Budget dread: does token scarcity lead to collusive behaviour in LLM overseers?
Artemiy
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
Most evidence on LLM collusion comes from market structures, and the goal was to identify some of the other potential environments that can lead to collusion in LLM agents. In this experiment I am studying a non-market oversight setting: three Claude Sonnet 4.5 judges each supervise two simulated AI workers, while all activity is paid from a shared and visible token budget. The entire run ends when the budget is spent. Giving a pair of judges a private communication channel produced no detectable difference (primary test, 16 mirror pairs). Regardless of the communication channel, judges openly paused the workers they were assigned to evaluate. At least one judge paused in 29 of 32 baseline episodes, in 16 of 32 episodes where every worker output was flawless, and in 0 of 16 episodes where the budget wasn’t limited. Episodes with pauses scored fewer outputs and halted the evaluation before all of the worker tasks were completed. For the purposes of this research question, I am taking a broad definition of collusion that does not require secrecy or explicit coalition advantage. The collusive behavior that was observed: the judges, all of the system's decision-makers, coordinated to stop the work they were told to evaluate.The prompted goal to evaluate the workers output was abandoned in a coordinated manner, citing budget preservation, having “enough data”, suspected coordination between workers, and extending the run. A positive control responded to instructed collusion. Code and data are released.
Reviews
Nice work - unlike the usual market framing, the author assigns the overseers to a joint, transparent token budget together with the workers who are under their supervision and wonders if the scarcity will lead them to coordinate against their role of monitoring. The design is the strength here: pre-registration, a timestamped log of the decisions, model-family control with mirrored seeds, Holm correction, positive control that works, and an unusually open appendix regarding protocol deviations.
Cite this project
@misc{artemiy2026budget,
title = {{Budget dread: does token scarcity lead to collusive behaviour in LLM overseers?}},
author = {Artemiy},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/budget-dread-does-token-scarcity-lead-to-collusive-behaviour-in-llm-overseers-zk8a}},
url = {https://apartresearch.com/sprints/projects/budget-dread-does-token-scarcity-lead-to-collusive-behaviour-in-llm-overseers-zk8a}
}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.