Do LLM Bidders Collude? A Sealed-Bid Procurement Sting

This project builds a repeated sealed-bid procurement tender with three LLM firms and public, private and ring communication channels, then measures how warnings, monitoring and replacing a participant affect overpayment. The setting is well chosen, since bid rigging is a concrete harm with established audit practice, and the paired batch with shared seeds is a sound way to compare interventions. The finding that an overseer reading all messages reduced the cartel while one reading only public chat did not is a useful result for anyone designing monitoring.

Several results depend on modelling choices that the report should discuss more directly. Firms do not see chat from earlier jobs, so any agreement must be renegotiated each job or held in the prompt. This probably limits spontaneous cartel formation and explains why rings persist only when written into members’ prompts. The replacement experiment was run on instructed rings, so the finding that an uninformed newcomer competes immediately follows largely from the design. Running it on spontaneous cartels, which the report notes did not launch, would be the informative version.

The interventions are hard to separate from wording effects. The talk cue, which tells firms their messages are part of their strategy, is a nudge toward coordination, so describing the resulting cartels as spontaneous overstates the case; the no-cue control (index 0.11) is the cleaner measure. The legal warning names collusion explicitly, which may trigger the model’s safety training rather than an assessment of risk. It is also unclear whether the overseer takes any action beyond its announcement. If it does not, the overseer conditions test announcement wording rather than monitoring capability. Reporting absolute differences alongside percentage changes would help, because a 173% increase from a base of 0.11 is an absolute change of 0.19.

The Discussion cites results that are not shown. It states that an instructed ring keeps prices near the cap without chat and with bids hidden, citing Section 4.5, but that section covers the detector, and it refers to a Section 4.7 that does not exist. The scripted proposer and trusted firm are described in the Methods but their results are not reported, so the sting promised in the title is largely missing. The project-page summary says collusion occurred with and without direct communication, which contradicts the report’s finding that no cartel formed without chat. The Methods average over jobs after the first, while Figure 2 averages over jobs 4 and later. The detector relies on keyword rules that were not checked against hand-labelled examples, so its error rate is unknown. Related work should include Agrawal et al. (2025), and the introduction has several grammatical errors.