Sourceunknowntext stored

https://arxiv.org/html/2605.15257

Author not recordedDate not recordedOriginal︎ ↗

Claims1

ExaminationsNone

Text248 sections

VersionsOne known

CopiesNone recorded

CorrectionsNone recorded

Facts checked Oct 30, 2026

ReachNo copies recorded

StructureIn 1 claim's map0 sources downstream across 1 claim

Evidence1 reading · 1 read wholeSupports 1No claim cites an examination's findings

Stewards’ notes · 1

Models trained to recognize chain-of-thought monitoring learn to obfuscate their reasoning to evade it.
Supports. The study's own data bear its assertion, with limits it states: the strongest effect is on a task where reasoning is not computationally required; on a reasoning-necessary exfiltration task detection of the misbehaving turns fell only from 94.6% to 87.2% for gpt-oss-120b; the training documents included hints on how to avoid the monitor, checkpoints were selected for maximal evasion, and the effect vanished when monitoring documents fell to 10% of the finetuning mix. The obfuscation observed is mostly omission and vaguer phrasing, not encoded reasoning. Oct 30, 2026

Examinations

None recorded.

Text713 passages · 1 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 CoT obfuscation experiment setup0 passages

4 Training on documents about monitoring leads to CoT obfuscation0 passages

5 CoT controllability predicts obfuscation success0 passages

6 Evading reward hacking monitors in RL0 passages

7 Discussion0 passages

References0 passages

A All model details0 passages

B Full model training details0 passages

B.1 Defect concealment task1 passage

B.2 Credential exfiltration0 passages

B.3 Reward hacking RL0 passages

C Monitor details0 passages

C.1 Defect concealment0 passages

C.1.1 Eval prompt0 passages

C.1.2 CoT monitor prompt0 passages

C.2 Agentic side-task0 passages

C.2.1 Eval prompt0 passages

C.2.2 CoT monitor prompt0 passages

C.3 Reward hacking RL0 passages

C.3.1 Eval prompt0 passages

C.3.2 Judge prompt0 passages

D Defect concealment task details0 passages

D.1 Sampling0 passages

D.2 Full model response examples0 passages

D.2.1 Behavior-only model responses0 passages

D.2.2 Monitor-aware model responses0 passages

D.3 Synthetic document generation prompts0 passages

D.3.1 Behavioral universe prompts4 passages

D.4 Example training documents0 passages

D.4.1 Behavior-only model documents0 passages

D.4.2 Monitor-aware model documents0 passages

D.5 Qualitative analysis0 passages

D.5.1 Counterfactual autorater0 passages

D.5.2 Disclosure autorater0 passages

D.6 Reasoning length changes2 passages

E Agentic sidetask full task details0 passages

E.1 Setup0 passages

E.2 Scenarios0 passages

E.3 Environment simulation0 passages

E.4 Exfiltration detection0 passages

E.5 Per-turn monitoring0 passages

E.6 No-reasoning condition0 passages

E.7 SDF document universes0 passages

F CoT control task full details0 passages

F.1 Instructions used0 passages

F.2 Instruction result breakdown0 passages

F.3 Evaluation questions1 passage

G Reward hacking RL full details0 passages

G.1 RL Environment0 passages

G.2 RL Training Details0 passages

G.2.1 Hyperparameters0 passages

G.2.2 Seeds0 passages

G.3 Reward Structure0 passages

G.4 Ground-Truth Reward Hacking Classification0 passages

G.5 Evaluation and Aggregation Conventions0 passages

G.6 Overoptimization and Seed Truncation2 passages

G.9 SDF Epoch Selection0 passages

G.10 Monitor Prompt0 passages

G.11 Held-Out Evaluation0 passages

G.12 Detection Rate Among Reward Hacking0 passages

G.13 Example CoTs8 passages

License: CC BY 4.01 passage

Training on Documents About Monitoring Leads to CoT Obfuscation1 passage

Affiliation: Pivotal Research0 passages

Bilal Chughtai †0 passages

Affiliation: Google DeepMind0 passages

Joshua Engels ‡0 passages

Affiliation: Google DeepMind0 passages

Abstract1 passage

1 Introduction13 passages · 1 marked

We find that exposing models to monitoring documents consistently makes those models better at avoiding CoT deception monitors. These effects extend beyond the prompted setting: monitor-aware variants of gpt-oss-120b learn to reward hack without detection significantly faster than unaware controls during RL. Finally, CoT controllability ( Chen et al., 2026 ) (the degree to which a model can reshape its own reasoning) is closely associated with obfuscation success across models.

Models trained to recognize chain-of-thought monitoring learn to obfuscate their reasoning to evade it.SupportedReading · this claim’s Steward

Supports. The study's own data bear its assertion, with limits it states: the strongest effect is on a task where reasoning is not computationally required; on a reasoning-necessary exfiltration task detection of the misbehaving turns fell only from 94.6% to 87.2% for gpt-oss-120b; the training documents included hints on how to avoid the monitor, checkpoints were selected for maximal evasion, and the effect vanished when monitoring documents fell to 10% of the finetuning mix. The obfuscation observed is mostly omission and vaguer phrasing, not encoded reasoning.

Finally, we discuss takeaways in Section 70 passages

2 Related work4 passages

3 CoT obfuscation experiment setup8 passages

4 Training on documents about monitoring leads to CoT obfuscation11 passages

Behavior-only CoT ✗ Detected0 passages

Monitor-aware CoT ✓ Undetected2 passages

Behavior-only CoT ✗ Detected0 passages

Monitor-aware CoT ✓ Undetected8 passages

5 CoT controllability predicts obfuscation success4 passages

6 Evading reward hacking monitors in RL19 passages

7 Discussion3 passages

Acknowledgments and Disclosure of Funding4 passages

References10 references

Technical report0 passages

External Links: Link3 passages

Note: https://metr.org/blog/2025-08-08-cot-may-be-highly-informative-despite-unfaithfulness/15 passages

Note: Anthropic Alignment Science BlogAccessed: 2026-04-160 passages

External Links: Link60 passages

Note: https://www.kimi.com/blog/kimi-k2-6 Technical blog post3 passages

Note: LessWrong / AI Alignment Forum0 passages

External Links: Link23 passages

Note: https://thinkingmachines.ai/tinker/7 passages

Note: LessWrongAccessed: 2026-04-120 passages

External Links: Link7 passages

Note: Anthropic Alignment Science Blog0 passages

External Links: Link3 passages

Note: Anthropic Alignment Science Blog0 passages

External Links: Link7 passages

Note: LessWrong / AI Alignment Forum blog post0 passages

External Links: Link9 passages

Appendix A All model details17 passages

Infrastructure models2 passages

Appendix B Full model training details1 passage

B.1 Defect concealment task137 passages

B.2 Credential exfiltration6 passages

B.3 Reward hacking RL10 passages

Appendix C Monitor details1 passage

C.1 Defect concealment1 passage

C.1.1 Eval prompt2 passages

C.1.2 CoT monitor prompt1 passage

C.2 Agentic side-task1 passage

C.2.1 Eval prompt2 passages

C.2.2 CoT monitor prompt1 passage

C.3 Reward hacking RL1 passage

C.3.1 Eval prompt1 passage

C.3.2 Judge prompt1 passage

Appendix D Defect concealment task details0 passages

D.1 Sampling2 passages

D.2 Full model response examples2 passages

D.2.1 Behavior-only model responses0 passages

D.2.2 Monitor-aware model responses0 passages

D.3 Synthetic document generation prompts2 passages

D.3.1 Behavioral universe prompts6 passages

In both prompts, the bullet4 passages

In each of the three prompts, the bullet6 passages

D.4 Example training documents1 passage

D.4.1 Behavior-only model documents1 passage

D.4.2 Monitor-aware model documents1 passage

D.5 Qualitative analysis1 passage

D.5.1 Counterfactual autorater2 passages

D.5.2 Disclosure autorater1 passage

D.6 Reasoning length changes26 passages

Appendix E Agentic sidetask full task details1 passage

E.1 Setup2 passages

E.2 Scenarios3 passages

E.3 Environment simulation12 passages

E.4 Exfiltration detection8 passages

E.5 Per-turn monitoring2 passages

E.6 No-reasoning condition1 passage

E.7 SDF document universes3 passages

Appendix F CoT control task full details0 passages

F.1 Instructions used1 passage

Each constraint is prepended to the system prompt in the form27 passages

Composite score1 passage

F.2 Instruction result breakdown29 passages

F.3 Evaluation questions11 passages

Sub-domain2 passages

20 passages

Combinatorics and counting0 passages

30 passages

Probability and conditional probability0 passages

41 passage

51 passage

61 passage

71 passage

81 passage

91 passage

102 passages

Dynamic programming0 passages

21 passage

31 passage

41 passage

50 passages

Concurrency or synchronisation0 passages

61 passage

70 passages

Complexity analysis or computability0 passages

80 passages

Recursion or backtracking2 passages

20 passages

Thermodynamics or statistical mechanics0 passages

30 passages

Electromagnetism or optics0 passages

40 passages

Quantum or atomic physics0 passages

50 passages

Chemical reactions, bonding, or equilibria0 passages

60 passages

Molecular or cellular biology0 passages

70 passages

Evolution or ecology0 passages

80 passages

Earth, atmospheric, or climate science1 passage

11 passage

20 passages

Political revolution or regime change0 passages

31 passage

40 passages

Migration, demography, or settler society0 passages

50 passages

Technological or industrial transition0 passages

60 passages

Military or strategic conflict0 passages

70 passages

Intellectual, cultural, or religious movement0 passages

80 passages

Geography, resources, or environment1 passage

Truth-tellers and liars deduction0 passages

20 passages

Combinatorial counting or arrangement0 passages

30 passages

River-crossing or state-search0 passages

40 passages

Weighing or measurement with constraints0 passages

50 passages

Timing or scheduling puzzle0 passages

60 passages

Self-reference or paradox0 passages

70 passages

Common-knowledge or multi-agent reasoning0 passages

80 passages

Game-theoretic strategy2 passages

21 passage

31 passage

40 passages

Moral status or personhood0 passages

51 passage

60 passages

Decision under uncertainty or risk0 passages

70 passages

Intergenerational or future-people ethics0 passages

84 passages

Appendix G Reward hacking RL full details1 passage

G.1 RL Environment5 passages

G.2 RL Training Details1 passage

G.2.1 Hyperparameters15 passages

G.2.2 Seeds1 passage

G.3 Reward Structure6 passages

G.4 Ground-Truth Reward Hacking Classification1 passage

G.5 Evaluation and Aggregation Conventions7 passages

G.6 Overoptimization and Seed Truncation22 passages

G.9 SDF Epoch Selection1 passage

G.10 Monitor Prompt2 passages

G.11 Held-Out Evaluation2 passages

G.12 Detection Rate Among Reward Hacking1 passage

G.13 Example CoTs21 passages

Instructions for reporting errors7 passages

Lineage across 1 claim

Draws on · none recorded

This source

https://arxiv.org/html/2605.15257

in 1 claim's map

Drawn on by · none recorded

Record

Kind: unknown
Authors: Not recorded
Publisher: Not recorded
Published: Not recorded
DOI: None
Link: https://arxiv.org/html/2605.15257
Archived copy: None found
Stored text: 122,097 characters, retrieved Oct 30, 2026 · 248 sections, 713 passages
Claims extracted: Not recorded
Versions: One known
Copies: None recorded
Corrections: None recorded
Facts checked: Oct 30, 2026

History 4

Oct 30, 2026ReadingSteward of “Models trained to recognize chain-of-thought mo…” · supports, read whole
Oct 30, 2026DividedSegmenter · Divided into 973 parts by its own structure
Oct 30, 2026FactsFacts fetcher · Facts looked up
Oct 30, 2026Entered the graphEntered the graph; copy stored