reconstruction

This paper demonstrates that reconstruction-based tests for activation explanations can be gamed, producing high scores while specific claims remain false. It proposes RECAP, which trains linear heads alongside the target model to keep designated internal content reliably decodable and independently verifiable against probes, improving safety auditing.

Introduces a new open-source toolkit, img2threejs, that converts photos into Three.js models and automatically compares the rendering result with the original image to calibrate reconstruction quality, especially suitable for hard-surface objects.

This paper introduces Lorentz Encoding (LE), a physics-informed framework that uses implicit neural representations and physical constraints to reconstruct high-resolution CEST MRI from sparsely sampled data, achieving superior performance over existing methods.

The author strongly criticizes the concept of an infinite canvas, arguing that the essence of AI video is horizontally arranged film strips, and announces a complete refactoring from first principles before release.

This paper investigates the impact of inference stride (disjoint vs. overlapping windows) on reconstruction-based time series anomaly detection, finding that overlapping windows consistently improve performance by up to 28% across multiple model architectures on the TSB-AD and UCR benchmarks.

MRAgent introduces a novel graph-based memory framework that dynamically reconstructs memory during reasoning, achieving up to 23% improvement on long-horizon benchmarks while reducing computational costs.

This paper presents a comprehensive taxonomy of 3D vision research, covering geometric representations, datasets, learning paradigms, and applications in reconstruction, generation, and video modeling.

This paper identifies and explains a spectral bias in reconstruction-based EEG foundation models, where embeddings over-represent aperiodic and low-frequency components while under-representing oscillatory components, especially at higher frequencies, leading to poor performance in low-resource settings.

This paper proposes CoAD, a novel framework that unifies Outlier Exposure (classification) and Masked Autoencoder (reconstruction) paradigms for time series anomaly detection, addressing their respective limitations. Extensive experiments show that CoAD significantly outperforms state-of-the-art methods while being lightweight and fast.

ShapeCodeBench is a synthetic benchmark for perception-to-program reconstruction where models generate executable drawing programs from raster images, evaluated on metrics like exact match and pixel accuracy. The benchmark is designed to be renewable via seeded RNG, and current models still achieve low exact match rates, indicating room for improvement.