Can Transparency In Pricing Policy Facilitate Collusion Among LLM Agents?

The work asks whether keeping pricing strategies secret is a deterrent for collusion. The results proposed are enough to indicate that this is in fact true and warrant further research. The modelling, experimental setup and experiments are clear, correct and well presented. The main limitation I would like to see addressed is the fact that agents’ policies are deterministic in nature, while, even in game-theoretical settings, stochasticity can have a big impact.

This is a focused, well-executed study with a genuinely fresh question. Can LLM pricing agents collude simply by reading each other's strategies, with no messages exchanged? Earlier work (Agrawal et al. 2025) showed that natural language messages increase collusion. This paper adds a second channel, policy transparency, and shows it raises prices about as much as talking. The gap is $4.7 versus $6.4 above the no channel baseline.

Two findings stand out. First, the collusion runs through matching rather than punishment. Sellers stop undercutting and copy a shared price, with almost no threats or trigger rules. That means detectors built to hunt for punishment schemes could miss this entirely.

Second, in the transparent policy arm there are no messages at all, yet prices stay high. A monitor that only reads messages would see nothing. The governance lesson is clear and a little counterintuitive. Strategy transparency between competitors should be treated as a possible collusion channel, not assumed to be a safeguard.

The design is clean. The prompt is identical across the four arms except for the clause describing the channel. Runs are paired by seed to cancel noise. Scripted controls bound the measure. The no channel price reproduces the parent paper's number to the cent. The presentation is a real strength. Honest limitations, a dual-use section, a full worked example, and the verbatim prompt. The work is reproducible from the PDF alone.

The main constraints are scope, and the author states them plainly. One stylized market, 8 seeds, two models from a single provider, and an open admission that the matching may partly be prompt-anchoring rather than strategic reasoning. The clean price effect also leans on passive, price-taking buyers. With strategic LLM buyers it nearly vanished and mostly cut trade volume instead. These bound how far the result travels, but they do not undermine it. A strong, honest contribution that opens a neglected question.

Cool conceptual angle bringing program equilibrium into LLM pricing. My main concern is an identification problem in the prompts. Because the visible pricing policies include explicit price numbers, it's really hard to tell if the models are actually doing strategic commitment/reciprocity, or if they're just anchoring on the numbers in the prompt. If you run an ablation where the policy logic is shared without literal price examples, your argument would be much more convincing.