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

S-Space Steering for Eval-Awareness Control in Reasoning Models

Michael J Clark · Team AntiPaSTO

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

Read the report

Report: S-Space Steering for Eval-Awareness Control in Reasoning Models

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

We can no longer trust evaluations of frontier models because they detect when they are being evaluated, and this problem is getting worse as models get more capable. Recent work replicated this eval-awareness in open-weight models and showed it responds to activation steering in Qwen3-32B, but standard steering vectors are unreliable and the proposed rank-1 weight surgery makes things worse on Qwen3. We apply S-space steering, a novel method that steers in the singular-value basis of weight matrices (the transformation space) rather than in activation space (the data space), making the perturbation input-dependent. With eval-awareness suppressed, the gap between how the model behaves on eval-looking vs real-looking prompts shrinks to +1pp, down from +7pp at baseline and +26pp with the prior method. We can steer a model for eval-unawareness, then run evaluations with higher confidence that the results reflect real behavior.

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. Nice proof of concept that training-free steering in the singular-value basis of weight matrices with contrastive persona pairs can successfully reduce compliance gap between real and eval-looking tasks. It suggests that the method can successfully reduce scenario recognition in the model, and I am very interested to see S-Space steering validated across more models and tasks.

    In the presented work, I would also be interested in a follow-up analysis on why "hypothetical" scenarios have a higher compliance rate than "real" scenarios, but decreasing awareness leads to higher absolute compliance rates. This suggests multiple compliance mechanisms might be at play simultaneously.

  2. This is quite frankly, amazing work. I would like the author to maybe dive deeper into the BashArena demo. The side task result where eval-unaware direction actually suppressed compliance with an injected malicious instruction is the most important finding in the paper if replicated. But with n=1, and the author's own caveat - it's a POC right now. Maybe scaling this to more samples will be useful

Cite this project

@misc{clark2026sspace,
  title = {{S-Space Steering for Eval-Awareness Control in Reasoning Models}},
  author = {Michael J Clark},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/sspace-steering-for-evalawareness-control-in-reasoning-models-7j1i}},
  url = {https://apartresearch.com/sprints/projects/sspace-steering-for-evalawareness-control-in-reasoning-models-7j1i}
}

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