Published in arxiv, 2023
Explores efficient solutions for transportation optimization via model based curriculum learning
Recommended citation: Shah, A., Tran, D., Tang, Y. (2023). "Efficient Mitigation of Bus Bunching through Setter-Based Curriculum Learning." http://shavidan123.github.io/files/CS285_Bus_Bunching_Final_Project.pdf
Published in NAACL, 2025
A novel algorithm leveraging model refusal representation for automated jailbreaking suffix generation on LLMs
Recommended citation: Huang, D., Shah, A., Araujo, A., Wagner, D., & Sitawarin, C. (2025). Stronger universal and transfer attacks by suppressing refusals. NAACL 2025. https://shavidan123.github.io/files/NAACL__Stronger_Universal_and_Transferable_Attacks_by_Refusal_Suppression.pdf
Published in ICML (Spotlight), 2026
Practical prompt-injection-based poisoning attacks against the Rapid Response jailbreak detection framework, including a novel Omission Attack that flips ~90% of target labels with only 1% poisoning rate.
Recommended citation: Huang, D., Chang, J., Shah, A., Mittal, P., & Sitawarin, C. (2026). Rapid Poison: Practical poisoning attacks against the Rapid Response framework. ICML 2026. https://shavidan123.github.io/files/Rapid_Poison__ICLR_Workshop_ (3).pdf
Published in arxiv, 2026
On-Policy Consistency Training (OPCT) supervises a model on its own responses conditioned on contrastive prompts, reducing sycophancy and maintaining jailbreak defense without the capability degradation that supervised fine-tuning induces.
Recommended citation: Han, A., Fujimoto, K., Shah, A., Nguyen, K., Xu, K., Yueh-Han, C., Sucholutsky, I., & Angell, R. (2026). On-policy consistency training improves LLM safety with minimal capability degradation. arXiv preprint arXiv:2605.21834. https://shavidan123.github.io/files/OPCT_On_Policy_Consistency_Training.pdf
Published in arxiv, 2026
Activation Consistency Training (ACT) supervises internal representations rather than outputs, providing competitive defense against jailbreaks and prompt injection in reasoning LLMs while remaining robust to adaptive attacks.
Recommended citation: Shah, A., Brinkmann, J., & Angell, R. (2026). Mitigating adaptive attacks against reasoning models with activation consistency training. arXiv preprint arXiv:2605.28467. https://shavidan123.github.io/files/ACT_Activation_Consistency_Training.pdf
Published in Preprint, 2026
Characterizes covert influence between language models, where behavioral traits transfer through channels undetectable by humans, across supervised fine-tuning, on-policy distillation, and in-context learning, using inference-time per-sample attribution to select carriers that amplify influence.
Recommended citation: Shah, A., Chooi, J., Ou, J., & Feng, S. (2026). Covert influence between language models. Preprint. https://shavidan123.github.io/files/Covert_Influence_Between_Language_Models.pdf
Undergraduate course, University 1, Department, 2014
This is a description of a teaching experience. You can use markdown like any other post.
Workshop, University 1, Department, 2015
This is a description of a teaching experience. You can use markdown like any other post.