Employer Learning, Sorting, and Productivity: Evidence from Computer Scientists

How does employer learning affect the allocation of talent and aggregate productivity? I study this question in the labor market for computer science (CS) Ph.D.s, using the job histories and post-Ph.D. publications of 31,000 graduates from 2000 to 2021. Publishing a CS conference paper raises the probability that a researcher moves to a top tech firm in the following year, especially for workers who start at less productive firms in industry. The increase in upward mobility upon publication is larger for less experienced workers and for scarcer signals, such as first authorship or papers with a matched patent. These patterns are consistent with predictions from an equilibrium search model with public employer learning. Estimating the model, I find that learning from post-Ph.D. publications accounts for 14% of overall CS publications, as it reallocates high-ability researchers to employers that provide more opportunities to publish. This contribution exceeds that of initial sorting, underscoring that much of worker ability is revealed on the job rather than at entry.