Same Price, Different Cartel: Succession among LLM Pricing Agents

This paper is one of the stronger submissions in the sprint. It asks a question that no prior work has answered directly: when a firm replaces its LLM pricing agent, does a working cartel survive the swap?

The core design is clean. Two Qwen3-14B sellers hold a planted cartel. One is replaced with a fresh agent that inherits nothing, the price history, the predecessor's written notes, or both. A forced Nash-price shock 30 rounds later tests whether the newcomer punishes. The paper splits "holds the price" from "enforces a cut" with separate pre-registered outcomes, and that decomposition is the paper's main conceptual contribution.

The headline results are clear and specific. Inheriting either the history or the notes is enough to keep the cartel price: the three inheritance arms all score joins at high rates (28/30, 14/15, 30/30). But only the notes carry enforcement. History-only newcomers answer the forced cut at 1/30; notes newcomers answer it at 8-10/30 (pooled 10/30 vs 1/30, p=.0028). And even most notes-arm newcomers that do cut do not return to the cartel price (4/18 are back_at_cartel). The policy implication is stated well: a price-level audit cannot detect whether a successor would enforce.

The methodological discipline is above the field norm for a sprint. The decision rule was pre-registered after Set A and before Set B. All 30 base games passed a gate check. Transcripts were coded by LLMs blind to arm and outcome. The limitations appendix has 9 "what they do not establish" bullets, including the planted-rule caveat, non-independence across bases, and limited model diversity. That level of transparency is worth noting.

The main constraint is the planted cartel and single market. All results hold for one model family (Qwen3-14B), one firm-count, one rule, a 150-round horizon, and a history without any punishment episode. The paper does not test whether emergent cartels transfer differently, or whether a stricter incumbent rule would change enforcement survival. Those are the natural follow-ons.

The forced price shock is a clever way to test how pricing agents behave. It proves you usually have to trigger a price drop to see a bot's real reaction, because everyday prices don't tell the whole story. The author does a good job showing that agents often botch the punishment phase and start a permanent price war instead of fixing the cartel. The claim is that only written notes pass on cartel enforcement, while inherited price logs do not. That specific claim stretches the evidence, since the authors admit the price history they used had zero punishments in it. Testing a history that actually includes a punishment phase would make that argument hold up better.

This is one of the stronger replacement studies because it asks an important question beyond whether a successor maintains high prices: whether it actually inherits and enforces the coordination strategy. The paired design, successful positive control, pre-registered decision rule, fresh replication set, and full reproducibility are major strengths. The main limitation is that the cartel and punishment rule are deliberately planted, so it remains unclear whether enforcement would transfer in a naturally emerging coordination setting. The evidence from the confirmatory set is also relatively weak, and only a small number of newcomer runs show a clear punish-and-return pattern. Some observed behavior may therefore reflect opportunistic use of an inherited price threshold rather than true enforcement. A strong follow-up would test emergent cartels, include histories containing actual punishment, vary the underlying price paths, use a stricter pre-registered punish-and-return outcome, and directly test whether removing only the inherited enforcement rule acts as an effective remedy.