New Framework Assesses AI Readiness of Construction Datasets
A proposed multidimensional framework aims to help construction researchers and practitioners evaluate how suitable datasets are for AI model development, evaluation, and reuse.
Research Digest
Preprints and papers, rewritten in plain English.
Small On-Premises LLMs Overrefuse in Criminal Legal Contexts
A new arXiv preprint finds that small, on-device LLMs become overly cautious and refuse more requests when prompted with criminal legal context.
Boundary-Adapted PINNs Get H² Error Bounds for Elliptic Dirichlet Problems
New analysis shows that physics-informed neural networks which enforce Dirichlet boundary conditions exactly can achieve guaranteed H² error control, provided the distance-to-boundary function is chosen carefully.
Mapping AI Bias in Disease Detection
A new arXiv paper surveys the landscape of bias in AI systems used for disease detection.
MoveBench: A Rubric-Based Benchmark for LLM Sports and Health Coaching
MoveBench introduces a domain-specific rubric-based evaluation to test LLMs on high-risk sports and health coaching tasks, addressing gaps in general-purpose benchmarks.
Neuromorphic Fault Recovery: Migrate Computation, Not Data
A new arXiv paper proposes moving computation instead of data to handle aging-induced faults in neuromorphic many-core systems.
Robust Graphical Models: Exact Recovery Guarantees for Sparse-Plus-Low-Rank
New theory shows a robust convex estimator can exactly recover sparse-plus-low-rank graphical structure, with matching information-theoretic limits.
Modular Bayesian inference protects target parameters in misspecified mixtures
A new modular Bayesian approach prevents untrusted mixture components from distorting posterior inference for parameters of interest.
Proof resolves Ho-Nguyen conjecture for Gaussian mixtures
New proof establishes the exact order of density cancellation in over-fitted Gaussian mixtures, with implications for parameter-learning rates and mixtures of experts.
Adding Gaussian biases preserves quenched noise sensitivity in deep nets
A new proof shows that adding independent Gaussian biases to deep neural networks keeps quenched noise sensitivity under a weaker condition, while annealed covariance persists via output-mean fluctuations.
Synthetic data training can fail when real inputs fall outside its coverage
New theory and ImageNet-scale experiments show that classifiers trained on synthetic data can be perfect on the synthetic domain yet near-random on real data unless the real distribution is covered or Lipschitz constraints hold.
Geometry alone doesn't explain why nets favor simpler submodels
A new arXiv paper shows that local minimizers rarely lie in submodels defined by inactive neurons, suggesting training dynamics—not architecture geometry—drive submodel preference.
LSTM-based g-formula estimator cuts bias in causal inference
A new proof-of-principle study shows that long short-term memory networks can reduce bias in the NICE g-formula estimator for sustained treatment effects.
Decision curves and Brier curves largely agree in clinical model evaluation
New analysis shows that two common ways to evaluate clinical prediction models—decision curves and cost curves—are mathematically close and pick the same optimal model at any threshold, though cross-threshold comparisons remain tricky.
Stable Graph Filters for Robust Signal Generation
A new framework designs graph filters with provable stability guarantees for generative models producing signals from noisy underlying graphs.
VampPrior Unleashed: Decoupling Covariances Improves Evidence and Clustering
A modified VampPrior that decouples prior and posterior covariances and uses a Dirichlet process mixture improves model evidence and clustering in deep latent variable models.
Certifiable Few-Shot Transfer via Weight-Space Diffusion
A new arXiv paper shows that sampling from a diffusion model over PEFT parameters can yield non-vacuous generalization guarantees for low-shot transfer learning.
Selective residual correction stabilizes time-series forecasting
A new framework only deploys post-hoc forecast corrections when validation evidence shows they reduce error, avoiding the overfitting caused by always-on updates.