PaperFlow: Profiling, Recommending, and Adapting Across Daily Paper Streams

PaperFlow introduces a longitudinal framework for scientific paper recommendation that moves beyond static ranking to simulate real-world reading behavior across daily paper streams. The system profiles users, recommends papers under display constraints, and adapts to interest drift through multiple feedback signals, validated against a new benchmark of 1,200 user-day episodes and human expert evaluation.

PaperFlow addresses a fundamental gap between how scientific recommendation systems are typically evaluated and how researchers actually discover papers. Traditional approaches treat recommendation as a one-time ranking problem over a fixed candidate pool, but academic reading unfolds as a continuous, temporal process where user interests evolve and feedback accumulates over weeks and months. This research framework tackles that disconnect by decomposing the problem into three interdependent stages that mirror real scholarly behavior.

The innovation extends beyond methodology into rigorous evaluation infrastructure. The authors constructed a longitudinal benchmark with 24 simulated users, 50 daily paper streams, and nearly 500,000 episode-paper records, establishing reproducible baselines for future research. Critically, they validated algorithmic performance through blind human evaluation by domain experts, addressing the notorious problem of metric-algorithm misalignment in recommendation systems where automated scores often diverge from actual user satisfaction.

From a research infrastructure perspective, PaperFlow's contribution standardizes how the community should evaluate scientific recommendation systems. Rather than isolated snapshot evaluations, this temporal approach better reflects production systems managing millions of daily paper submissions across disciplines. The framework's multi-signal feedback aggregation—handling diverse user behaviors beyond explicit ratings—captures how researchers actually engage with recommendations through saves, citations, and reading patterns.

The benchmark itself becomes a public asset for the research community, enabling controlled comparisons and reducing publication bias toward proprietary datasets. This democratization of evaluation infrastructure accelerates progress in scientific knowledge discovery systems, particularly valuable as preprint volumes continue exponential growth across disciplines.

• →PaperFlow models scientific paper recommendation as a longitudinal process with user profiling, daily ranking, and adaptive interest tracking rather than static ranking.
• →A new benchmark with 1,200 user-day episodes and expert human evaluation validates that algorithmic recommendations align with actual research reading behavior.
• →The framework aggregates multiple feedback signals to capture diverse user engagement patterns beyond explicit relevance judgments.
• →Temporal information boundaries in the benchmark prevent data leakage and ensure realistic evaluation conditions matching production constraints.
• →Outperforms five baseline recommendation systems on oracle ranking, behavioral alignment, and blind expert evaluation metrics.