PAL-Bench: Evidence-Grounded Profile Reconstruction from Longitudinal Personal Albums

A new benchmark called PAL-Bench aims to test how well AI systems can reconstruct personal profiles from long-term photo collections without accessing private ground-truth data, according to a paper posted to arXiv on June 15, 2026 [1]. Longitudinal personal albums function as weak-schema multimodal databases, where key facts require joining information across faces, text, timestamps, locations, and repeated events [2]. Existing visual, video, document, and lifelog benchmarks test only sub-problems, not album-scale profile reconstruction with social identity binding and evidence citation [2]. The PAL-Bench benchmark addresses this gap by operating under a public-record contract. Its Evidence Compiler builds latent private worlds, programs target-level evidence paths, renders album pixels, re-measures them through perception pipelines, and exports audited public and private views [2]. Agents receive only perception-derived public records; targets, identifier maps, and evidence paths remain hidden [2]. The benchmark contains 50 synthetic users, 36,659 public photo records, and 2,799 targets spanning owner facts, identities, and relations [1][2]. A privacy-preserving audit with 10 participants confirmed that PAL-Bench evidence structures match real private albums, though the authors note that equivalent releases remain privacy-prohibitive [1][2]. Across seven systems and two compute-matched diagnostics, a seven-metric protocol revealed a gap between plausible profile summarization and faithful social reconstruction [1][2]. Systems recovered some owner facts but struggled with recurring identities and evidence citation [2]. PAL-TRACE, a reference framework that freezes identity bindings before owner-fact mining, performed best but left hard identity resolution far from solved [2]. The paper positions PAL-Bench as a testbed for perceptual entity resolution, multimodal data integration, temporal evidence aggregation, and provenance-aware structured prediction [1][2]. The paper was submitted to arXiv, a preprint repository that accounts for roughly 95 percent of the paper URLs Hugging Face users have linked in their repositories organically [4]. Hugging Face and arXiv have collaborated to embed interactive demos directly alongside papers on arXiv abstract pages, allowing users to try community-built demos in a browser without writing code [5]. The PAL-Bench paper is listed under arXiv's Computer Science and Artificial Intelligence category [1].