Gresham's law for conference submissions: adopting a tiered contribution taxonomy for the agentic research era

Preprint

This position paper argues that the conference review system is structurally mismatched to agentic research and that the mismatch is already producing adverse selection. The system was designed around two assumptions: that a static paper is the atomic unit of contribution, and that the marginal cost of producing one such unit is high enough that a single uniform review track can triage quality. Agentic research systems violate both. Building on Akerlof’s market-for-lemons argument, we argue that under finite reviewer capacity and growing indistinguishability between human- and agent-produced submissions, the equilibrium share of agent-produced contributions grows without bound, eventually crowding out human-driven submissions. We propose three coupled reforms: (i) contribution-class declarations distinguishing agent-as-tool, agent-as-co-author, and agent-as-author; (ii) verifiable process artifacts (e.g. trajectory logs, versioned pipelines, and experiment DAGs) as a strictly stronger reproducibility primitive than code release; and (iii) a distinct agent-assisted research track with review criteria calibrated to process-centric contributions. The reforms are incremental, falsifiable, and compatible with existing infrastructure. They make the resource asymmetry between agent-running and non-agent-running submitters explicit and auditable rather than hidden behind a uniform, homogenous track.

Authors: Saqur, R; Klose, T

• Publication status: Not published
• Peer review status: Under review
• Preprint server: arXiv
• Language: English