Cognition · Data artifact
Self-Consistency Sampling Policy
Data artifactCognitionCognition & Memoryarc:SelfConsistencySamplingPolicy
A configuration artifact mapping problem classes or difficulty tiers to sample count k, decoding parameters (temperature, top-k, top-p) and aggregation method for self-consistency.
Responsibility. Specifies sampling budget and decoding parameters per problem class.
Also known as: Tiered sampling configuration, Sample budget policy
Relationships
configures structural
is read by dependency
is evaluated by assurance
Design guidance
- SHOULD scale sample count with problem difficulty and required accuracy, derived from the cost of an error versus the cost of extra samples.
- SHOULD record which sampling configurations work for which problem classes as procedural knowledge.
Quantitative guidance
As stated by the sources; verify before use.
- Easy problems (85% base): k=5 gives 91%, k=20 only 93%; hard problems (45% base): k=5 58%, k=10 65%, k=20 71% (Ch5.3).
- Medical diagnosis example: k=8 with quality weighting met ~90% accuracy within a 10-15 s latency budget (~12 s parallel); k=8 to k=15 gained only 1.2 points for 87% more cost (Ch5.3).
Classification
- Patterns
- Tiered deployment (easy k=3-5 majority; medium k=5-10 light weighting; hard k=10-40 full RASC)
- Quality attributes
- Cost efficiencyFunctional suitability: correctness and validity (ISO/IEC 25010 | NIST AI RMF: valid)
- Risks mitigated
- Uniform over-sampling of easy queriesUnder-sampling of hard queries
Sources
- Ch5.3: T. Nguyen, "Self-Consistency Fundamentals," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 5.3. ISBN: 9798244538229.