2026

CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment

Chengxiao Wang*, Enyi Jiang*, Xiaojing Liao, Sanmi Koyejo (* equal contribution)

arXiv preprint 2026

CLEAR uses a lightweight hidden-state gate to continuously control the activation strength of a safety low-rank adapter, cutting HarmBench attack success on Llama-3-8B-Instruct from 32.3% to 0.5% while preserving utility -- up to 7.1 points higher GSM8K accuracy than standard safety fine-tuning or LoRA.

CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment

Chengxiao Wang*, Enyi Jiang*, Xiaojing Liao, Sanmi Koyejo (* equal contribution)

arXiv preprint 2026

CLEAR uses a lightweight hidden-state gate to continuously control the activation strength of a safety low-rank adapter, cutting HarmBench attack success on Llama-3-8B-Instruct from 32.3% to 0.5% while preserving utility -- up to 7.1 points higher GSM8K accuracy than standard safety fine-tuning or LoRA.

When Behavioral Safety Evaluation Fails: A Representation-Level Perspective

Enyi Jiang*, Anders Gjølbye*, Yibo Jacky Zhang, Sanmi Koyejo (* equal contribution)

arXiv preprint 2026

Behavioral safety audits miss an "audit gap": we construct models that pass every static refusal test yet comply with harmful requests under small perturbations of their internal states, with dissociated models scoring 2.5-3.1x higher vulnerability than their bases.

When Behavioral Safety Evaluation Fails: A Representation-Level Perspective

Enyi Jiang*, Anders Gjølbye*, Yibo Jacky Zhang, Sanmi Koyejo (* equal contribution)

arXiv preprint 2026

Behavioral safety audits miss an "audit gap": we construct models that pass every static refusal test yet comply with harmful requests under small perturbations of their internal states, with dissociated models scoring 2.5-3.1x higher vulnerability than their bases.

Environment-Adaptive Preference Optimization for Wildfire Prediction

Enyi Jiang, Wu Sun

arXiv preprint 2026

EAPO predicts rare wildfire events under shifting environmental conditions by retrieving distribution-aligned training data with $k$-NN and combining supervised fine-tuning with preference optimization, reaching 0.7310 ROC-AUC with better detection in extreme regimes.

Environment-Adaptive Preference Optimization for Wildfire Prediction

Enyi Jiang, Wu Sun

arXiv preprint 2026

EAPO predicts rare wildfire events under shifting environmental conditions by retrieving distribution-aligned training data with $k$-NN and combining supervised fine-tuning with preference optimization, reaching 0.7310 ROC-AUC with better detection in extreme regimes.

Latent Adversarial Regularization for Offline Preference Optimization

Enyi Jiang, Yibo Jacky Zhang, Yinglun Xu, Andreas Haupt, Nancy Amato, Sanmi Koyejo

arXiv preprint 2026

GANPO regularizes offline preference optimization in latent space, using a GAN-inspired adversarial objective to penalize divergence between the policy and reference model's internal representations instead of token-level similarity. This gives more robust structural feedback under distribution shift and noise at minor computational cost.

Latent Adversarial Regularization for Offline Preference Optimization

Enyi Jiang, Yibo Jacky Zhang, Yinglun Xu, Andreas Haupt, Nancy Amato, Sanmi Koyejo

arXiv preprint 2026

GANPO regularizes offline preference optimization in latent space, using a GAN-inspired adversarial objective to penalize divergence between the policy and reference model's internal representations instead of token-level similarity. This gives more robust structural feedback under distribution shift and noise at minor computational cost.

2025

Towards Generalized Certified Robustness with Multi-Norm Training

Enyi Jiang, David S. Cheung, Gagandeep Singh

Transactions on Machine Learning Research (TMLR) 2026

CURE is the first multi-norm certified training framework, lifting certified training beyond a single $l_p$ threat model via bound alignment and a natural-to-certified training connection. It improves union robustness by 32.0% on MNIST, 25.8% on CIFAR-10, and 10.6% on TinyImageNet, and generalizes to unseen geometric and patch perturbations.

Towards Generalized Certified Robustness with Multi-Norm Training

Enyi Jiang, David S. Cheung, Gagandeep Singh

Transactions on Machine Learning Research (TMLR) 2026

CURE is the first multi-norm certified training framework, lifting certified training beyond a single $l_p$ threat model via bound alignment and a natural-to-certified training connection. It improves union robustness by 32.0% on MNIST, 25.8% on CIFAR-10, and 10.6% on TinyImageNet, and generalizes to unseen geometric and patch perturbations.

Robust Answers, Fragile Logic: Probing the Decoupling Hypothesis in LLM Reasoning

Enyi Jiang, Changming Xu, Nischay Singh, Tian Qiu, Gagandeep Singh

Transactions on Machine Learning Research (TMLR) 2026

MATCHA, an answer-conditioned probing framework, shows that reasoning robustness and answer robustness are largely decoupled: under imperceptible perturbations LLMs keep the correct answer while producing inconsistent or nonsensical rationales, and the resulting adversarial examples transfer to black-box models.

Robust Answers, Fragile Logic: Probing the Decoupling Hypothesis in LLM Reasoning

Enyi Jiang, Changming Xu, Nischay Singh, Tian Qiu, Gagandeep Singh

Transactions on Machine Learning Research (TMLR) 2026

MATCHA, an answer-conditioned probing framework, shows that reasoning robustness and answer robustness are largely decoupled: under imperceptible perturbations LLMs keep the correct answer while producing inconsistent or nonsensical rationales, and the resulting adversarial examples transfer to black-box models.

2024

RAMP: Boosting Adversarial Robustness Against Multiple $l_p$ Perturbations for Universal Robustness

Enyi Jiang, Gagandeep Singh

Neural Information Processing Systems (NeurIPS) 2024

RAMP uses a logit pairing loss and gradient projection to balance the accuracy/robustness tradeoff across multiple $l_p$ threat models at once, reaching 53.3% union accuracy on CIFAR-10 by fine-tuning and 44.6% from scratch with 81.2% clean accuracy on ResNet-18, plus better robustness to unseen adversaries and corruptions.

RAMP: Boosting Adversarial Robustness Against Multiple $l_p$ Perturbations for Universal Robustness

Enyi Jiang, Gagandeep Singh

Neural Information Processing Systems (NeurIPS) 2024

RAMP uses a logit pairing loss and gradient projection to balance the accuracy/robustness tradeoff across multiple $l_p$ threat models at once, reaching 53.3% union accuracy on CIFAR-10 by fine-tuning and 44.6% from scratch with 81.2% clean accuracy on ResNet-18, plus better robustness to unseen adversaries and corruptions.

Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting

Enyi Jiang*, Yibo Jacky Zhang*, Sanmi Koyejo (* equal contribution)

International Conference on Learning Representations (ICLR) 2024

A theoretical framework for federated domain adaptation that reveals a "gradient projection" insight, yielding two practical methods -- Federated Gradient Projection and an auto-weighting scheme that optimally combines source gradients -- with strong gains when the target data is scarce.

Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting

Enyi Jiang*, Yibo Jacky Zhang*, Sanmi Koyejo (* equal contribution)

International Conference on Learning Representations (ICLR) 2024

A theoretical framework for federated domain adaptation that reveals a "gradient projection" insight, yielding two practical methods -- Federated Gradient Projection and an auto-weighting scheme that optimally combines source gradients -- with strong gains when the target data is scarce.

2023

Federated Domain Adaptation for Healthcare

Enyi Jiang

M.S. Thesis, University of Illinois Urbana-Champaign 2023

My master's thesis studies the federated domain adaptation problem for healthcare applications, both empirically and theoretically.

Federated Domain Adaptation for Healthcare

Enyi Jiang

M.S. Thesis, University of Illinois Urbana-Champaign 2023

My master's thesis studies the federated domain adaptation problem for healthcare applications, both empirically and theoretically.

2020

EVIDENCEMINER: Textual Evidence Discovery for Life Sciences

Xuan Wang, Yingjun Guan, Weili Liu, Aabhas Chauhan, Enyi Jiang, Qi Li, David Liem, Dibakar Sigdel, John Caufield, Peipei Ping, Jiawei Han

Annual Meeting of the Association for Computational Linguistics (ACL), System Demonstrations 2020

EvidenceMiner is a web-based system that retrieves sentence-level textual evidence for a natural-language query from a large life-science literature corpus, built without human annotation using distantly supervised named entity recognition and open information extraction.

EVIDENCEMINER: Textual Evidence Discovery for Life Sciences

Xuan Wang, Yingjun Guan, Weili Liu, Aabhas Chauhan, Enyi Jiang, Qi Li, David Liem, Dibakar Sigdel, John Caufield, Peipei Ping, Jiawei Han

Annual Meeting of the Association for Computational Linguistics (ACL), System Demonstrations 2020

EvidenceMiner is a web-based system that retrieves sentence-level textual evidence for a natural-language query from a large life-science literature corpus, built without human annotation using distantly supervised named entity recognition and open information extraction.