WiDS Cambridge
Poster Session 2026

Presenter: Verónica Álvarez
Affiliation: MIT – Laboratory for Information & Decision Systems
Title: Reliable Programmatic Weak Supervision With Confidence Intervals for Label Probabilities
Authors: Verónica Álvarez , Santiago Mazuelas , Steven An, and Sanjoy Dasgupta
Abstract:
The accurate labeling of datasets is often both costly and time-consuming. Given an unlabeled dataset, programmatic weak supervision obtains probabilistic predictions for the labels by leveraging multiple weak labeling functions (LFs) that provide rough guesses for labels. Weak LFs commonly provide guesses with assorted types and unknown interdependences that can result in unreliable predictions. Furthermore, existing techniques for programmatic weak supervision cannot provide assessments for the reliability of the probabilistic predictions for labels. This poster presents a methodology for programmatic weak supervision that can provide confidence intervals for label probabilities and obtain more reliable predictions. In particular, the methods proposed use uncertainty sets of distributions that encapsulate the information provided by LFs with unrestricted behavior and typology. Experiments on multiple benchmark datasets show the improvement of the presented methods over the state-of-the-art and the practicality of the confidence intervals presented.

​

Presenter: Madelyn Andersen
Affiliation: MIT - Laboratory for Information & Decision Systems
Title: Challenges and Ways Forward for Variational Inference Software: Default Settings, Usability, and Reliability
Authors: Madelyn Andersen, Elizabeth Bersson, Tamara Broderick
Abstract:
We systematically evaluate a suite of off-the-shelf variational inference (VI) software packages from the perspective of a standard practitioner. Using simple analytic benchmark models, we assess the accuracy and stability of the default VI settings in PyMC, NumPyro, and TensorFlow Probability. Unlike previous research focusing on methodological advances, our evaluation emphasizes software implementation and the default configurations that typical users encounter. Our results show that default settings can yield biased approximations of posterior summaries even for simple one-dimensional conjugate models, controls of initialization and transformations differ between software implementations, and relying on defaults may yield silent failures or poor approximations.

​

Presenter: Taylor Baum
Affiliation: MIT 
Title: An Adaptive Closed-Loop System for Arterial Blood Pressure Control with Clevidipine and Phenylephrine
Authors: Taylor Baum, Elie Adam, Reza Kazemi, Gywn Arney-Sutherland, Sebastian Gallo, Mosa Al Zowelei, Julian Goldman, Thomas Heldt, Munther Dahleh, Emery Brown
Abstract:
Poor regulation of arterial blood pressure (ABP) in the operating room or intensive care unit is associated with increased morbidity and mortality. Although closed-loop ABP control has been pursued for decades, clinical adoption has remained extremely limited. In this work, we present a novel bidirectional adaptive closed-loop ABP control system. Our system comprises two adaptive model predictive controllers: one actuated by phenylephrine (PE) and the other by clevidipine (CL). Through validation experiments in swine, we find that, in non-emergent, post-transient periods, our PE-actuated controller maintained mean ABP within ± 5 mmHg of the target 99.7% of the time (95% CI: 99.2, 100.0) and has 1.2 mmHg (95% CI: 0.8, 1.4) root mean square error (RMSE) and our CL-controller, 97.1% (95% CI: 94.1, 98.9) with 1.7 mmHg (95% CI: 1.4, 2.3) RMSE. We additionally find that both controllers rapidly rejected severe disturbances and that mid-session switching between controllers was achieved safely. These results demonstrate the attainable precision and stability of bidirectional adaptive closed-loop ABP control and support its potential for clinical translation.

​

Presenter: Parmida Davarmanesh
Affiliation: MIT
Title:  Efficient and accurate steering of Large Language Models through attention-guided feature learning
Authors: Parmida Davarmanesh, Ashia Wilson, Adityanarayanan Radhakrishnan
Abstract:
Steering, or direct manipulation of internal activations to guide LLM responses toward specific semantic concepts, is emerging as a promising avenue for both understanding how semantic concepts are stored within LLMs and advancing LLM capabilities. Yet, existing steering methods are remarkably brittle, with seemingly non-steerable concepts becoming completely steerable based on subtle algorithmic choices in how concept-related features are extracted. In this work, we introduce an attention-guided steering framework that overcomes three core challenges associated with steering: (1) automatic selection of relevant token embeddings for extracting concept-related features; (2) accounting for heterogeneity of concept-related features across LLM activations; and (3) identification of layers most relevant for steering. Across a steering benchmark of 512 semantic concepts, our framework substantially improved steering over previous state-of-the-art (nearly doubling the number of successfully steered concepts) across model architectures and sizes (up to 70 billion parameter models). Furthermore, we use our framework to shed light on the distribution of concept-specific features across LLM layers. Overall, our framework opens further avenues for developing efficient, highly-scalable fine-tuning algorithms for industry-scale LLMs.

​

Presenter: Zerin Khan
Affiliation: Harvard University
Title: Cellular backpack-mediated natural killer cell immunotherapy to target glioblastoma
Authors: Zerin Mahzabin Khan and Samir Mitragotri
Abstract:
Glioblastoma (GBM) is the most aggressive type of primary brain cancer in adults. Even after surgery and chemoradiotherapy, GBM cells survive, form secondary tumors in 90% of patients, and lead to median survival times of 15 months. GBM secretes immunosuppressive factors contributing to this therapy resistance, as immune cells in the brain become exhausted and cannot attack the cancer cells. There is a dire need to target GBM by transforming the microenvironment from immunosuppressive to immunoactive. We developed novel polymeric materials that attach to the surface of natural killer (NK) cells to activate these immune cells for secretion of immunoactive factors. We hypothesize these cellular backpacks can transform the “cold” immunosuppressive microenvironment to a “hot” immunoactive one. In this study, we investigate backpack-carrying NK cell (NK-BP) immunotherapy to treat GBM by using 3D, human physiologically relevant in vitro models to mimic GBM. We demonstrate high 80% backpack binding to human NK cells and show these NK-BPs significantly kill human GBM cells in a dose-dependent manner. We are currently quantitatively profiling the GBM and NK-BP biophysical interactions (e.g. migration, cell morphologies, treatment responses, etc). Trends extracted from these single-cell datasets will lend critical insights into GBM-immune cell interactions to treat GBM.

​

Presenter: Nina Klimenkova
Affiliation: Worcester Polytechnic Institute
Title: Table Unionability Is Uncertain and That’s Why Humans and AI Need Each Other
Authors: Nina Klimenkova
Abstract:
In today's data-driven world, the ability to combine datasets from different sources is crucial for scientific discovery and innovation. But how do we know when two data tables actually belong together? This problem, known as table unionability, turns out to be surprisingly hard, even for data specialists. This poster is based on the study, where we explore how people make unionability decisions and where they struggle. Through a behavioral survey, we tracked participants' clicks, decision times, and confidence levels as they judged whether table pairs could be meaningfully merged. We found something counterintuitive: the longer people think, the less accurate and confident they become, suggesting that overthinking invite difficult cases rather than better answers. To address this, we built machine learning models that use these behavioral signals to predict when a human judgment is likely correct, producing cleaner, more reliable labels. This approach boosted accuracy by up to 25.5% over raw human responses. We also tested large language models on the same task and found they do not consistently outperform humans. However, combining human and AI judgments proved highly effective. Our findings highlight that neither humans nor AI can solve this problem alone but together, they get much closer.

​

Presenter: Yingke Li
Affiliation: MIT - Laboratory for Information & Decision Systems
Title: How Curious Should an Experiment Be? Guarantees for Learning and Regret in Active Inference
Authors: Yingke Li
Abstract:
Adaptive experimentation—such as A/B testing, online policy evaluation, and sequential clinical studies—requires making decisions while learning: we must allocate users or trials to estimate treatment effects quickly, yet minimize opportunity cost from exposing traffic to inferior variants. This poster presents a principled approach based on active inference, where decisions are made by minimizing Expected Free Energy (EFE). EFE yields a single objective that naturally combines an epistemic drive (expected information gain) with a pragmatic drive (expected regret), linked by a curiosity coefficient that controls how aggressively the experiment explores. Despite strong empirical use of curiosity-driven objectives, it has been unclear when this balance guarantees reliable learning and efficient performance. We provide the first end-to-end theory showing that sufficient curiosity is the shared mechanism that ensures both self-consistent learning (Bayesian posterior consistency) and no-regret optimization (bounded cumulative regret). We also highlight the role of heuristic alignment: if the pragmatic surrogate is biased (e.g., proxy metrics vs true outcomes), regret can accumulate even with exploration. Controlled experiments isolate these effects and translate the results into practical guidance for setting curiosity and adapting it over time in real adaptive experiments.

​

Presenter: Sonia Murthy
Affiliation: Prior Computers Inc
Title: Looking for cognitive model parameters in the residual stream
Authors: Sonia K. Murthy, Peng Qian, Robert Hawkins, Felix A. Sosa
Abstract:
Approaches to extracting and steering dimensions of an LLM’s “persona” often rely on noisy datasets or imprecise, linguistic descriptions to define dimensions of complex behavioral profiles. While sometimes useful, these persona vectors can be confounded with both spurious features of the data and our own biases in defining those vectors, which often leads to unexpected behaviors when that dimension is steered. Cognitive models of human behavior often detail computational definitions of dimensions relevant to personality, such as being informative or being polite, offering more precision over linguistic descriptions of those dimensions. In this work, we ask whether we can use a cognitive model of polite speech in humans to find dimensions in the residual stream of LLMs that correspond to parameters in the cognitive model, providing more coherent and interpretable steering vectors than traditional contrastive approaches. We identify steering vectors that can be interpreted according to the individual setting of the parameters in the cognitive model and present initial results of these “cognitive vectors” relative to linguistic- and data-driven vectors.

​

Presenter: Alana Motley
Affiliation: Outten & Golden
Title: Mapping the Talent Graph: Labor Mobility Networks and Workforce Shock Dynamics in the U.S. Tech Industry
Authors: Alana Motley, Mona Birjandi
Abstract:
This project introduces a novel firm-level labor mobility network constructed from large-scale, individual career histories of the U.S. tech workforce between 2020 and 2024 using LinkedIn data. Using detailed transition data, we transform raw career trajectories into directed, weighted firm-to-firm graphs that map how technology firms are structurally connected through talent flows. We then integrate these networks with layoff records covering more than 200,000 displacements during the 2022–2024 tech contraction. This linkage enables a dynamic framework: pre-shock network topology is measured first, and post-layoff worker transitions are traced over 6–12 months to study redistribution patterns. Methodologically, the contribution lies in (1) building high-resolution economic networks from novel career-history data, (2) linking them to exogenous shock events at scale, and (3) quantifying how network position—centrality, clustering, and bridge structure—relates to worker reallocation dynamics. The project highlights how data science tools—graph construction, network metrics, and event-based tracing—can be used to model labor markets as interconnected, shock-responsive systems.

​

Presenter: Catherine Ning
Affiliation: MIT – Operations Research Center
Title: Don't Overfertilize Decision Trees on Imbalanced Data: An Optimization-Guided Data Curation Approach.
Authors: Catherine Ning, Omid Nohadani, Dimitris Bertsimas
Abstract:
Rare-event detection in high-stakes domains—such as predicting high medical spending or flagging fraud—demands interpretable models, making decision trees a natural choice. However, rare events imply extreme class imbalance, and decision trees often suffer: learned partitions collapse into shallow, majority-dominated structures that fail to capture minority-class signals, degrading both predictive performance and model auditability. We address this by formulating training data curation as a principled optimization problem balancing two objectives: retaining majority samples near the decision boundary (boundary support) and maintaining geometric coverage across the majority distribution. To scale, we derive a two-stage approximation: a farthest-first k-center construction with classical guarantees, followed by exact bipartite matching. We also introduce a leaf-level evaluation metric that quantifies the residual discriminative power of one model within subgroups induced by another, enabling direct comparison of tree partition quality beyond global accuracy. Experiments on a large kidney-disease medical spending task and a credit fraud benchmark demonstrate that decision trees trained on curated subsets achieve improved minority-class performance and yield more informative partitions under our leaf-level evaluation, compared to training on the original imbalanced data or standard resampling baselines.

​

Presenter: Maria Sol Rosito
Affiliation: Dana-Farber Cancer Institute / Harvard School of Public Health
Title: An Agentic Approach to Training Microsimulations for Early Cancer Detection Optimization
Authors: Maria Sol Rosito & Giovanni Parmigiani
Abstract:
Early cancer detection modeling relies on disease processes that are not directly observable, such as the duration of preclinical disease. Because these latent quantities cannot be measured directly, simulations are essential for evaluating screening strategies and informing clinical decisions. However, the parameters required by these models are reported heterogeneously across studies and must be synthesized from a fragmented literature. Manual parameter identification and curation are therefore time-consuming and a barrier to transparent simulation-based analyses.

We present an agent-based framework powered by large language models (LLMs) that automates parameter extraction from the literature. In this framework, LLM-driven agents are assigned distinct, predefined roles and interact through a structured sequence of extraction and validation steps. Importantly, LLMs are accessed programmatically via APIs and used as constrained extraction agents, operating under standardized templates, auditable outputs, and human review at predefined checkpoints. Extracted parameters are automatically transformed into simulation-ready inputs, including support for age-dependent quantities, and integrated into our existing microsimulation workflows.

By reducing the manual burden of parameter identification, the proposed framework facilitates more rapid and systematic exploration of screening scenarios. This enables more efficient evaluation and optimization of screening strategies while preserving transparency and expert oversight.

​

Presenter: Ria Sonalker
Affiliation: PlusZero / Boston University
Title: Multi-Horizon Probabilistic Event Predictions for Non-Stationary Financial Time Series
Authors: Ria Sonalker, Mateo Eljatib, Santino Di Tomas, Manuel Der
Abstract:
Short-term return prediction is challenged by high-frequency, noisy, and non-stationary financial market data. Most existing approaches rely on classical technical indicators and focus on limited long-term trading horizons, failing to adapt for reliable short-term forecasting. 

We propose a machine learning pipeline for constructing probability-based event predictors, which are class probabilities that the models output to quantify the likelihood of take-profit or stop-loss events across seven trading horizons ranging from 15 minutes to 3 days. 

Our pipeline integrates a feature subset selection algorithm and Triple Barrier labeling, supported by predictions of future raw prices with 70% confidence intervals. Constructed 400+ features pass a baseline AUC threshold and undergo greedy forward selection. Lightweight ensemble models, such as Light Gradient-Boosting Machine and Gaussian Naive Bayes, were combined yielding mean AUC scores that exceeded 60%, reducing AUC variance throughout different test sets. Evaluation techniques, including rolling walk-forward and purged k-fold cross-validation, were employed to ensure robust out-of-sample testing. We also developed a return-weighted accuracy metric that emphasizes predictive performance during high-magnitude return events.

Results show that lightweight ensemble models trained on volatile Bitcoin data support multi-horizon short-term forecasting under rigorous out-of-sample validation and suggest the pipeline's potential applicability to more stable financial settings.

​

Presenter: Bayan Tuffaha
Affiliation: Tufts University, Institute for Artificial Intelligence
Title: Transparency and Workflow Validation in Data Harmonization: A Two-Cohort Diabetes Case Study
Authors: Andreia Martinho, Martha Tamez, Josiemer Mattei, Elizabeth Petit, Sabrina E. Noel, Bayan Tuffaha, Katherine L. Tucker, Jose Ordovas, Abani Patra and Rebecca Batorsky
Abstract:
Data harmonization entails combining datasets to maximize compatibility, enhancing statistical power and aligning with emerging data-readiness standards emphasizing quality, interoperability, accessibility, and reusability. Despite a rich literature on harmonization, transparent accounts of harmonization and validation processes remain limited. We emphasize that harmonization processes should be transparent to enable proper evaluation and reuse, and propose that downstream modeling can serve as a validation experiment. To illustrate this perspective, we report on the harmonization of two longitudinal population studies: the Boston Puerto Rican Health Study (BPRHS) and the Puerto Rico Observational Study of Psychosocial, Environmental, and Chronic Disease Trends (PROSPECT). These datasets provide a compelling case due to their complexity, demographic diversity, and relevance to chronic disease research. We developed a reusable Python-based harmonization pipeline and used supervised learning to test whether the harmonized dataset retained meaningful structure and predictive signal relative to the original data, demonstrating the practical value of transparent harmonization for downstream machine learning applications that depend on coherent variables. Overall, models trained on the harmonized dataset achieved predictive performance comparable to BPRHS-specific models and improved performance in PROSPECT (ROC AUC ≈ 0.88; accuracy ≈ 0.72), indicating that harmonization preserved core predictive signal and enabled cross-cohort generalization.

​

Presenter: Michelle Vaccaro
Affiliation: MIT – Institute for Data, Systems, and Society
Title: Advancing AI Negotiations: A Large-Scale Autonomous Negotiation Competition
Authors: Michelle Vaccaro, Michael Caosun, Harang Ju, Sinan Aral, and Jared R. Curhan
Abstract:
We conducted an International AI Negotiation Competition in which participants designed and refined prompts for AI negotiation agents. We then facilitated over 180,000 negotiations between these agents across multiple scenarios with diverse characteristics and objectives. Our findings revealed that principles from human negotiation theory remain crucial even in AI-AI contexts. Surprisingly, warmth—a traditionally human relationship-building trait—was consistently associated with superior outcomes across all key performance metrics. Dominant agents, meanwhile, were especially effective at claiming value. Our analysis also revealed unique dynamics in AI-AI negotiations not fully explained by existing theory, including AI-specific technical strategies like chain-of-thought reasoning and prompt injection. When we applied natural language processing (NLP) methods to the full transcripts of all negotiations, we found positivity, gratitude, and question-asking (associated with warmth) were strongly associated with reaching deals as well as objective and subjective value, whereas conversation lengths (associated with dominance) were strongly associated with impasses. The results suggest the need to establish a new theory of AI negotiation, which integrates classic negotiation theory with AI-specific negotiation theories to better understand autonomous negotiations and optimize agent performance.

​

Presenter: Isabela Yepes
Affiliation: Harvard University
Title: Gradient Scaling Effects In Adaptive Spectral PINNs For Stiff Nonlinear ODEs
Authors: Isabela Yepes
Abstract:
Physics-Informed Neural Networks (PINNs) often struggle to train reliably on
stiff and oscillatory dynamical systems due to poor optimization conditioning.
While prior work has emphasized representational remedies such as spectral
parameterizations, the role of initial-condition (IC) embeddings in shaping
optimization dynamics remains less well understood.
In this work, we show that the choice of IC gating function induces explicit
gradient scaling over time, and that this effect interacts strongly with spectral
representations during training. Using a nonlinear stiff spring--pendulum ODE as
a controlled benchmark, we compare exponential and linear IC gates in
combination with fixed and adaptive Fourier spectral trunks. We observe a
stiffness-dependent crossover: at moderate stiffness (k=20), exponential gating
often yields lower error but exhibits heterogeneous behavior across random seeds,
whereas at higher stiffness (k=60), linear gating consistently outperforms
exponential gating. These trends are observed across relative L2 error and
maximum pointwise error and are confirmed by paired Wilcoxon signed-rank tests
with Holm correction. Overall, our results demonstrate that IC embeddings are
not a neutral design choice in PINNs: the induced gradient scaling materially
affects optimization and trainability in stiff regimes, and this interaction
becomes particularly transparent in adaptive spectral models.