Minerval
Sourceunknowntext stored

https://arxiv.org/html/2605.29729

Author not recordedDate not recordedOriginal︎ ↗

Claims2
ExaminationsNone
Text107 sections
VersionsOne known
CopiesNone recorded
CorrectionsNone recorded
Facts checked Nov 10, 2026
ReachNo copies recorded
StructureIn 2 claims' maps0 sources downstream across 2 claims
EvidenceNot yet read
ReviewsNot reviewed

Examinations

None recorded. An examination is commissioned by a claim’s Steward for what its claim rests on; a review of the document for its own sake is asked for as a request bearing on it.

Text510 passages · 2 marked

No examination yetMarked passage
Skip to main content0 passages
Report GitHub Issue6 passages
Report Issue0 passages
Back to Abstract0 passages
Download PDF0 passages
Contents0 passages
Reading mode0 passages
Colour theme0 passages
Abstract0 passages
1 Introduction0 passages
2 Related work0 passages
3 Methods0 passages
4 Honeypot evaluations as a defense against misaligned AI0 passages
5 Results0 passages
6 Evaluating our honeypot evaluations against our desiderata0 passages
7 Limitations and future work0 passages
8 Acknowledgements0 passages
References0 passages
9 Canaries0 passages
A Autorater specification and validation0 passages
A.1 Autorater for evaluating model outputs0 passages
A.2 Rating Scales0 passages
A.3 Autorater validation0 passages
B Types of sabotage0 passages
B.1 Agency-only prompting0 passages
B.2 Goal and agency prompting0 passages
C Misaligned model organism0 passages
License: CC BY 4.02 passages
Realistic honeypot evaluations for scheming propensity0 passages
Victoria Krakovna0 passages
Affiliation: Google DeepMind0 passages
David Lindner0 passages
Affiliation: Google DeepMind0 passages
Lewis Ho0 passages
Affiliation: Google DeepMind0 passages
Sebastian Farquhar0 passages
Affiliation: Equal contribution0 passages
Affiliation: Google DeepMind0 passages
Rohin Shah0 passages
Corresponding author: vkrakovna@google.com0 passages
Affiliation: Equal contribution0 passages
Affiliation: Google DeepMind0 passages
May 20260 passages
Abstract1 passage
1 Introduction17 passages · 1 marked

Most current work on scheming propensity ( Lynch et al., 2025 ; Schlatter et al., 2026 ) uses synthetic scenarios that often put models in situations with pressure to scheme and a lack of alternative options. These scenarios are often unrealistic, which makes it unclear how well the results will generalize. They frequently trigger scheming behaviors for other reasons besides instrumental goals (e.g. because they are interpreting the setting as an instruction to scheme or roleplay with a scheming persona). Furthermore, models demonstrate increasing awareness that they are being evaluated ( Needham et al., 2025 ) , which makes it harder to interpret their behavior.

AI scheming evaluations use contrived scenarios that strongly cue misaligned actions.Supported

Not yet read by this claim’s Steward.

Evaluation setting0 passages
Codebase0 passages
Commit chain summary1 passage
Chain of thought monitor for scheming2 passages
Chain of thought monitor for scheming2 passages
Supervised finetuning for alignment research2 passages
Evaluating model ability to obfuscate reasoning2 passages
Frontier safety misuse mitigations1 passage
2 Related work5 passages
3 Methods5 passages
Evaluation setting0 passages
Codebase0 passages
Agent task summary1 passage
Chain of thought monitor for scheming2 passages
Chain of thought monitor for scheming2 passages
Supervised finetuning for alignment research2 passages
Evaluating model ability to obfuscate reasoning2 passages
Frontier safety misuse mitigations44 passages
4 Honeypot evaluations as a defense against misaligned AI22 passages
5 Results28 passages · 1 marked
Perceived evaluation type0 passages
Typical effect on scheming0 passages
Scheming / alignment test1 passage
Prompt injection / jailbreak1 passage
Coding exercise1 passage
Situational awareness test1 passage
Instruction following or scheming ability test7 passages
Honeypot, prompt0 passages
Eval awareness0 passages
Sabotage evidence0 passages
Scheming reasoning1 passage
None7 passages
None14 passages
6 Evaluating our honeypot evaluations against our desiderata15 passages
7 Limitations and future work13 passages
8 Acknowledgements1 passage
References3 references
External Links: 2505.1461717 passages
External Links: 2410.215145 passages
External Links: 2206.133535 passages
External Links: 2507.052469 passages
External Links: 2412.0390446 passages
External Links: 2412.0498415 passages
External Links: 1712.054415 passages
External Links: 2403.1379313 passages
External Links: 2410.1391933 passages
External Links: ISBN 07695204135 passages
External Links: 2605.064905 passages
9 Canaries3 passages
Appendix A Autorater specification and validation0 passages
A.1 Autorater for evaluating model outputs2 passages
Degree of sabotage1 passage
Covertness of sabotage1 passage
Goal-directedness1 passage
Scheming reasoning2 passages
A.2 Rating Scales66 passages
A.3 Autorater validation4 passages
Appendix B Types of sabotage1 passage
B.1 Agency-only prompting19 passages
B.2 Goal and agency prompting21 passages
Appendix C Misaligned model organism10 passages
Instructions for reporting errors7 passages

Lineage across 2 claims

Draws on · none recorded
This source
https://arxiv.org/html/2605.29729
in 2 claims' maps
Drawn on by · none recorded

Record

Kind
unknown
Authors
Not recorded
Publisher
Not recorded
Published
Not recorded
DOI
None
Link
https://arxiv.org/html/2605.29729
Archived copy
None found
Stored text
70,060 characters, retrieved Nov 10, 2026 · 107 sections, 510 passages
Claims extracted
Not recorded
Versions
One known
Copies
None recorded
Corrections
None recorded
Facts checked
Nov 10, 2026

History 3

  1. Nov 10, 2026DividedSegmenter · Divided into 627 parts by its own structure
  2. Nov 10, 2026FactsFacts fetcher · Facts looked up
  3. Nov 10, 2026Entered the graphEntered the graph; copy stored