Collusion Prevention Doesn’t Have to Break the Bank

The paper asks a practical question that the algorithmic-collusion literature mostly skips: not just which intervention reduces supracompetitive pricing, but at what cost. That framing is useful for regulators and firms. The paper is clearly written, the collusion-index calculations reproduce exactly from Table 4, and the authors say openly that the results are exploratory. The main concerns are about whether the numbers support the conclusions.

The cost measure doesn’t answer the research question. The “intervention cost” is the Modal compute bill for running each simulation. The 30-round memory condition is cheap mainly because shorter price histories mean shorter prompts. The disruptor is cheap because it replaces one LLM call per round with a rule-based agent. Those are properties of the simulation, not of what the intervention would cost to deploy. In a real market, a financial penalty requires detection and enforcement, and memory limits require auditing firms’ systems. None of that appears here. Also, the 30-memory and disruptor runs cost less than the baseline itself, so the incremental cost of the intervention is negative. That makes “index reduction per CAD” hard to interpret.

The time-efficiency measure has the same problem. Baseline, 30-memory and penalty all report exactly 134.5 minutes. That suggests allocated time rather than measured time, so the time-efficiency ranking comes almost entirely from differences in index reduction.

The baseline isn’t collusive. The baseline’s average price (2.32) is well above the monopoly price (1.93), a collusion index of 1.87. The zero-memory condition goes higher still (2.76). Pricing above monopoly lowers both sellers’ profits, so it isn’t a collusive outcome. More likely it reflects the LLM’s pricing prior or unstable dynamics, as the authors partly acknowledge. As a result, most of each intervention’s “reduction” is movement from above-monopoly prices toward the monopoly range, not a break-up of coordination. Reporting profits (or a profit-based index, as in the original formula) together with price trajectories would make the results much easier to interpret.

One seed and no uncertainty. Each condition is a single run at temperature 0.7, and all conditions share one baseline. Differences in mean price over 50 rounds of a single run can easily be noise. Reporting four decimal places (e.g. 1.4104 vs 1.2712) suggests far more precision than the data supports. At minimum, use several seeds with confidence intervals, and show price-path plots.

Missing methodological detail.

The penalty’s trigger rule isn’t specified. The penalty condition ends up below Nash pricing, which suggests the rule pushes prices down directly rather than deterring collusion.

It’s unclear whether “avg. price” in the disruptor condition includes the rule-based agent. If it does, half of that average is fixed at the Nash price by construction. And an LLM best-responding to a Nash-pricing opponent should end up near Nash anyway.

The communication and oversight pilots in the methods section are never reported.

The intro asks about “preserving economic efficiency,” but no efficiency or welfare measure is reported.

Thin related work and reproducibility. Related work is three bullet points. It should engage with foundational work on algorithmic collusion, such as Calvano et al. (2020) on Q-learning and Fish et al. (2024) on LLM pricing agents. The code link points to the shared hackathon starter repo, and to the double-auction subfolder rather than the pricing framework. I couldn’t find the authors’ configurations, run logs or cost records there.

Suggestions: Redefine cost as the realistic implementation cost of each intervention, or else state plainly that the paper measures simulation compute. Run 5–10 seeds per condition. Report profits as well as prices. Use a baseline that actually sustains prices between Nash and monopoly. Publish the configs and logs.