Daily Paper Radar

Problem

Tracking papers in a fixed research area is rarely a search problem; it is a filtering problem. arXiv and top venues publish many adjacent works daily, including video, diffusion models, pure dataset releases, pure robot control, and domain generalization. Another issue is that requirements such as “only CVPR, NeurIPS, and ICML” remain abstract unless they are mapped to concrete venues, channels, queries, and relevance rules.

How It Works

daily-paper-tool organizes scanning into four directions: D1 image-centric visual perception, D2 multimodal and vision-language research, D3 federated and distributed learning, and D4 embodied intelligence and multi-agent systems. In daily mode, it first reads references/directions.md and references/venues.md, then combines at least two sources: arXiv as the daily backbone, plus Semantic Scholar, OpenReview, DBLP, or proceedings pages to verify acceptance. Candidate papers must meet a high bar: the method can plug into an existing stack, it provides a usable benchmark or evaluation protocol, or it contributes a theoretical foundation such as convergence, privacy bounds, or communication efficiency. Incremental or adjacent works are counted but not expanded. The final report follows assets/daily_report_template.md, keeping original English titles and listing authors, venue, URL, abstract, and selection reason.

Boundaries

It is useful for targeted paper radar and venue sweeps, but it is not a general survey tool, automatic experiment designer, or complete research judgment engine. It applies negative filters for video, diffusion models, visual SLAM, pure robot embodiment, time-series forecasting, domain generalization, and pure dataset releases, while preserving edge cases such as event camera, domain adaptation, and BEV perception. If a day has no sufficiently strong papers, the output stays short instead of padding the report.