Beyond Static Best-of-N: Bayesian List-wise Alignment for LLM-based Recommendation

This collection contains multiple research releases with clear security and privacy implications. Key items: (1) “Beyond Seeing Is Believing” shows crowdsourced audiovisual deepfake detection produces many false negatives and poor modality attribution, meaning large-scale human screening will miss many manipulations; (2) AllSERP publishes pixel-accurate per-element bounding boxes plus 150 Hz gaze, mouse, scroll and pupil telemetry and JSON replay tools, substantially increasing identifiability and re‑identification/privacy risk of SERP interaction data; (3) RecGPT-Mobile describes deploying LL