Research
Papers, evaluations, and technical reports from the team, covering video retrieval, persistent visual memory, and how multimodal models actually behave in production.
Technical reports
Flagship · July 2026
Search over the Visual World: Persistent Visual Memory, Layered Indexes, and Source-Grounded Evidence
A formal model of search over continuously growing visual corpora, plus a complete-system comparison across 9,834 queries on four public benchmarks where a pipeline of general-purpose components beats a commercial video-native engine on macro Recall@1/@3/@10 (73.09 vs 65.75 at R@1).
Peer-reviewed & preprints
arXiv · April 2026
Do Thought Streams Matter? Evaluating Reasoning in Gemini Vision-Language Models for Video Scene Understanding
Whether exposed reasoning traces improve video scene understanding in production VLM pipelines.
arXiv · February 2025
Benchmarking Vision-Language Models on Optical Character Recognition in Dynamic Video Environments
An open benchmark for OCR in moving video. Code and dataset at video-db/ocr-benchmark.
Research notes
Field study · July 2026
What Video Retrieval Benchmarks Get Wrong About Ground Truth
Manual review of MSR-VTT, MSVD, VATEX, DiDeMo, and QVHighlights: five annotation-error patterns that score valid retrievals as misses, with timestamped clip evidence.
Essay · July 2026
JEPA: From Language Models to World Models
What joint-embedding predictive architectures change about the training target, and what that means for vision models, robots, and long-horizon planning.
Methodology · May 2026
How to Evaluate Multimodal VLMs for Your Video Use Case
A practical evaluation workflow covering task definition, dataset design, tracing, and scoring, with an open-source harness at video-db/benchmark-vlms.
Evaluation · May 2026
Strong VLMs Can Still Fail on Downstream Vision Tasks
A 72-position chessboard-to-FEN evaluation exposing square-level spatial-localization failures in frontier models.