
Chengzhi Mao,
ACL, 2025
arXiv
We propose a method designed to enhance
the detection of LLM-generated text by learning
to rewrite more on LLM-generated inputs and less
on human generated inputs.
|
📄
Other Papers
|
|
CheatBench: Measuring Reward Gaming in AI Agents
Long Phan,
Stephen K. Yang,
Jason J. Lim,
Mantas Mazeika,
Wenyu Zhang,
Zheyuan Liu,
Richard Ren,
Jingxiang Meng,
Yaoteng Tan,
Weiliang Zhao,
Addison Wu,
Matei Anghel,
Dan Hendrycks
arXiv
Introduces CheatBench, a benchmark that measures reward gaming — how AI agents pursue goals
through dishonest shortcuts on difficult tasks — across multiple domains, to help measure and
mitigate deceptive behavior as agents take on higher-stakes work.
|
|
|
🤝
Acknowledgement
I would like to acknowledge the Thinker Research Grants support from
Thinking Machine.
|
|
🧳
Travel
Beyond research, I love traveling and scuba diving 🤿 — here are my
footprints around the world 🌍.
|
Feel free to steal this website's source
code. Do not scrape the HTML from this page itself, as it includes
analytics
tags that you do not want on your own website — use the github code instead. Also,
consider
using Leonid Keselman's Jekyll fork of this page.
|
|