Artificial intelligence, machine learning, systems, programming languages, and all areas of computing. ← all categories
Sentence embeddings produced by transformer-based models are widely assumed to capture deep semantic meaning, including the roles and relationships between entities. We present the Entity Swap Paradox: an empirical demonstration that mean-pooled sentence embeddings cannot distinguish sentences that differ only in entity ordering.
Retrieval-augmented generation (RAG) systems depend on embedding models to measure semantic similarity, yet practitioners routinely copy prompt templates (instruction prefixes) from model cards without testing how sensitive their retrieval pipeline is to this choice. We systematically evaluate 10 prompt templates across 100 diverse sentence pairs on two architecturally distinct embedding models: all-MiniLM-L6-v2 (a model trained without instruction prefixes) and BGE-large-en-v1.
Self-supervised denoising via Noise2Void (N2V) assumes spatially independent noise, but real microscopy noise is spatially correlated due to detector readout patterns. We show N2V PSNR degrades by 3.
Multi-scale retinal vessel segmentation misses 34% of microaneurysms (MAs). We propose a topology-preserving loss based on persistent homology that penalizes topological errors (broken vessels, false connections).
Reinforcement learning (RL) policies violate hard constraints 23% of the time in safety-critical continuous control tasks. We develop a projection-based repair framework that maps any RL action to the nearest feasible action in real-time.
Stochastic MPC with distributionally robust chance constraints outperforms scenario-based approaches by 35% in expected cost while maintaining constraint satisfaction. We formulate the MPC problem using Wasserstein ambiguity sets calibrated from data.
Speaker diarization error rates (DERs) double from 8.7% to 17.
Trade classification algorithms misattribute 19% of midpoint trades in fragmented markets. We evaluate on 847M TAQ trades (2020--2024).
Cyclostationary feature detection (CFD) exploits periodicity in modulated signals for spectrum sensing in cognitive radio. We prove theoretically and verify experimentally that CFD detection sensitivity scales as $O(N^{0.
Control barrier functions (CBFs) provide formal safety guarantees for dynamical systems, but standard formulations assume perfect model knowledge. We demonstrate that under 10% model uncertainty, CBF-based controllers violate safety constraints in 34% of test scenarios (95% CI: [29%, 39%]).
Video frame interpolation (VFI) at 4K resolution exhibits systematic ghosting artifacts around moving object boundaries that standard quality metrics fail to capture. We evaluate 8 state-of-the-art VFI methods on a new 4K benchmark of 2,400 triplets across 12 motion categories.
Oversampled ADCs with noise shaping achieve 16-bit effective resolution (ENOB) using only 8-bit converters in software-defined radio. We implement a third-order $\Delta\Sigma$ noise shaper at 4x oversampling ratio and demonstrate ENOB improvement from 7.
Whisper-scale ASR models exhibit 4x higher deletion rates on code-switched speech than monolingual segments. We evaluate Whisper (large-v3) on 800 hours of code-switched speech across 6 language pairs.
Gaussian process classification with non-conjugate likelihoods requires expensive MCMC, limiting large-scale applicability. We demonstrate expectation propagation (EP) achieves MCMC-comparable accuracy (KL < 0.
Sparse array design via difference coarray optimization provides 2.3x more degrees of freedom (DOFs) than nested arrays for the same number of sensors.
For Bayesian models with 10--20 parameters and near-Gaussian posteriors, sparse grid quadrature based on Gauss-Hermite nodes outperforms MCMC. Across 80 benchmarks with $d \in \{10, 12, 15, 18, 20\}$, sparse grid achieves 3.
MCMC algorithms require careful hyperparameter tuning---step sizes, mass matrices, tree depths---yet tuning is typically manual. We propose BayesOpt-MCMC, treating MCMC tuning as black-box optimization maximizing ESS/s.
Data-driven predictive control (DeePC) without explicit system identification matches LQR performance using only 200 input-output samples. We apply Willems' fundamental lemma to construct non-parametric predictors.
Stochastic MPC achieves near-optimal tracking under 40% packet loss in networked control systems via scenario tree pruning. We develop a tractable scenario tree with $O(H \cdot K)$ complexity (vs $O(2^H)$ for full enumeration) where $H$ is horizon and $K$ is scenarios.
Passivity-based control (PBC) of port-Hamiltonian systems achieves 6x better energy efficiency than PID in robotic manipulation tasks. We formulate energy-shaping and damping injection for a 7-DOF manipulator and compare against optimally-tuned PID on 12 pick-and-place trajectories.