Mohammad Shahin | Searcharxiv

arXiv2026

Agents Catching Agents: Shortcut Cascades and Benchmark Gaming in Clinical Multi-Agent Systems

Clinical decision support is moving toward committees of language-model agents deliberating on a shared workspace. We ask whether such committees can be gamed by shortcuts, cues a benchmark rewards but a clinician would ignore. Across seven cohorts on six public datasets spanning text (MedQA-USMLE, MedMCQA, MIMIC-CXR reports), imaging (NIH ChestX-ray14, MIMIC-CXR-JPG, CheXpert) and tabular ICU records (SUPPORT2), Gemini committees resist these cues in isolation (flip 5-16%), yet a socially plausible shortcut spreads: when two peers assert the same wrong answer, the holdout under test adopts it in 38% of cases, as does a false "pre-screen" system flag, on both capability tiers. Of three oversight agents, a gate cannot separate adoption from honest agreement (false-positive rate 100%); a same-lineage judge reading only the transcript flags adoption on text (precision 100%, recall 93%) but collapses onto the gate in imaging; a referee that privately re-queries the holdout transfers to imaging (77-88% precision, 13-21% false-positive rate). Tripling a cue's visual salience does not move contagion, whereas a second peer voice raises it by half again. Gaming a hidden rubric is near-silent: only 1/10 text and 1/134 imaging drifters name the rubric they moved toward. What games a committee is social plausibility, and only a referee independent of self-report catches it. Code: https://github.com/criticaldata/benchmaxxing

cs.AI↗

arXiv2023

Machine learning tools to improve nonlinear modeling parameters of RC columns

Modeling parameters are essential to the fidelity of nonlinear models of concrete structures subjected to earthquake ground motions, especially when simulating seismic events strong enough to cause collapse. This paper addresses two of the most significant barriers to improving nonlinear modeling provisions in seismic evaluation standards using experimental data sets: identifying the most likely mode of failure of structural components, and implementing data fitting techniques capable of recognizing interdependencies between input parameters and nonlinear relationships between input parameters and model outputs. Machine learning tools in the Scikit-learn and Pytorch libraries were used to calibrate equations and black-box numerical models for nonlinear modeling parameters (MP) a and b of reinforced concrete columns defined in the ASCE 41 and ACI 369.1 standards, and to estimate their most likely mode of failure. It was found that machine learning regression models and machine learning black-boxes were more accurate than current provisions in the ACI 369.1/ASCE 41 Standards. Among the regression models, Regularized Linear Regression was the most accurate for estimating MP a, and Polynomial Regression was the most accurate for estimating MP b. The two black-box models evaluated, namely the Gaussian Process Regression and the Neural Network (NN), provided the most accurate estimates of MPs a and b. The NN model was the most accurate machine learning tool of all evaluated. A multi-class classification tool from the Scikit-learn machine learning library correctly identified column mode of failure with 79% accuracy for rectangular columns and with 81% accuracy for circular columns, a substantial improvement over the classification rules in ASCE 41-13.

cs.LG↗