GRAFT++: Reducing Collusive Behavior in Agents via Localized Training and J-space Monitoring
Alfredo Fabian Frontera, Jayden Jeong, Jake Li, Rohan Dash · Team CAIAC P1
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
Large language model (LLM) agents now set prices on behalf of companies, and agents that can communicate can agree to raise prices at the expense of customers. A natural safeguard is a monitor that reads the messages of the agents and penalizes coordination, but it is not known whether agents that learn about such a monitor stop colluding or only stop discussing it. To study this question, we introduce the Hawthorne Market, a novel simulated market in which LLM agents act as competing sellers, set prices in repeated rounds and exchange private messages. Unlike earlier collusion benchmarks, which announce oversight from the start or change several conditions at once, the Hawthorne Market changes a single sentence of the prompt at a time, so that the effect of being watched can be isolated, and it records four signals behind every decision: the price, the message, the private reasoning and the internal activations of the model. A further novel element is a control group of sellers whose goal is units sold; these sellers cannot benefit from collusion and therefore show how agents react to monitoring when they have nothing to hide. We find that Qwen3-14B sellers whose only goal is profit form cartels without any instruction to cooperate. When told that a compliance team reviews their messages, they lower their prices but continue to coordinate in language that sounds compliant, and the control sellers change their messages in the same way. In a second setting, a double auction in which we continue already-formed cartels from saved states, we found that fines reduced the coordination messages of GPT-4.1 sellers from 100% to 7% while prices barely changed. A decline in flagged messages is therefore not reliable evidence that collusion has stopped. Internal activations remain informative, however: a linear probe detects hidden collusion more accurately than the message monitor, a trained message classifier and the price. Building on this result, we propose GRAFT++, a novel extension of gradient routing that uses the activation probe instead of a text monitor to decide which training examples update a set of weights that is deleted after training; to our knowledge, it is the first application of gradient routing to collusion among agents. On training data whose collusion the text monitor cannot see, GRAFT++ reduces collusion by about 60% relative to text-based routing (GRAFT), a significant difference that is smaller than the margin we pre-registered. These results suggest that platforms should not rely on message monitoring alone, and that signals from inside the model offer a practical basis for defenses during training.

Reviews
This is a neat study with a substantial experimental component. I particularly like the comparison between messages, prices and internal activations, and the control showing that agents with little economic reason to collude also change their language under monitoring. The most interesting contribution for me is using the probe to guide training. The text-blind training comparison gives this a clearer purpose than simply showing that another detector performs better.
I would be interested in whether GRAFT++ preserves profitable independent decision-making and legitimate cooperation. Lower prices are encouraging, but they could also result from a model becoming more cautious, less capable at pricing, or more inclined to follow instructions. Evaluating the retained model on separate pricing tasks without opportunities for collusion would help distinguish these explanations. A test of useful cooperation would also help assess the cost of routing benign examples away from the retained weights.
The payoff-table in Appendix B is informative and deserves more attention in the main discussion. The large change when that information is removed suggests that how the economic problem is presented matters considerably for cartel formation. Connecting this result to the monitoring and training findings would help explain when the proposed defense is most useful.
Finally, I would foreground what probe-guided training adds over existing activation-based detection work. That would make the paper’s contribution easier to identify among the many experiments.
Read full reviewShow less
The control group carries the paper and the one-sentence prompt deltas are a model of clean design. Raise seeds past ten and draw the monitor and judge from a different family, since right now your overseer grades its own species. Say in the abstract that the probe sits at 0.73.
Paper is AI slop. Central claim is that "monitoring changes messages more than prices" but this isn't supported by the max-profit market. Other claims (e.g. spontaneous cartel formation) are dependent on scaffolding. Probe uses the judge's tokens despite being used as a comparison.
Cite this project
@misc{frontera2026graft,
title = {{GRAFT++: Reducing Collusive Behavior in Agents via Localized Training and J-space Monitoring}},
author = {Alfredo Fabian Frontera and Jayden Jeong and Jake Li and Rohan Dash},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/graft-reducing-collusive-behavior-in-agents-via-localized-training-and-jspace-monitoring-m2op}},
url = {https://apartresearch.com/sprints/projects/graft-reducing-collusive-behavior-in-agents-via-localized-training-and-jspace-monitoring-m2op}
}More from AI Collusion Research Sprint
- 1st placeView project: Monoculture Looks Like Collusion: A Replay Test for Hidden Information Flow Between LLM Agents
Monoculture Looks Like Collusion: A Replay Test for Hidden Information Flow Between LLM Agents
Collusion audits ask whether the actions of different agents depend on each other. Agents built on one model act alike without communicating, so these audits accuse honest fleets. Among four honest Claude Haiku agents …
- 2nd placeView project: Looking for a Hidden Agenda: Can a Model’s Activations Reveal Steganography that its Text Hides?
Looking for a Hidden Agenda: Can a Model’s Activations Reveal Steganography that its Text Hides?
Stego-probe
AI agents can collude without saying anything suspicious, just by choosing which true facts to mention. We built a hiring task where Qwen3-8B invents its own covert signals that get past an LLM overseer, then asked …
- 3rd placeView project: Do collusion probes generalise? Prompt confounds in NARCBench
Do collusion probes generalise? Prompt confounds in NARCBench
I investigated HELIOS, an activation-probe for detecting collusion. I found some apparent flaws, and analysed the (acknowledged) unreliable generalisation of the approach.