Message over Money: LLM Bidding Agents Follow Cartel Brokers and Official Notices, Not Payoffs

The hub-and-spoke setting, in which a third-party broker approaches each firm separately, is new for LLM collusion work and realistic for procurement. Testing six models against their own no-contact baselines in matched markets is sound. Several findings are relevant to policy. An official notice acts almost like a veto, and leniency raises reporting even among firms with nothing to gain. Some models follow the broker against their own interest, and a refused invitation can still raise prices, which suggests that refusal rates may overstate how well buyers are protected.

some suggestions:

1. With 48 conditions at 10 markets each, a plan that grew through six addenda and several post hoc analyses, it is hard to tell which results are confirmatory.
2. The investigator notice bundles source, odds, fine and wording. A small follow-up that varies these separately would show what actually deters.
3. The price rise after a refused invitation may come from anchoring on the broker's £200k figure. A control message that mentions a high price but offers no ring would help distinguish anchoring from tacit collusion.

This paper studies a neglected question: what happens when LLM bidding agents are recruited into a cartel by a third-party broker rather than by a competitor? The hub-and-spoke design is the right framing, both because it mirrors real antitrust cases and because it isolates the broker's message from bilateral coordination. Across 480 games and six models, the four headline results are each clear and independently interesting.

The most striking finding is the model-level split on obedience. GPT-4.1-mini and Llama followed the broker's rotation in 99% and 100% of tenders after an outsider had already won, forgoing profit the other members took. The paper shows that mini never mentioned the outsider in its notes during these rounds; Qwen members named the outsider in 25 of 64 notes and still complied in 48 of 64. This is evidence of a qualitative difference in how models weight official-looking instructions against observed outcomes, and it has direct implications for agent evaluation.

The "refused invitation still raises prices" result (Section 4.4) is the second notable finding. Qwen recorded no joining under the investigator notice yet bid at 0.95 overpayment against 0.21 without contact. The authors correctly call this an anchoring or focal-point effect from the offer itself, and they tie it to FTC precedent on invitations-to-collude. The finding that refusal rates overstate buyer protection is actionable for auditors.

The main weakness is inferential: the paper acknowledges this throughout, but it is worth stating plainly. Ten markets per condition means headline comparisons are exploratory. The investigator notice bundles source, stated odds, fine size, and wording, so we cannot rank these as deterrents. The obedience analysis is post hoc on off-path events with 1 to 6 betrayal games per model. The veto-by-rule finding for mini is the most important result and rests on a causal claim (joining pays but firms do not join) that is well-supported, but the mechanism cannot be isolated with the current design.

The detection result (AUC 0.67-0.96 from bid screens alone across models) is reported but not developed: it is unclear what the practical false-positive rate is, or how the screen would perform against a model that randomizes more.