List your company

Question Answering on TruthfulQA (Primary Truthfulness Metric)

58.4TruthfulQAGRPO
16.90427.67738.4549.223Oct 1, 2025Nov 1, 2025Dec 2, 2025Jan 2, 2026Feb 2, 2026Mar 5, 2026Apr 5, 2026

Evaluation results

MethodLinks
GRPOBackbone=Qwen3-8B, Scenario=Strategic Deception2026.0358.4
CoT MonitorBackbone=Llama-3.1-8B, Scenario=Strategic Deception2026.0357.98
GRPOBackbone=Llama-3.1-8B, Scenario=Strategic Deception2026.0356.42
SARBackbone=Qwen3-8B, Scenario=Strategic Deception2026.0356.11
CoT MonitorBackbone=Qwen3-8B, Scenario=Strategic Deception2026.0355.77
SARBackbone=Qwen3-8B, Scenario=Sycophancy2026.0354.87
GRPOBackbone=Llama-3.1-8B, Scenario=Sycophancy2026.0354.74
CoT MonitorBackbone=Llama-3.1-8B, Scenario=Sycophancy2026.0354.56
Qwen3-8BScenario=Strategic Deception2026.0354.44
Qwen3-8BScenario=Sycophancy2026.0354.44
CoT MonitorBackbone=Qwen3-8B, Scenario=Sycophancy2026.0354.38
SARBackbone=Llama-3.1-8B, Scenario=Sycophancy2026.0354.23
GRPOBackbone=Qwen3-8B, Scenario=Sycophancy2026.0354.22
Llama-3.1-8BScenario=Strategic Deception2026.0354.05
Llama-3.1-8BScenario=Sycophancy2026.0354.05
SARBackbone=Llama-3.1-8B, Scenario=Strategic Deception2026.0352.42
Base modelRuntime (min)=N/A2026.0430.2
POMERuntime (min)=10.62026.0429.2
MuonRuntime (min)=12.02026.0429
SIFTRuntime (min)=26.42026.0429
AdamWRuntime (min)=8.12026.0428.9
BLURRuntime (min)=8.62026.0427.1
GradDiff w/ HybridAlgorithm=GradDiff, Optimizer=Hybrid2025.1024.1
NPO w/ HybridAlgorithm=NPO, Optimizer=Hybrid2025.1023.8
NPO w/ signAdamAlgorithm=NPO, Optimizer=signAdam2025.1023.6
NPO w/ SAMAlgorithm=NPO, Optimizer=SAM2025.1023.6
NPO w/ RSAlgorithm=NPO, Optimizer=RS2025.1023.4
NPOAlgorithm=NPO2025.1023.3
GradDiffAlgorithm=GradDiff2025.1022.4
NPO w/ signSGDAlgorithm=NPO, Optimizer=signSGD2025.1022.2
GradDiff w/ RSAlgorithm=GradDiff, Optimizer=RS2025.1022.2
GradDiff w/ signSGDAlgorithm=GradDiff, Optimizer=signSGD2025.1021.9
Pre-unlearn2025.1021.4
GradDiff w/ SAMAlgorithm=GradDiff, Optimizer=SAM2025.1021.4
GradDiff w/ signAdamAlgorithm=GradDiff, Optimizer=signAdam2025.1020.8
GradDiff w/ ZOAlgorithm=GradDiff, Optimizer=ZO2025.1020.3
NPO w/ ZOAlgorithm=NPO, Optimizer=ZO2025.1018.5