Chain-of-thought monitoring

What is Chain-of-thought monitoring?

Oversight method that inspects intermediate reasoning traces to detect unsafe or policy-violating behavior

Also known as: Chain-of-thought monitorability

In other languages

한국어사고 과정 모니터링

모델이 답을 내기까지의 중간 추론 흐름을 관찰해 위험하거나 비정상적인 행동을 탐지하는 기법이다.

日本語思考連鎖モニタリング

モデルが答えに至る途中の推論過程を監視し、危険な意図や逸脱した振る舞いを検出する手法

Related Terms

• Runtime monitoringObserving a software system while it is running to detect behavior, errors, or policy violations.
• Automated Reasoning ChecksFormal verification method that tests whether system behavior satisfies explicitly defined logical policies or constraints
• LLM observabilityTooling for tracing, measuring, and debugging language model calls in applications
• Maker/checkerAn internal control pattern where one process performs an action and a second, independent process verifies its correctness
• Gate canaryMonitoring check that uses known-bad inputs to confirm a safety gate still detects failures
• Multi-step verificationA methodology in computational reasoning where a system evaluates the accuracy of individual intermediate steps and the overall logic of a multi-part process.
• Taint analysisA software testing technique that tracks the flow of untrusted or potentially compromised data through a system.
• Runtime criticsEvaluation components that assess a system's behavior while it is running, often to detect errors or guide corrections
• Prompt EnhancementTechnique for adding clearer requirements, constraints, and checks to an instruction before model execution
• Lean certificateMachine-checkable formal proof written in the Lean theorem prover, used to verify mathematical arguments with software
• LintingAutomated process of analyzing source code to flag programming errors, bugs, and stylistic inconsistencies
• Automated ReasoningA technique that uses mathematical logic to rigorously verify whether statements or program behaviors are correct.