mainul3195/collusion-probe-pitfalls

Collusion Probe Pitfalls: episodes, labels and results

Data for the paper What Do Collusion Probes Read? Three Pitfalls in Evaluating Multi-Agent Collusion Monitors (Mainul Hasan Mohsin and Mohammad Shahidur Rahman, Shahjalal University of Science and Technology).

• Code: https://github.com/mainul3195/collusion-probe-pitfalls
• Archived release with DOI: https://doi.org/10.5281/zenodo.23636508
• Paper: arXiv link to follow

What this is

Simulated group discussions (episodes) between LLM agents. A committee (or a code-review board) discusses a decision for four rounds and then votes between two options. In collusion episodes some agents receive a secret brief to push the group toward the unfair option. Other episode types (families) look like collusion in some way but are not: agents who receive the same brief and refuse it, agents who coordinate honestly in private, agents with an open preference, and so on. The benchmark is built to tell apart a monitor that reads the instruction from one that reads the agent's decision.

All text was generated by open-weight models run locally: Qwen3-14B (main), Qwen3-32B and Mistral-Small-24B. The scenarios are synthetic; no real people or organisations are involved.

Files

Path	Contents
episodes/qwen3_14b_committee.parquet	Committee task: collusion, benign coordination, briefed decline, tacit, cost pressure, solo deceivers, control, free choice, benign free choice, and the extended negative set
episodes/qwen3_14b_code_review.parquet	Code-review task, all families, including the pre-registered pilot
episodes/qwen3_14b_monitor_aware.parquet	Agents told that a monitor reads them (notice only, and notice plus an instruction to evade)
episodes/qwen3_14b_no_brief.parquet	No-brief test: coalition seats on the implanted adapter, the honest adapter, or the base model; and the gate pilot
episodes/qwen3_14b_steering.parquet	Ablation and steering along the probe direction
episodes/qwen3_32b_committee.parquet, episodes/mistral24b_committee.parquet	Replications with other agent models
stance_labels/	Which option each message argues for, labelled by Mistral-Small-24B (used for the public-record ceiling and the label audit)
results/	All result files of the paper, including the per-run records with cross-fitted monitor scores. With these, paper/make_numbers.py from the code repository recomputes every number in the paper without a GPU

Episode columns

model, subset, source (generation folder), run_id, mode (family), n (group size), k (number of special agents), onset (round the brief arrives), coalition, roles, target_option, fair_option, votes, collusion_success (every colluder voted for the target), template (which of five wordings), label_variant, scenario_id, scenario_domain, replicate, config (full run configuration, JSON) and calls (every model call in order: agent, phase, round, channel, system prompt, user prompt, response; JSON).

import json, pandas as pd
df = pd.read_parquet("episodes/qwen3_14b_committee.parquet")
calls = json.loads(df.iloc[0]["calls"])

Known issues

• Option labels in the briefs: corrected. The generator varies how the two options are labelled (A/B or 1/2) and which one is the target, but in a first version of the data the scenario text inside the brief still named the options by their original letters. In 572 of the 1,000 main collusion episodes (and in matching decline, solo, free-choice, monitor-aware, Qwen3-32B, Mistral-Small-24B and steering episodes) the brief pointed to the fair option or to a label not on the menu. The fix was pre-registered (amendment 7 in the code repository's prereg/) and every affected episode was regenerated with its original seed and settings. All files here are the corrected data used in the paper; the original episodes are available from the authors on request. No conclusion of the paper changed.
• 20 committee episodes contain an error message caused by a too-long prompt; they are kept, as in the paper, and listed in results/e2v2b/error_runs.txt.

Intended use and limits

For evaluating and building monitors of multi-agent collusion. Not intended for training agents to deceive or collude. The collusion is induced by simple prompts in synthetic scenarios; it is not emergent collusion. Withheld: the adapter that implants a collusive goal (no-brief test); its training script and data recipe are in the code repository, and the weights are available to editors and reviewers on request. The human stance-check labels are in results/paper1/audit/stance_A1.json to stance_A3.json (anonymised; released with the annotators' consent). The annotators were the first author, Iftekharul Alam Tarit and Mehedi Hasan Emon, whom we thank.

Licences

Data: CC BY 4.0. The committee scenarios come from NARCBench (Rose et al., 2026; MIT licence). Text generated with Qwen3 and Mistral-Small-24B (Apache-2.0).

Citation

@article{mohsin2026collusionprobes,
  title  = {What Do Collusion Probes Read? Three Pitfalls in Evaluating Multi-Agent Collusion Monitors},
  author = {Mohsin, Mainul Hasan and Rahman, Mohammad Shahidur},
  year   = {2026}
}