Shahriar Noroozizadeh
ML+ISM PhD @ Carnegie Mellon Unviersity
snoroozi [AT] cs.cmu.edu

Deep sequence models tend to memorize geometrically; it is unclear why.
[Preprint: ArXiv]

Abstract: Deep sequence models are said to store atomic facts predominantly in the form of associative memory: a brute-force lookup of co-occurring entities. We identify a dramatically different form of storage of atomic facts that we term as geometric memory. Here, the model has synthesized embeddings encoding novel global relationships between all entities, including ones that do not co-occur in training. Such storage is powerful: for instance, we show how it transforms a hard reasoning task involving an ℓ-fold composition into an easy-to-learn 1-step navigation task.
From this phenomenon, we extract fundamental aspects of neural embedding geometries that are hard to explain. We argue that the rise of such a geometry, as against a lookup of local associations, cannot be straightforwardly attributed to typical supervisory, architectural, or optimizational pressures. Counterintuitively, a geometry is learned even when it is more complex than the brute-force lookup.
Then, by analyzing a connection to Node2Vec, we demonstrate how the geometry stems from a spectral bias that -- in contrast to prevailing theories -- indeed arises naturally despite the lack of various pressures. This analysis also points out to practitioners a visible headroom to make Transformer memory more strongly geometric. We hope the geometric view of parametric memory encourages revisiting the default intuitions that guide researchers in areas like knowledge acquisition, capacity, discovery, and unlearning.

SurvHTE-Bench: A Benchmark for Heterogeneous Treatment Effect Estimation in Survival Analysis
[Preprint Link Available Soon] [Data & Code]

Abstract: Estimating heterogeneous treatment effects (HTEs) from right-censored survival data is critical in high-stakes applications such as precision medicine and individualized policy-making. Yet, the survival analysis setting poses unique challenges for HTE estimation due to censoring, unobserved counterfactuals, and complex identification assumptions. Despite recent advances, from causal survival forests to survival meta-learners and outcome imputation approaches, evaluation practices remain fragmented and inconsistent. We introduce SurvHTE‐Bench, the first comprehensive benchmark for HTE estimation with censored outcomes. The benchmark spans (i) a modular suite of synthetic datasets with known ground truth, systematically varying causal assumptions and survival dynamics, (ii) semi-synthetic datasets that pair real-world covariates with simulated treatments and outcomes, and (iii) real-world datasets from a twin study (with known ground truth) and from an HIV clinical trial. Across synthetic, semi-synthetic, and real-world settings, we provide the first rigorous comparison of survival HTE methods under diverse conditions and realistic assumption violations. SurvHTE‐Bench establishes a foundation for fair, reproducible, and extensible evaluation of causal survival methods.

Forecasting Clinical Risk from Textual Time Series: Structuring Narratives for Temporal AI in Healthcare
[Paper: AAAI-2026] [Code]

Abstract: Clinical case reports encode rich, temporal patient trajectories that are often underexploited by traditional machine learning methods relying on structured data. In this work, we introduce the forecasting problem from textual time series, where timestamped clinical findings -- extracted via an LLM-assisted annotation pipeline -- serve as the primary input for prediction. We systematically evaluate a diverse suite of models, including fine-tuned decoder-based large language models and encoder-based transformers, on tasks of event occurrence prediction, temporal ordering, and survival analysis. Our experiments reveal that encoder-based models consistently achieve higher F1 scores and superior temporal concordance for short- and long-horizon event forecasting, while fine-tuned masking approaches enhance ranking performance. In contrast, instruction-tuned decoder models demonstrate a relative advantage in survival analysis, especially in early prognosis settings. Our sensitivity analyses further demonstrate the importance of time ordering, which requires clinical time series construction, as compared to text ordering, the format of the text inputs that LLMs are classically trained on. This highlights the additional benefit that can be ascertained from time-ordered corpora, with implications for temporal tasks in the era of widespread LLM use.

PubMed Open Access Textual Times Series Corpus: Reconstructing patient trajectories from clinical case reports using LLMs
[Preprint: ArXiv [PMOA-TTS]] [Code] [Data]
[Preprint: ArXiv [Textual Time Series corpus for Sepsis (T2S2)]]

Abstract: Understanding temporal dynamics in clinical narratives is essential for modeling patient trajectories, yet large-scale temporally annotated resources remain limited. We present PMOA-TTS, the first openly available dataset of 124,699 PubMed Open Access (PMOA) case reports, each converted into structured (event, time) timelines via a scalable LLM-based pipeline. Our approach combines heuristic filtering with Llama 3.3 to identify single-patient case reports, followed by prompt-driven extraction using Llama 3.3 and DeepSeek R1, resulting in over 5.6 million timestamped clinical events. To assess timeline quality, we evaluate against a clinician-curated reference set using three metrics: (i) event-level matching (80% match at a cosine similarity threshold of 0.1), (ii) temporal concordance (c-index > 0.90), and (iii) Area Under the Log-Time CDF (AULTC) for timestamp alignment. Corpus-level analysis shows wide diagnostic and demographic coverage. In a downstream survival prediction task, embeddings from extracted timelines achieve time-dependent concordance indices up to 0.82 ± 0.01, demonstrating the predictive value of temporally structured narratives. PMOA-TTS provides a scalable foundation for timeline extraction, temporal reasoning, and longitudinal modeling in biomedical NLP. The dataset is available at: https://huggingface.co/datasets/snoroozi/pmoa-tts.

The Impact of Medication Non-adherence on Adverse Outcomes: Evidence from Schizophrenia Patients via Survival Analysis
[Paper: CHIL-2025] [Code]

Abstract: This study quantifies the association between non-adherence to antipsychotic medications and adverse outcomes in individuals with schizophrenia. We frame the problem using survival analysis, focusing on the time to the earliest of several adverse events (early death, involuntary hospitalization, jail booking). We extend standard causal inference methods (T-learner, S-learner, nearest neighbor matching) to utilize various survival models to estimate individual and average treatment effects, where treatment corresponds to medication non-adherence. Analyses are repeated using different amounts of longitudinal information (3, 6, 9, and 12 months). Using data from Allegheny County in western Pennsylvania, we find strong evidence that non-adherence advances adverse outcomes by approximately 1 to 4 months. Ablation studies confirm that county-provided risk scores adjust for key confounders, as their removal amplifies the estimated effects. Subgroup analyses by medication formulation (injectable vs. oral) and medication type consistently show that non-adherence is associated with earlier adverse events. These findings highlight the clinical importance of adherence in delaying psychiatric crises and show that integrating survival analysis with causal inference tools can yield policy-relevant insights. We caution that although we apply causal inference, we only make associative claims and discuss assumptions needed for causal interpretation.

mRNA-LM: full-length integrated SLM for mRNA analysis
[Paper: Nucleic Acids Research Journal] [Code]
[Patent]

Abstract: The success of SARS-CoV-2 (severe acute respiratory syndrome coronavirus 2) messenger RNA (mRNA) vaccine has led to increased interest in the design and use of mRNA for vaccines and therapeutics. Still, selecting the most appropriate mRNA sequence for a protein remains a challenge. Several recent studies have shown that the specific mRNA sequence can have a significant impact on the translation efficiency, half-life, degradation rates, and other issues that play a major role in determining vaccine efficiency. To enable the selection of the most appropriate sequence, we developed mRNA-LM, an integrated small language model for modeling the entire mRNA sequence. mRNA-LM uses the contrastive language–image pretraining integration technology to combine three separate language models for the different mRNA segments. We trained mRNA-LM on millions of diverse mRNA sequences from several different species. The unsupervised model was able to learn meaningful biology related to evolution and host–pathogen interactions. Fine-tuning of mRNA-LM allowed us to use it in several mRNA property prediction tasks. As we show, using the full-length integrated model led to accurate predictions, improving on prior methods proposed for this task.

T5-generated clinical-Language summaries for DeBERTa Report Analysis (TLDR)
[Paper: SemEval-2024 at NAACL] [Code]

Abstract: This paper introduces novel methodologies for the Natural Language Inference for Clinical Trials (NLI4CT) task. We present TLDR (T5-generated clinical-Language summaries for DeBERTa Report Analysis) which incorporates T5-model generated premise summaries for improved entailment and contradiction analysis in clinical NLI tasks. This approach overcomes the challenges posed by small context windows and lengthy premises, leading to a substantial improvement in Macro F1 scores: a 0.184 increase over truncated premises. Our comprehensive experimental evaluation, including detailed error analysis and ablations, confirms the superiority of TLDR in achieving consistency and faithfulness in predictions against semantically altered inputs.

Temporal-Supervised Contrastive Learning: Modeling Patient Risk Progression
[Paper: ML4H] [Paper: AAAI - R2HCAI Workshop] [Code]

Abstract: We consider the problem of predicting how the likelihood of an outcome of interest for a patient changes over time as we observe more of the patient’s data. To solve this problem, we propose a supervised contrastive learning framework that learns an embedding representation for each time step of a patient time series. Our framework learns the embedding space to have the following properties: (1) nearby points in the embedding space have similar predicted class probabilities, (2) adjacent time steps of the same time series map to nearby points in the embedding space, and (3) time steps with very different raw feature vectors map to far apart regions of the embedding space. To achieve property (3), we employ a nearest neighbor pairing mechanism in the raw feature space. This mechanism also serves as an alternative to “data augmentation”, a key ingredient of contrastive learning, which lacks a standard procedure that is adequately realistic for clinical tabular data, to our knowledge. We demonstrate that our approach outperforms state-of-the-art baselines in predicting mortality of septic patients (MIMIC-III dataset) and tracking progression of cognitive impairment (ADNI dataset). Our method also consistently recovers the correct synthetic dataset embedding structure across experiments, a feat not achieved by baselines. Our ablation experiments show the pivotal role of our nearest neighbor pairing.

Contrastive Learning Based Interpretable Hospital Discharge Delay Prediction

Abstract: We addressed the significant challenge of delays in patient discharge across hospitals. Over an 11-month period, more than 63% of discharges at four UPMC hospitals were delayed, leading to costs of an estimated $6.6 million in the sampled hospital units. These delays adversely affect patient experience and health outcomes, exacerbated by issues like the lack of post-discharge patient transportation and ineffective capacity management in the health system. Throughout the CMLH fellowship, we aimed to mitigate these issues by developing a discharge delay prediction module. This initiative was divided into two phases: (1) Length of Stay Prediction: Various regression models were benchmarked using prehospital data. Predicting longer lengths of stays posed challenges, mainly due to their infrequent occurrence in the dataset. (2) Predictability Analysis: Building on initial insights, the prediction task was refined based on length of stay percentiles, identifying patients with more predictable stays versus those harder to forecast. A key innovation in this study was the application of a contrastive learning approach. This methodology significantly outperformed traditional models, including Random Forest, XGBoost, Support Vector Machines, Logistic Regression, and Fully-Connected Neural Networks. By leveraging the contrastive learning paradigm, the study offers a robust solution to predict patient discharge times, providing valuable guidance for hospital management and optimizing patient flow.

Pre-trained CLIP Encoder for Embodied Instruction Following in ALFRED
[Paper]

Abstract: We introduce a method employing pre-trained CLIP encoders to enhance model generalization in the ALFRED task. In contrast to previous literature where CLIP replaces the visual encoder, we suggest using CLIP as an additional module through an auxiliary object detection objective. We validate our method on the recently proposed Episodic Transformer architecture and demonstrate that incorporating CLIP improves task performance on the unseen validation set. Additionally, our analysis results support that CLIP especially helps with leveraging object descriptions, detecting small objects, and interpreting rare words.

Automatic Brain Pathology Analysis for Traumatic Brain Injury
[Paper]

Abstract: Traumatic brain injury (TBI) is one of the leading causes of death and disability worldwide. Detailed studies of the microglial response after TBI require high throughput quantification of changes in microglial count and morphology in histological sections throughout the brain. In this paper, we present a fully automated end-to-end system that is capable of assessing microglial activation in white matter regions on whole slide images of Iba1 stained sections. Our approach involves the division of the full brain slides into smaller image patches that are subsequently automatically classified into white and grey matter sections. On the patches classified as white matter, we jointly apply functional minimization methods and deep learning classification to identify Iba1-immunopositive microglia. Detected cells are then automatically traced to preserve their complex branching structure after which fractal analysis is applied to determine the activation states of the cells. The resulting system detects white matter regions with 84% accuracy, detects microglia with a performance level of 0.70 (F1 score, the harmonic mean of precision and sensitivity) and performs binary microglia morphology classification with a 70% accuracy. This automated pipeline performs these analyses at a 20-fold increase in speed when compared to a human pathologist. Moreover, we have demonstrated robustness to variations in stain intensity common for Iba1 immunostaining. A preliminary analysis was conducted that indicated that this pipeline can identify differences in microglia response due to TBI. An automated solution to microglia cell analysis can greatly increase standardized analysis of brain slides, allowing pathologists and neuroscientists to focus on characterizing the associated underlying diseases and injuries.

A.I. PhD Student Researcher [2025]: • Isolated a clean and analyzable instance of implicit in-weights reasoning in Transformers, demonstrating that their memory is better characterized by global geometric structure rather than purely local associative storage.
• Provided both empirical and theoretical evidence connecting this emergent geometric memory to spectral bias in Node2Vec-style dynamics, offering new insights into how Transformers generalize beyond local associations. [Paper] – Under Review
• Investigating the sufficiency of next-token prediction (NTP) as a training paradigm for large language models, and analyzing alternative multi-token objectives to enable deeper reasoning.
• Combining theoretical and empirical analysis in controlled synthetic settings and downstream language tasks to identify representational and learning bottlenecks inherent to NTP.
• Designing and executing large-scale experiments leveraging Google’s compute infrastructure, aiming to refine methods and produce publishable advances in sequence modeling objectives for language models.

A.I. Research Scientist Intern [2024]: • Co-led the development of the mRNA-LM model, a language model built from scratch and pretrained on millions of full-length mRNA sequences, achieving state-of-the-art performance on various mRNA prediction tasks, including structure prediction, localization, and translation efficiency.
• Designed and implemented a contrastive learning-based multimodal joint representation inspired by CLIP, which enhanced the alignment of embeddings from different mRNA regions (5' UTR, CDS, and 3' UTR) and significantly improved the downstream predictive performance of the full-length mRNA language model.
• Spearheaded submitting a paper on mRNA-LM to Nucleic Acids Research journal (IF: 16.8) and supported the filing of a patent for the mRNA-LM project, showcasing innovative methodologies and findings. [Paper] [Patent]
• Contributed to the project Many-Shot In-Context Learning for Molecular Inverse Design, developing a semi-supervised learning method utilizing Large Language Models (LLMs) to improve molecular design and lead optimization. Implemented a multi-modal LLM framework for interactive molecular structure modification using text instructions.
• Collaborated on integrating Large Language Models (LLMs) into the Bayesian Optimization framework to guide optimization directions for reaction yield in drug discovery, achieving superior performance compared to human experts in selecting optimal reactions and refining design pipelines.

Software Development Engineering Intern [2015]: Main focus areas researched and worked on during this internship included: Windows 10 Universal Application Platform (UAP), Windows 10 NFL Application, Development of a Key Performance Indicator (KPI) System, Mocking Framework Development, Coded User Interface (UI) Automation and Build Machine Automation Development.

Research and Development Intern [2016]: Developed an electronic nose sensor that is capable of selectively and sensitively detect biomarkers in exhaled breath to improve the emergency diagnosis of lung infections for patients with respiratory diseases including Acute Respiratory Distress Syndrome (ARDS). I designed a standalone signal processing algorithm and application tailored for gas chromatography data, which effectively isolated the presence of octane—a critical biomarker of ARDS in exhaled breath. This application was instrumental in enhancing the accuracy and reliability of the electronic nose sensor, especially given the challenges of non-real-time data processing.