Answers are grounded in this paper and cite their sources.
We study tacit bid suppression by learning bidders, and run a partly pre-registered comparison of the first-price, second-price and Dutch formats, with and without disclosure of the winning bid, for two Q-learning or language model (LLM) bidders that share a common known value. All LLM results are for one model, claude-haiku-5-5 at low effort, and for 100 rounds unless stated. With a hint that stresses that paying less raises profit, a first-price auction that hides the rival's bid holds the revenue flat at 0.51 of the value, disclosure lifts it to 0.89, and second-price reaches 0.97, a gain over disclosure that is not established, and zero on the proportional grid. This hidden-bid level is a suppressed outcome, and at most 0.33% of the agents' notes in any cell of the main arm (and at most 0.94% in any claude-haiku-5-5 cell we ran) match a list of coordination words, which we did not check against hand labels. There, a fixed bid would raise a bidder's profit by 0.2344 of the value per round, compared to 0.2445 realised. In a registered probe on a finer grid, where the same market sits at 0.57 of the value, a forced one-round deviation of 0.10 of the value reaches the rival only as one extra loss when it is upward (in 4 of 12 runs; in the other 8 the rival sees nothing) or as one extra win when it is downward (in 8 of 12 runs), and in both cases the rival's mean bid in the next five rounds moves by less than 0.02 of the value. We detect no enforcement, and call the outcome bid suppression without coordination reasoning, not collusion. This flat level is specific to that hint and to the cent grid: on the finer grid of the probe, the same hint gives 0.57 and is still rising (slope +0.00195 [0.00142, 0.00248] per round). When the hint is dropped or replaced by one that stresses winning, the hidden-bid revenue is about 0.2 of the value higher at round 100 and still rising. In a registered 300-round run without a hint (8 of 12 runs so far), it climbs from 0.70 in rounds 71 to 100 to 0.88 [0.80, 0.96] in rounds 271 to 300, still rising and still below the competitive level. We ran no disclosed cell without a hint, so for A0 the effect of hiding the rival's bid is not measured; for A1 and A2, showing it lifts the revenue at round 100 by 0.38 and 0.24 of the value (the latter across batches). The flat level near half of the value appears only with the hint that stresses paying less, a contrast across batches. For Q-learning bidders with optimistic initialisation, a hindsight update from the winning bid lifts first-price revenue about as much as the payment rule, and with one round of memory, second-price revenue falls below its benchmark in every run, although both bidders visit every state-bid pair at least 100 times. Version 1 reports an interim read of the 300-round no-hint cell (8 of 12 runs); a later version will report all 12. Version 2: extended runs to 300 rounds, a second model, and corrected claims after review.
No takes yet. Share an insight, caveat, or question.
Ryan Steubs (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: