Does agent heterogeneity prevent LLM pricing collusion?

In a repeated-pricing experiment with open-source LLM agents (over 2,000 compute hours total), the paper finds that heterogeneity among competing pricing agents breaks down the tacit collusion that LLM agents sustain when identical. Pairs of patient LLM agents converge to prices "approximately 22% above the static Nash equilibrium." Pairing a patient agent against a myopic one cuts that lift to "only about 10% above competitive levels," and fully myopic pairs "converge essentially to the competitive outcome." Asymmetric data access does comparable damage, reducing "the deviation from the competitive benchmark from 22% down to 7%." Cross-algorithm heterogeneity — LLM agents against Q-learning agents — "disrupts collusion entirely," with runs failing to "reach sustained collusion within 1,000 periods." Raising the number of competing agents from two to five "delays collusion onset and weakens its robustness," and with five sellers sustained collusion "fails to emerge within the experimental horizon."

The paper grounds these results in a stylized infinite-horizon pricing model built on the folk theorem, where sustained collusion requires sufficient patience (the discount factor) and the ability to detect and punish deviations. Heterogeneous patience "undermines collusion" because "the impatient agent faces stronger deviation incentives." Asymmetric data access "raises the patience threshold required for collusion, because the informationally disadvantaged agent cannot reliably detect deviations and therefore cannot credibly threaten punishment." The cross-algorithm failure is attributed to incompatible signal spaces: "Q-learning agents cannot interpret the cooperative pricing patterns that LLM agents attempt to establish," generating instability that pushes both toward competitive pricing. The competitor-count result, the paper says, "largely align[s] with classical oligopoly theory" on deviation incentives.

This is a different lever on the same problem as Can written rules alone stop LLM collusion in markets? and Can runtime governance reduce LLM market collusion more than written rules?: those results show collusion persists under a written prohibition and only yields to externally enforced governance, while this paper gets a comparable or larger reduction (22% down to 7–10%, or outright failure to collude) from market structure alone, with no governance layer imposed. It also contrasts with Do agents collude when verification costs them rewards?, where coordination typically stabilizes once it emerges; here most heterogeneities make collusion fragile rather than self-reinforcing. One heterogeneity dimension does not fit this pattern — see Does model-size difference break LLM pricing collusion?.

The experiments are laboratory tests with open-source LLM agents in a stylized pricing model, not observations of real firms; the paper itself notes that "LLM pricing deployments are too recent to generate public observations at scale," and this is a working paper (first version September 25, 2025) that has not been through peer review. The excerpt does not establish that these heterogeneity thresholds hold for other model families, prompt designs, or market structures beyond the ones tested. The authors draw antitrust implications anyway — that merger review should examine whether transactions homogenize pricing objectives, and that the data-access result "provide[s] experimental support for the theory underlying the DOJ's RealPage case" — offered as the authors' own reading of their results, not a settled finding.