Vatsal Jha

Patient world models are usually evaluated by predictive fidelity, but their clinical value depends on the decisions their beliefs support under finite sensing, human attention and intervention capacity. We introduce AdaptiveCareArena, a real-trajectory replay benchmark that scores patient models by the resource-constrained decisions they support, instantiated on two public cohorts (CrossCheck, CES). Same-day review triage cuts missed high-severity reports relative to a neutral cyclic baseline at every tested budget, but ranking on each patient’s last report recovers nearly all of the belief’s advantage, so same-day triage mostly exercises the harness. A forward-looking variant finds within-person information in the moving-average belief on CES and a small decision-level gain over last report (+0.03 to +0.04 episode recall; intervals include zero). For measurement, one adaptive item-selection policy shows no practical advantage over a fixed short form, and the oracle gap is an upper bound inflated by selection. A tuned Kalman filter improves one-step accuracy by 6%, while 88 of 90 policy–budget–metric comparisons retain their direction and statistical resolution. We also name observation-action entanglement, show that gating the belief on review more than doubles missed events at the smallest budget, and show that an id-ordered cyclic baseline roughly doubles apparent attention disparity.

Cement production emitted about 2.4 Gt CO2 in 2024, and specifying lower-carbon concrete is a near-term procurement lever. Yet model-based material search can use predicted embodied carbon twice: first to rank candidates, then to quantify the selected product’s saving. We test that reuse on 108,326 ready-mix concrete Environmental Product Declarations (EPDs) from 413 mostly North American manufacturers, holding out whole manufacturers. Ranking retains useful signal under supplier shift while valuation degrades: mean within-class pairwise accuracy is 0.591, but mean log global warming potential (GWP) R² is 0.082. Selected mixes average 10.4% lower declared GWP than their class median, yet their declared GWP averages 9.1% (about 31 kgCO2e/m³ at the corpus median) above the value reported for them. We derive a conditional identity linking this claim gap to search depth; across pools of 2–50 products the gap is linear in depth (R² ≥ 0.97 in all three pool compositions). Changing the selection rule buys little, at a cost in realized advantage; replacing the point claim with a selection-conditional conformal floor keeps the chosen product and cuts overstatement above 0.05 log units from 51% to 11%; a simple class-quantile shift does comparably. The implied saving floor is positive in only 16.5% of cases, however, and the bound misses its 10% exceedance target on 15.2% of unseen-supplier cases. The result motivates a rank-then-report architecture for carbon recommenders.

Connected vehicles differ not only in what they observe but also in which targets they can supervise: a phone-grade client may expose sparse GPS, whereas an OBD-equipped vehicle can support a telemetry-derived powertrain-state label. We introduce BehaviorFL-MT, a federated multi-task formulation that predicts destination region and a proxy powertrain-state composite (a thresholded score over RPM, engine load, and MAF variability, not a diagnostic fault code) from an early recorded-sample prefix. Two explicit masks separate missing inputs from missing targets, structured OBD dropout exposes the state head to equipped and GPS-only views, and discovery-aware participation increases rare-task exposure without prior client-label metadata. None of these ingredients is new on its own; the contribution is their combination for partially supervised vehicle federations and a controlled study of which one drives the result. On 102,843 VED, GeoLife, and T-Drive trips with client-disjoint evaluation, only VED supplies state labels (188 of 5,617 training clients), and discovery-aware participation raises natural/GPS-only state AUPRC from 0.512/0.452 (best uniform-participation federated baseline) to 0.613/0.552 over three seeds. Ablations attribute this gain to participation rather than to the masking components. An oracle with complete eligibility metadata from round 1 reaches 0.613/0.550, so online discovery recovers essentially all of the observed oracle gain despite discovering only 114–129 of the 188 eligible clients by round 60. Destination HR@5 stays at 0.748 versus 0.751, a fixed geometry prior already reaches 0.749, and macro-F1 drops from 0.339 to 0.325. The study isolates how partial task supervision, unequal sensors, and cohort composition interact in federated behavioral learning.

Internet of Things (IoT) devices in home environments present unique security challenges that traditional authentication methods fail to address effectively. This paper presents Context-Aware Adaptive Multi-Factor Authentication (CA-AMFA), an innovative edge-based security system that dynamically adjusts authentication demands based on real-time risk analysis. The proposed approach uses an integrated risk model and leverages three contextual bandit algorithms—Upper Confidence Bound (UCB), Thompson Sampling, and Epsilon-Greedy—to dynamically update risk factor weights through continuous learning mechanisms. The hardware implementation on Raspberry Pi demonstrates the system’s relevance to resource-constrained environments. Through rigorous evaluation across a variety of scenarios, it is demonstrated that the bandit-based approaches significantly outperform fixed-weight schemes, lowering false acceptance rates by up to 44.75% and lowering unnecessary authentication friction by up to 20.18%, while maintaining strong security guarantees. The UCB algorithm demonstrated the best performance, achieving perfect accuracy in high-risk scenarios where traditional approaches failed. The open-source deployment and reproducible evaluation platform make CA-AMFA a computationally efficient solution that well balances security and usability for next-generation IoT authentication systems.