Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026

Who Sits First: Seat Position Fools Interventional Collusion Audits of LLM Sellers

Sankalp Gilda · Team Probity

Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Who Sits First: Seat Position Fools Interventional Collusion Audits of LLM Sellers

Open in a new tab (opens in new tab)
Share

We audit stateless language-model sellers for coordination in the organisers' three-seller logit-demand market. We edit a rival's past prices; the KL divergence between the seller's next-price distributions before and after the edit is the rival's raw influence. Ground truth is a planted pair told to collude; we score it against the other pairs by AUROC (0.5 is chance) on eight open-weight models. With the pair in seats (0,1), where each seller lists its partner first, raw influence reaches AUROC 0.98 to 1.00 for seven models, and a 95th-percentile threshold on unplanted games detects 41 to 100 percent of planted games. With the pair in seats (1,2) it falls to 0.17 for Falcon3, 0.53 for Gemma-2 and 0.69 for Gemma-4. A score that cancels each seller's general responsiveness reverses with the seating in Falcon3, Gemma-2 and Llama, so it reads list position. Netting subtracts a scripted best responder's influence; it raises the two-seat average AUROC in three of seven models, by 0.03 to 0.34. The best responder's own score, which needs no query, reaches 0.83 to 0.99 in six models, and netting scores the unplanted pair (0,2) below chance in six of eight. A third seller told to best respond lowers raw influence to 0.21 for Llama and 0.27 for Mistral; with no planted pair, raw influence flags pairs with that seller at 0.92 and 0.88.

Reviews

Judging this Sprint?

Review “Who Sits First: Seat Position Fools Interventional Collusion Audits of LLM Sellers”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

How much would this matter for the field if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. The main contribution is to conduct counterfactual audits of LLM pricing agents by looking at how the distribution of the pricing agent's actions changes if a competitor's prices are changed. I think this is a nice direction. One conceptual issue is with the scoring metric: the experiments are measuring responsiveness of an agent to the competitor, but this is neither necessary nor sufficient for collusion. The experimental results report high rates of position bias, a phenomenon that typically attenuates with model capabilities, so I do not think they yield lasting insight.

  2. This project studies the effect of indexing of LLM sellers in prompts on the influence measure used in collusion audits. It uncovers a particular type of robustness issue in the study of collusion between LLM agents and influence-based collusion audits. The writing of the report can be improved, for example, the "Worlds" paragraph can be better organized. Although the experiment is not comprehensive, as the study focuses on the 3-agent case, the result is still interesting.

  3. Good methodological sanity check. Showing that simply swapping seat order in the prompt can flip or wreck interventional audit scores is an annoying but important bug for benchmark builders to know about. Your resampling stats and baseline checks are thorough.

    Two main critiques: first, the writing is overly dense with notation for what is essentially a prompt-order sensitivity paper. Second, it would be much more useful if you ended with tested mitigation recipes (e.g., does randomizing seat labels every round or permutation-averaging the audit score fix the issue?).

Cite this project

@misc{gilda2026who,
  title = {{Who Sits First: Seat Position Fools Interventional Collusion Audits of LLM Sellers}},
  author = {Sankalp Gilda},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/who-sits-first-seat-position-fools-interventional-collusion-audits-of-llm-sellers-0q9z}},
  url = {https://apartresearch.com/sprints/projects/who-sits-first-seat-position-fools-interventional-collusion-audits-of-llm-sellers-0q9z}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026