Ahmad A. Rushdi — Projects

LLM Evaluation: Uncertainty & Trust

Recent work stress-testing large language and vision-language models — whether they genuinely reason about numerical functions or merely recall, and whether a model's stated confidence stays warranted as its perceptual evidence degrades.

Many agents, how many independent witnesses? When several LLM agents agree, the agreement is read as corroboration — as if each were an independent measurement. Agents built from similar models make correlated mistakes, and debate makes them more alike, so a five-agent panel is worth barely more than one independent witness. Only mixing models from different developers buys real independence.

Five ways to build a five-agent panel, from the assumption (five independent votes) to round 0 to one round of debate. Edge darkness is the agents' error correlation ρ_c; the meter is the effective number of independent agents, N_eff = N / (1 + (N−1)ρ_c). MMLU-Pro, measured values. Paper at the AIMS Workshop, COLM 2026.

Reasoning or recall? A blind protocol hides the identity of a function and asks LLMs to predict, integrate, and optimize it from sampled points alone — separating genuine numerical reasoning from memorized recall.

blind numerical-function protocol — Blind protocol across three tasks (multifidelity, integration, optimization): sampled points → LLM → predicted function, integral, or optimum.

recall vs reasoning spectrum — Responses span a spectrum — recall, code, genuine reasoning, or no capability — by error metric and sample budget. paper

Confidence as testimony. As perceptual evidence decays under noise and occlusion, a faithful model should lose confidence — yet some confabulate a confident, wrong class instead.

VLM class and confidence across six images under increasing noise and occlusion — Six items × seven defeater levels (noise + occlusion): the model's class (*italic* = wrong) and verbalized confidence. Faithful items abstain; others confabulate (cup→“tambourine”, clock→“mirror”). paper

Robot Learning: Decision Making under Uncertainty

Does uncertainty-aware machinery around a generative robot policy actually help it succeed? Testing world-model selection and runtime failure monitors over whole episodes, not single decisions.

Best-of-K world-model selection. A diffusion policy proposes K candidate actions, a world model predicts each outcome, and a scoring rule executes the best. Across five world models and K from 5 to 100, selection that improves single decisions can lower success over whole episodes — on Push-T, from 56.7% to as low as 32.5%. paper

One Push-T episode, decision by decision: the policy proposes K candidates (gray), the world model picks one (purple), the robot executes it.

One episode, six ways to choose: the policy alone and the policy with each of five world models, from the same start and seed. Barcodes show each model's 20 candidates sorted by true score and shaded by predicted score. Paper at the UncertainWM Workshop, CoRL 2026.

Beyond AND and OR. Runtime monitors join several failure scores with a fixed rule — alarm if any fires, or only if all do. Calibrating the scores jointly instead learns one no-alarm region in score space that keeps the false-alarm guarantee; a diagonal region catches more out-of-distribution failures on Push-T (0.79 vs. 0.60 at the same alarm rate). With L. Lützow and M. J. Kochenderfer, UncertainWM Workshop, CoRL 2026.

AI Ethics & Governance

Where regulation meets the pace of AI — measuring how governance actually works, and where it stalls.

Bridging AI development and regulation. Only 4.23% of U.S. AI bills (2017–2025) reach a terminal outcome; a position paper argues for adaptive, anticipatory legislation with independent oversight. position paper

bill-flow funnel — Flow of 1,419 state AI bills (2017–2025) from introduction through committee, floor vote, and final outcome. paper

High-d sampling & visualization

A library of high-dimensional test functions for optimization, uncertainty quantification, and numerical integration problems.

Model evaluation & UQ for AI/ML

Quantifying uncertainty in AI/ML predictive estimates, decomposing it into aleatoric and epistemic components, and using it to guide active learning. Applications across vision, image classification, regression, and time-series.

Vision. Uncertainty in object detection, segmentation, and image generation/captioning workflows.

DALL·E 3 generation — Text-to-image generation with DALL·E 3 (prompt: "A nature reserve with giraffe, elephant, zebra, lion, faceted style.")

YOLO detection — Object detection with YOLOv3 (COCO).

SAM segmentation — Image segmentation using Meta's Segment Anything (SAM).

BLIP captioning — Image captioning with BLIP (via Hugging Face inference API).

Image classification. Accuracy & uncertainty impact of rotation and AWGN on NN classifiers (MNIST / CIFAR).

MNIST under rotation.

MNIST under additive Gaussian noise.

1-d regression. Uncertainty bands with deep ensembles, MC dropout, and conformal prediction over limited noisy observations.

Uncertainty bands across methods.

Multifidelity fusion. Convolutional encoder/decoder networks that fuse many cheap low-fidelity samples with a few expensive high-fidelity ones to reconstruct full fields and their predictive uncertainty.

multifidelity field predictions — True vs. predicted pressure fields with error and uncertainty maps (Poiseuille flow), reconstructed mostly from low-fidelity data. paper

Voronoi Piecewise Surrogate (VPS) models

Leveraging Voronoi diagrams and Delaunay graphs for global surrogate modeling in high-dimensional UQ and adaptive sampling under limited sample budgets (multifidelity / costly / high-stakes simulations).

VPS neighbors — Finding significant Voronoi neighbors in high dimension. paper

VPS surrogate — A global high-d surrogate stitching local patches in Voronoi cells. paper

gradients — Gradient samples for local active subspaces.

POF darts — Adaptive sampling for probability-of-failure estimation. paper

Meshing & mesh tuning

Creating and tuning 2-d / 3-d meshes with guaranteed quality properties.

VoroCrust: unclipped Voronoi mesh conforming to a high-quality surface mesh. paper

All-quad mesh — Robust all-quad meshing of domains with connected regions, with angle & edge-length guarantees. paper

Fourier & signal processing in bio / bioinformatics

Feature extraction (e.g., period-3 in DNA), object localization (e.g., echoes in ultrasound), and spectral analysis of high-d sampling techniques via Nd-FFT.

persistence spectrum — Persistence spectrum: a narrowband interference signal embedded in broadband.

Digital notch filtering to detect wideband ultrasound contrast echoes in blood.

Spoke-darts spectrum — Fourier spectrum of high-d sample patterns. paper

From computational methods to engineering applications

Extending mesh-tuning methods (e.g., non-obtuse triangulation, paper) to model fiber-reinforced polymers for elastic & failure simulations; and ML for collision detection and motion planning.

fiber simulation — Cross-section of a fiber micrograph and model at peak load; simulated tensile response of different packings. paper

Collision detection phase model.

Projects

⚖︎LLM Evaluation: Uncertainty & Trust

⚙︎Robot Learning: Decision Making under Uncertainty

§AI Ethics & Governance

✦High-d sampling & visualization

σModel evaluation & UQ for AI/ML

⬡Voronoi Piecewise Surrogate (VPS) models

△Meshing & mesh tuning

∿Fourier & signal processing in bio / bioinformatics

◈From computational methods to engineering applications