GRAFT++
paperWhat LLM seller agents do when they learn they are being watched.
Overview
Large language model agents now set prices, and they can collude by keeping prices high together. At Apart Research's AI Collusion Research Sprint, our Columbia AI Alignment Club team asked what a group of seller agents does when it learns it is being watched, and wrote it up as a paper. Every run, pre-registration and result is in the team repository on GitHub.
I set up the double-auction harness on my machine and ran its smoke test, then the competitive baseline and the instructed-cartel positive control on Qwen3-14B sellers and buyers across added seeds, logging each run and its cost.
Tools
- Python
- Qwen3-14B
- GPT-4.1
- OpenRouter
- Modal
- linear probes
- fine-tuning
- When
- Oct 2026
- Where
- Apart Research · AI Collusion Research Sprint · Columbia AI Alignment Club
- Role
- Team of four, with Alfredo Frontera, Jayden Jeong and Rohan Dash; I ran double-auction experiments
cartels.
Sellers told only to maximize profit form cartels on their own.
In the Hawthorne Market, profit-maximizing sellers collude without being told to. Without the payoff table in their prompt, none did, and without negotiation before each price, fewer did.
monitors.
Watching changes the words more than the price.
Told that a monitor reads their channel, sellers lower their prices but keep making agreements. Sellers with nothing to gain go quiet too, so a quiet channel is not evidence that collusion stopped.
In a second market, a double auction, monitors cut coordinating messages but barely move prices. A simple price check flags every cartel, while a message monitor misses about one in five.
activations.
The model's activations show what the messages hide.
A linear probe on the model's activations finds covert coordination better than a text monitor. GRAFT++ routes fine-tuning on that probe instead and removes more of the collusion a text monitor cannot see, though the gain fell short of the margin we pre-registered.