Two-Sided State-Space Models for Sequential Recommendation with Non-Random Multimodal Review Feedback
Ziwen Pan*, Zihan Liang*, Ruoxuan Xiong
Accepted at EMNLP 2026 Findings
- TS-SSM jointly tracks evolving user preferences and item states, treating multimodal reviews—and how they are observed and expressed—as non-random feedback, while capturing cross-item spillovers and the asymmetric persistence of positive and negative reviews.
- Across six Amazon datasets, TS-SSM outperforms strong sequential, multimodal, and debiasing baselines, achieving 14.8%–18.8% relative improvements in Recall@20 over the strongest sequential baseline.
DART: Mitigating Harm Drift in Difference-Aware LLMs via Distill-Audit-Repair Training
Ziwen Pan*, Zihan Liang*, Jad Kabbara, Ali Emami
Accepted at ACL 2026 Findings
- Identified the problem of harm drift in safety-tuned LLMs—where fine-tuning improves decision accuracy but makes rationales more harmful—and introduced DART, a Distill-Audit-Repair pipeline for improving difference-awareness classification while preserving safer explanations.
- On eight benchmarks, DART improved Llama-3-8B-Instruct accuracy from 39.0% to 68.8%, boosted equal-treatment accuracy from 11.3% to 72.6%, and reduced harm drift cases by 72.6%, showing that accuracy and safety can be improved together.
Learning Dynamic Representations and Policies from Multimodal Clinical Time-Series with Informative Missingness
Zihan Liang*, Ziwen Pan*, Ruoxuan Xiong
Accepted at ACL 2026 Findings
- Proposed an MNAR-aware multimodal framework for clinical time-series that treats observation patterns as signal rather than noise, combining explicit missingness features, adaptive fusion of sparse clinical text, and action-conditioned latent dynamics to learn patient states for both prediction and offline decision-making.
- Demonstrated strong results on MIMIC-IV and eICU: mortality prediction reached AUROC 0.876 on MIMIC-IV (+3.9% over GRU-D), while the learned treatment policy improved FQE by 20.3% over clinician behavior and delivered the largest gains for high-severity patients.
MambaDATG: Domain-Adaptive Tri-Plane-Gated Pre-training for 3D Abdominal Segmentation
Yaomin Shen*, Dongming Jiang*, Zihan Liang*, Yangbo Wei*, Wenkai Yang, Xiaoxin Sun, Zhen Huang, Suhua Wang, Qingsong Yao
Accepted at ICASSP 2026 Oral
- Proposed MambaDATG, a tri-plane-gated selective SSM pre-training that fuses axial, coronal, and sagittal scans via voxel gating—capturing directional anisotropy with linear-time efficiency and zero inference overhead.
- Introduced a domain-adaptive MIM stage adapting high-level layers on unlabeled CTs, improving small-organ segmentation (e.g., +1.8 Dice on BTCV) and revealing the tri-plane gate as the key performance driver.
Causal Representation Learning from Multimodal Clinical Records under Non-Random Modality Missingness
Zihan Liang*, Ziwen Pan*, Ruoxuan Xiong
Accepted at EMNLP 2025 Main
- Proposed CRL-MMNAR, a causal multimodal framework that treats modality missing-not-at-random as signal—combining missingness-aware fusion (gated by observation patterns) with cross-modal reconstruction + contrastive learning, and a multitask predictor with a cross-fitted rectifier to correct observation-pattern bias.
- Demonstrated consistent gains on MIMIC-IV and eICU, including AUC 0.8687→0.9824 for ICU admission (+13.1%) and 0.7989→0.8657 for readmission (+8.4%) on MIMIC-IV, and readmission AUC to 0.9294 (+13.8%) on eICU, with lower Brier scores indicating better calibration.
CareLab at #SMM4H-HeaRD 2025: Insomnia Detection and Food Safety Event Extraction with Domain-Aware Transformers
Zihan Liang*, Ziwen Pan*, Sumon Kanti Dey, Azra Ismail
Accepted at AAAI ICSWM 2025
- Ranked 1st on SMM4H–HeaRD 2025 Task 5, developing a RoBERTa + GPT-4–augmented system with classweighted training and ensembling, achieving F1 = 0.958 and revealing the limits of rule-based span extraction in clinical and food-safety texts.
Dynamic Policy Design for Autonomous Taxi Adoption: A Hierarchical Game-Theoretic Framework
Zihan Liang, Ziwen Pan
Accepted at TRB Annual Meeting 2026
- Formulated a three-tier dynamic Stackelberg game modeling interactions among government, platform, and workers in autonomous taxi adoption, deriving equilibrium policies that balance innovation incentives with social welfare.
- An extended version is under review at Transportation Research Record.