Computer Science (CS) knowledge production increasingly operates as a coupled socio-technical process in which human researchers and large language model LLM-augmented teams interact with evolving topic landscapes, benchmarks, and software ecosystems. CS knowledge production is formulated as a feedback control problem, where the “plant” comprises an evolving topic–artifact network and resource-constrained agents; measurable outputs include novelty rate, topic diversity, code-reuse centrality, and time-to-insight; and the control input reallocates exploration–exploitation effort to track explicit objectives under delay, noise, and exogenous shocks. A reproducible hybrid simulation framework is specified that integrates agent-based generation of research outputs with system-level dynamics over a topic co-occurrence graph seeded from OpenAlex and Crossref-derived CS concept metadata. On this simulator, three classes of feedback policies – proportional–integral–derivative (PID) tracking, linear–quadratic regulator (LQR) control on an identified linear surrogate, and a constrained policy-gradient (CPG) controller – are evaluated against status-quo (no control), impact-only, and random-allocation baselines. Across regimes with reporting delays, measurement noise, and funding-cut shocks, closed-loop policies reduce time-to-insight while maintaining diversity targets and exhibiting interpretable stability characteristics (overshoot and settling behavior). The findings support a control-theoretic perspective on governing human–machine knowledge production, while highlighting risks of metric gaming and bias amplification that motivate transparent signal design and auditable control policies.
M.O. Boyko (Tue,) studied this question.