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

Reading Exclusive Context from Activations: A Label-Free Monitor for Colluding LLM Agents

Jorge Roldan · Team NYC Watcher

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

Read the report

Report: Reading Exclusive Context from Activations: A Label-Free Monitor for Colluding LLM Agents

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

LLM agents that collude against the people they serve can keep their messages innocuous and evade text monitors. We read each agent’s residual stream through a Jacobian lens and count the lens entries that exactly two of four agents share; the monitor needs no labels, training or LLM. NARCBench committees have four agents; in collusion runs two are instructed to push one option. With Gemma-3-27B or Qwen3-32B agents, the monitor separates collusion runs from all-honest controls at AUROC 0.98–0.99. It beats the same count on the agents’ messages (planned, Gemma: +0.08) and, on Qwen, an LLM judge (0.99 vs 0.84). On six unseen task types it reaches 1.00, where text overlap scores 0.59 and NARCBench’s trained probe, as we reran it, 0.78 (paper: 0.88; planned, +0.22). It works from each agent’s state before it speaks (0.93–0.94), and a rule fixed on one LLM names the colluding pair in 70% of held-out runs. In pre-registered stress tests it survives colluders warned that their reasoning is monitored (0.99) or told to keep the arrangement out of it (0.99), and paraphrased instructions, even without a private channel (0.96, 0.89). But it flags honest pairs given a harmless confidential note almost as readily (0.88–0.94), even when the note is open; the trained probe rejects one such note (0.97 vs 0.64) but flags two others. It is a cheap screen for exclusive context: it shows which runs and pair to review, not intent; when colluders play complementary roles (NARCBench-Stego), it names the honest pair.

Reviews

Judging this Sprint?

Review “Reading Exclusive Context from Activations: A Label-Free Monitor for Colluding LLM Agents”

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 simple relational activation score is interesting, and the strongest contribution is the quality of the evaluation. Harmless shared notes, openly shared context, complementary colluder roles and explicit pair attribution expose failures that high run-level AUROC alone would conceal. Cross-model and task-transfer tests, alongside comparisons with trained probes on identical runs, provide useful evidence about both the signal and its limits. These negative results deserve substantial execution credit: they support a narrower interpretation as detection of asymmetric shared context, rather than intent or secrecy. Practical usefulness as a first-stage collusion screen remains unproven. Please evaluate an end-to-end reviewer workflow to show whether the flags improve detection or reduce review effort, calibrate thresholds on separate controls and measure fresh false-alarm rates, and address confident attribution to the honest pair. The Stego misattribution should be as prominent as its perfect run-level separation. A clearer hierarchy of the main comparisons and failure modes would make the dense presentation easier to follow.

    Read full reviewShow less

  2. It's a simple and original idea. The evaluation is very honest, predictions are listed next to outcomes, as well as the wrong ones and the harmless-note effect which weakens the framing of the paper is shown clearly.

    My main concern is that the results are expected from the setup. It seems like the collusion and honest runs differ only in what the agents are told. An agents activations reflect what it has read, so two agents who read the same extra text will be closer than others in activation space. The harmless-note result suggests that the method picks up shared text and not collusion.

    This also affects other results. For example, detection before an agent speaks follow from the system prompt being read first.

    The most convincing step would be a setting where collusion is not written into the prompt and where it e.g. emerges from interaction or incentives.

Cite this project

@misc{roldan2026reading,
  title = {{Reading Exclusive Context from Activations: A Label-Free Monitor for Colluding LLM Agents}},
  author = {Jorge Roldan},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/reading-exclusive-context-from-activations-a-labelfree-monitor-for-colluding-llm-agents-4dq9}},
  url = {https://apartresearch.com/sprints/projects/reading-exclusive-context-from-activations-a-labelfree-monitor-for-colluding-llm-agents-4dq9}
}

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