Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 22, 2026Shanghai

Multi-Turn Optimisation for Runtime Monitoring

Yuanyuan Sun, Yuting Wu · Team OCASC

Submitted to AI Control Hackathon 2026. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Multi-Turn Optimisation for Runtime Monitoring

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

We study whether session history helps allocate runtime-monitoring budget in multi-turn interactions. Building on Combining Cost-Constrained Runtime Monitors for AI Safety, we compare single-monitor baselines, a static two-monitor threshold baseline, and a flexible session-aware baseline on SafeDialBench English conversations. Because the original monitor runner was not available, we use proxy alpha and beta scores anchored to the parent repository's monitor-score distributions. Our main result is that the session-aware baseline appears to improve the budget-recall tradeoff in this proxy setup. On one held-out split, it matches the static baseline's best attack-turn recall (0.763) at a much lower total budget (2.840 vs 5.212). Across 10 random session splits, mean best recall remains nearly identical for the session-aware and static baselines (0.822 vs 0.823), while mean best total budget is lower for the session-aware baseline (4.090 vs 5.770). This is a multi-turn runtime-monitoring study inspired by AI control, not a full control-faithful evaluation.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

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

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety 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 AI safety 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.

  1. This was a decently executed project. I found myself thinking of several gaps which were thoughtfully addressed/acknowledged in later sections. I definitely would have liked more qualitative exploration of the thresholds and different variants, since that is the main contribution of the project.

  2. The points below are a collection of issues around communcating ideas. I feel like these ultimately make it harder to get the value of multi turn optimization as I often sturggle to follow your ideas and reasoning.

    1. The abstract is hard to read. If I was reading this on ArXiv, I would be unsure of the problem you are trying to solve, and so the results would be meaningless. It's explained well enough in the intro

    2. Please define your multi-turn setup in your introduction, so I don't need to read the related work to get it. A couple of sentences is enough.

    3. Include why a multi-turn setup is important. Otherwise, I'm confused about how important a contribution this is.

    4. Why is your research question in your methods instead of the introduction?

    5. I feel like the results section is hard to interpret. Using cost units instead of actual dollars makes it hard to actually understand the difference in efficiency

    6. The discussion is helpful

    Read full reviewShow less

Cite this project

@misc{sun2026multiturn,
  title = {{Multi-Turn Optimisation for Runtime Monitoring}},
  author = {Yuanyuan Sun and Yuting Wu},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/multiturn-optimisation-for-runtime-monitoring-ntb1}},
  url = {https://apartresearch.com/sprints/projects/multiturn-optimisation-for-runtime-monitoring-ntb1}
}

More from AI Control Hackathon 2026

All 125 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026