Guiding Worker Self-Selection in Crowdsourcing Contests: An LLM-Augmented Algorithmic Approach
TL;DR
GRAF is introduced, a greedy algorithmic framework that produces self-selection outcomes in polynomial time by ordering workers according to a score vector, with provable guarantees—zero worker regret and platform optimality—for special cases of SSTC, and LLMScore is proposed, an LLM-driven evolutionary framework that automatically designs the scoring algorithm GRAF uses.
Abstract
Crowdsourcing platforms coordinate large pools of online workers who compete in contests to produce solutions for clients, from logo design to machine-learning challenges. A defining feature of these platforms is that workers self-select: Each chooses which contests to enter and how much effort to invest, and these choices are strategic rather than aligned with what the platform needs. As a result, the platform may end up with too few participants, or too little effort, on the contests where high-quality solutions matter most—while individual workers may end up dissatisfied, regretting choices that left them worse off than an alternative they could have taken. To reconcile these competing interests, platforms increasingly offer workers recommendations on where to compete. We study how to generate such recommendations as the game of self-selection in Tullock contests (SSTC), a two-stage model in which workers first choose contests and then compete within them. We introduce GRAF, a greedy algorithmic framework that produces self-selection outcomes in polynomial time by ordering workers according to a score vector, with provable guarantees—zero worker regret and platform optimality—for special cases of SSTC. Because a good ordering is difficult to design by hand once workers are heterogeneous, we propose LLMScore, an LLM-driven evolutionary framework that automatically designs the scoring algorithm GRAF uses. LLMScore addresses two obstacles to applying existing LLM-based methods here: jointly optimizing solution quality and worker satisfaction, and the intractability of evaluating worker regret. It is trained only on small instances of a single setting yet transfers to larger and structurally different settings, and—because its output is human-readable code—a platform operator can inspect and adjust the recommendation logic it learns. Across 1,000 synthetic instances spanning four settings, GRAF augmented by LLMScore consistently yields high-quality or even near-optimal outcomes with low worker regret, benefiting platform and workers alike.
