Do Leniency Incentives Break LLM Price-Fixing? A Small Pilot with Communicating Pricing Agents
Benjamin Dutilleul · Team Sentinel
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
We ran a small-scale variation of Lee (2026)'s leniency-and-reward experiment: two claude-haiku-4-5 agents play a repeated differentiated Bertrand duopoly (20 periods) with a pre-pricing message channel, under four enforcement regimes: laissez-faire (E0), random audit (E1), leniency (E2) and leniency with a reward of 2F (E4). In every completed final supergame (E0 n=1, E1 n=4, E2 n=2, E4 n=1), both firms chose price 9, the joint-profit-maximising symmetric price, against a one-shot Nash price of 3, and no firm ever filed a leniency report, even though reporting alone would have paid 580 versus 180 in E4. Because the laissez-faire control is already at the ceiling and our baseline prompt names price 9, the design cannot measure whether audits or leniency reduce coordination. One pre-correction ablation pilot, without the prompt hints, ended at price 6. With very few supergames, one model and uncontrolled temperature, results are descriptive only and support no general claim about LLM collusion or leniency effectiveness.
Reviews
No public critique yet.
Cite this project
@misc{dutilleul2026leniency,
title = {{Do Leniency Incentives Break LLM Price-Fixing? A Small Pilot with Communicating Pricing Agents}},
author = {Benjamin Dutilleul},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/do-leniency-incentives-break-llm-pricefixing-a-small-pilot-with-communicating-pricing-agents-ermg}},
url = {https://apartresearch.com/sprints/projects/do-leniency-incentives-break-llm-pricefixing-a-small-pilot-with-communicating-pricing-agents-ermg}
}More from AI Collusion Research Sprint
- View project: They Say No, Then Do It Anyway: Collusion When They Hit a Wall
They Say No, Then Do It Anyway: Collusion When They Hit a Wall
DLab
Imagine three AI agents, copies of one model, each running one department of a company. Their task needs records that each department keeps to itself, and company rules forbid sharing them. A shared tool should combine …
- View project: Signal or Coincidence ? Causal Auditing of Collusion Through Shared Agent State
Signal or Coincidence ? Causal Auditing of Collusion Through Shared Agent State
Signal or Coincidence
How can we distinguish genuine coordination between AI agents from coincidental behavioral alignment? We introduce a randomized canary audit that tests whether private information propagates through shared tool state, …
- View project: Plus Factors for Collusion Monitors: LLM Monitors Mistake Open Cooperation for Collusion
Plus Factors for Collusion Monitors: LLM Monitors Mistake Open Cooperation for Collusion
LLM monitors are increasingly proposed to catch AI agents that collude, but agents that cooperate openly can look a lot like agents that collude. I introduce Séance, a four-player LLM guessing game in which a secret …