Research Platform Updated Mar 2026

LLM Hallucination
Benchmark & Ranking

TruthAnchor의 다층 검증 파이프라인으로 주요 LLM의 할루시네이션을 정량 분석합니다. 사실 정확도, 수치 정확도, 근거 신뢰도, 일관성, 불확실성 보정 5개 차원을 평가합니다.

View LeaderboardMethodology
#1 Ranked

Claude Opus 4.6

anthropic
94.2/ 100
96.1
factual
93.5
numerical
92.8
citation
95.0
consistency
91.7
uncertainty
12+
Models Evaluated
6
Domains Covered
5
Score Dimensions
500+
Benchmark Questions

Overall Ranking

Hallucination Defense Score across all domains

Full leaderboard
Claude Opus 4.6
anthropic
94.2
overall
96
Factual
94
Numerical
93
Citation
95
Consistency
92
Uncertainty
2
GPT-4o
openai
91.8
overall
93
Factual
91
Numerical
91
Citation
92
Consistency
89
Uncertainty
3
Claude Sonnet 4.6
anthropic
90.5
overall
93
Factual
89
Numerical
90
Citation
91
Consistency
89
Uncertainty
4
Gemini 2.0 Pro
google
89.1
overall
91
Factual
89
Numerical
88
Citation
90
Consistency
87
Uncertainty
5
DeepSeek V3
deepseek
87.3
overall
89
Factual
87
Numerical
86
Citation
89
Consistency
85
Uncertainty
6
Mistral Large
mistral
85.9
overall
87
Factual
85
Numerical
84
Citation
88
Consistency
84
Uncertainty
7
Llama 3.3 70B
meta
84.1
overall
86
Factual
83
Numerical
83
Citation
86
Consistency
83
Uncertainty
8
Qwen 2.5 72B
alibaba
83.5
overall
85
Factual
83
Numerical
82
Citation
85
Consistency
82
Uncertainty

Evaluation Framework

5 orthogonal dimensions measuring different aspects of hallucination

Factual Accuracy

35%

Claim extraction & NLI verification against ground-truth evidence

Numerical Accuracy

20%

Financial calculation verification with domain-specific tolerances

Citation Reliability

15%

Semantic similarity scoring between claims and source documents

Consistency

15%

Self-consistency measurement across multiple inference samples

Uncertainty Calibration

15%

Entropy-based confidence alignment with actual accuracy