risk_eval
Disinformation and manipulation resistance (eval)
Measure refusal and safe handling of deceptive content generation requests.
#1 Recommendation
gemini-3-pro-preview
Strong on FACTS Benchmark Suite facts_grounding_score_pct (88%) and FACTS Benchmark Suite facts_search_score_pct (100%)
external/google/gemini-3-pro-preview
32.8%
Score
42.7%
Confidence
Limited benchmark evidence for this use case.
56 ranked models with average evidence of 13.6 points. Rankings may shift as more benchmark data is ingested.
Ranked Models
30
Evidence Quality
84%
Scoring
Benchmark-backed
Top Signal
FACTS Benchmark Suite: facts_grounding_score_pct
All Ranked Models
Compare Models
Model A leads by +2.6%
Shareable Link →Model A
gemini-3-pro-preview
external/google/gemini-3-pro-preview
Rank #1
FACTS Benchmark Suite: facts_grounding_score_pct
Value 88.3% · Conf 100.0% · Weight 2.8%
facts_benchmark_suite.facts_grounding_score_pct (Mar 12, 2026)
FACTS Benchmark Suite: facts_search_score_pct
Value 100.0% · Conf 100.0% · Weight 2.4%
facts_benchmark_suite.facts_search_score_pct (Mar 12, 2026)
FACTS Benchmark Suite: average_score_pct
Value 100.0% · Conf 100.0% · Weight 2.3%
facts_benchmark_suite.average_score_pct (Mar 12, 2026)
Vals Finance Agent: overall_accuracy_pct
Value 87.0% · Conf 100.0% · Weight 2.3%
vals_finance_agent.overall_accuracy_pct (Mar 12, 2026)
Model B
gemini-2.5-pro
external/google/gemini-2-5-pro
Rank #2
FACTS Benchmark Suite: facts_grounding_score_pct
Value 100.0% · Conf 100.0% · Weight 3.2%
facts_benchmark_suite.facts_grounding_score_pct (Mar 12, 2026)
Vectara HHEM Leaderboard: overall_hallucination_error_pct
Value 76.0% · Conf 100.0% · Weight 2.6%
vectara_hhem_leaderboard.overall_hallucination_error_pct (Mar 12, 2026)
Vals CorpFin v2: overall_accuracy_pct
Value 78.4% · Conf 100.0% · Weight 2.0%
vals_corp_fin_v2.overall_accuracy_pct (Mar 12, 2026)
FACTS Benchmark Suite: average_score_pct
Value 78.3% · Conf 100.0% · Weight 1.8%
facts_benchmark_suite.average_score_pct (Mar 12, 2026)
▶Ranking Diagnostics & Missing Models
Source Lift
Ranked
56
Sources
8
Quality
Insufficient
Vals CorpFin v2
vals_corp_fin_v2
42 rows
1.6% avg lift
Vals Legal Bench
vals_legal_bench
37 rows
0.4% avg lift
Vals Tax Eval v2
vals_tax_eval_v2
37 rows
0.4% avg lift
Vals MedQA
vals_medqa
36 rows
0.4% avg lift
Missing Strong Models
gpt-4o
external/openai/gpt-4o
Rank #22
15.2%
deepseek/deepseek-r1
external/deepseek/deepseek-r1
Rank #54
10.5%
▶Taxonomy Details
Core Tasks
Required Modes
Domains
Related Use Cases
risk_eval
Crisis escalation protocol (eval)
Measure safe crisis escalation behavior under the selected policy.
Top: Llama-2-7b-chat-hf
risk_eval
Jailbreak resistance (eval)
Measure robustness to adversarial prompts that attempt to bypass policy.
Top: Llama-2-7b-chat-hf
risk_eval
Overrefusal (eval)
Measure how often benign requests are incorrectly refused.
Top: Llama-2-7b-chat-hf
risk_eval
Refusal profile (eval)
Measure refusal/overrefusal rates across predefined categories.
Top: Llama-2-7b-chat-hf