Erchin Serpedin

Retaining even a single structural relative slashes post-deletion retraining loss by $8\times$, revealing that target accuracy after unlearning is often driven by intrinsic dataset redundancy rather than unlearning failure.

A model-free benchmark reveals that vision-language models often misread rather than misreason, exposing a performance gap that a supervised vision model can surpass.

Culturally grounded multimodal evaluation reveals significant gaps in current AI systems' understanding of Arabic contexts, with implications for both accuracy and representation.

Integrating prompt-conditioned channel attention can boost segmentation accuracy by over 23% in challenging medical imaging tasks.

A single design choice—whether to include parity labels—can determine if a model predicts physically impossible outcomes, with staggering accuracy implications.

Natural-language descriptions can outperform traditional image statistics in predicting pulmonary nodule malignancy, enabling a calibrated LLM to triage cases effectively.

Detection, not repair, is the critical bottleneck in verifying chemical reasoning with large language models, revealing a 22% to 4% error reduction through targeted verification.

Achieving an order-of-magnitude improvement in force direction prediction, CliffordSTF redefines the capabilities of interatomic potentials.

Segmentation can be achieved directly from 4D Gaussian representations without the need for costly external masks, revealing that intrinsic scene cues hold significant object structure information.

Unsupervised learning of physical dynamics from video gets a much-needed dose of reality with IRIS, a benchmark dataset of 220 real-world videos, complete with ground truth and a standardized evaluation protocol.

Don't chase higher accuracy in scientific reasoning tasks; strategically abstaining when evidence is weak yields far greater reliability gains.