Question Answering on TruthfulQA MC1
88.8MC1 AccuracyLinUCB
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LinUCBFramework=QueryBandits, Backbone=GPT-4o, Source=Closed, Variant=Best (Dataset)2026.02 | 88.8 | 8.1 | |
| TS (Non-Contextual)Framework=QueryBandits, Backbone=GPT-4o, Source=Closed, Variant=Best (Non-Contextual)2026.02 | 88.7 | 8 | |
| Contextual TSFramework=QueryBandits, Backbone=GPT-4o, Source=Closed, Variant=Best (Overall)2026.02 | 85.2 | 4.5 | |
| SimplifyFramework=QueryBandits, Backbone=GPT-4o, Source=Closed, Notes=No learning, Variant=Best Static2026.02 | 83.4 | 2.7 | |
| GPT-4oBackbone=GPT-4o, Source=Closed2026.02 | 81.4 | 0.7 | |
| GPT-4Backbone=GPT-4, Source=Closed2026.02 | 81.3 | 0.6 | |
| No-Rewrite (Baseline)Framework=QueryBandits, Backbone=GPT-4o, Source=Closed, Notes=Baseline for ∆2026.02 | 80.7 | 0 | |
| ExpandFramework=QueryBandits, Backbone=GPT-4o, Source=Closed, Variant=Worst Static2026.02 | 67.9 | -12.8 | |
| GPT-4o miniBackbone=GPT-4o mini, Source=Closed2026.02 | 66.5 | -14.2 | |
| TruthXBackbone=Llama-2-7B-Chat, Source=Open2026.02 | 54.22 | -26.5 | |
| GPT-3.5 TurboBackbone=GPT-3.5 Turbo, Source=Closed2026.02 | 53.6 | -27.1 | |
| Induce-then-Contrast Decoding (ICD)Backbone=Llama-2-7B-Chat, Source=Open2026.02 | 46.32 | -34.4 | |
| SSQR-1%Avg. Bitwidth=3.445, Backbone=Qwen3-8B2025.07 | 38.19 | — | |
| SSQR-3%Avg. Bitwidth=4.085, Backbone=Qwen3-8B2025.07 | 37.82 | — | |
| SSQR-1%Avg. Bitwidth=4.445, Backbone=Qwen3-8B2025.07 | 37.58 | — | |
| SSQR-4%Avg. Bitwidth=4.405, Backbone=Qwen3-8B2025.07 | 37.21 | — | |
| RTNAvg. Bitwidth=4.125, Backbone=Qwen3-8B2025.07 | 36.84 | — | |
| Truth Forest (TrFr)Backbone=Llama-2-7B-Chat, Source=Open2026.02 | 36.7 | -44 | |
| Qwen3-8B2026.02 | 36.6 | — | |
| HRTNAvg. Bitwidth=4.125, Backbone=Qwen3-8B2025.07 | 36.47 | — | |
| SSQR-3%Avg. Bitwidth=5.085, Backbone=Qwen3-8B2025.07 | 36.47 | — | |
| SDPO2026.02 | 36.47 | — | |
| BF16 BaselineAvg. Bitwidth=16.000, Backbone=Qwen3-8B2025.07 | 36.35 | — | |
| GPTQAvg. Bitwidth=4.125, Backbone=Qwen3-8B2025.07 | 36.35 | — | |
| SSQR-4%Avg. Bitwidth=5.405, Backbone=Qwen3-8B2025.07 | 36.23 | — | |
| HPTQAvg. Bitwidth=3.125, Backbone=Qwen3-8B2025.07 | 36.11 | — | |
| GPTQAvg. Bitwidth=3.125, Backbone=Qwen3-8B2025.07 | 36.11 | — | |
| SSQR-5%Avg. Bitwidth=5.725, Backbone=Qwen3-8B2025.07 | 35.86 | — | |
| SSQR-2%Avg. Bitwidth=4.765, Backbone=Qwen3-8B2025.07 | 35.62 | — | |
| SSQR-5%Avg. Bitwidth=4.725, Backbone=Qwen3-8B2025.07 | 35.37 | — | |
| HRTNAvg. Bitwidth=3.125, Backbone=Qwen3-8B2025.07 | 35.13 | — | |
| HPTQAvg. Bitwidth=4.125, Backbone=Qwen3-8B2025.07 | 35.01 | — | |
| Llama-2-7B-Chat (base)Backbone=Llama-2-7B-Chat, Source=Open2026.02 | 34.64 | -46.1 | |
| Inference Time Intervention (ITI)Backbone=Llama-2-7B-Chat, Source=Open2026.02 | 34.64 | -46.1 | |
| RTNAvg. Bitwidth=3.125, Backbone=Qwen3-8B2025.07 | 34.03 | — | |
| Self-Highlighted Hesitation (SH2)Backbone=Llama-2-7B-Chat, Source=Open2026.02 | 33.9 | -46.8 | |
| SSQR-2%Avg. Bitwidth=3.765, Backbone=Qwen3-8B2025.07 | 33.9 | — | |
| Decoding by Contrasting Layers (DoLa)Backbone=Llama-2-7B-Chat, Source=Open2026.02 | 32.2 | -48.5 | |
| SSQR-4%Avg. Bitwidth=3.405, Backbone=Qwen3-8B2025.07 | 31.33 | — | |
| SSQR-5%Avg. Bitwidth=3.725, Backbone=Qwen3-8B2025.07 | 31.21 | — | |
| HPTQAvg. Bitwidth=2.125, Backbone=Qwen3-8B2025.07 | 31.09 | — | |
| HRTNAvg. Bitwidth=2.125, Backbone=Qwen3-8B2025.07 | 30.48 | — | |
| SSQR-3%Avg. Bitwidth=3.085, Backbone=Qwen3-8B2025.07 | 29.5 | — | |
| GPTQAvg. Bitwidth=2.125, Backbone=Qwen3-8B2025.07 | 28.4 | — | |
| SSQR-1%Avg. Bitwidth=2.445, Backbone=Qwen3-8B2025.07 | 28.4 | — | |
| SSQR-2%Avg. Bitwidth=2.765, Backbone=Qwen3-8B2025.07 | 27.54 | — | |
| Contrast-Consistent Search (CCS)Backbone=Llama-2-7B-Chat, Source=Open2026.02 | 26.2 | -54.5 | |
| Contrastive Decoding (CD)Backbone=Llama-2-7B-Chat, Source=Open2026.02 | 24.4 | -56.3 | |
| Supervised FinetuningBackbone=Llama-2-7B-Chat, Source=Open2026.02 | 24.2 | -56.5 | |
| RTNAvg. Bitwidth=2.125, Backbone=Qwen3-8B2025.07 | 24.11 | — | |
| GPT-2 1.5BBackbone=GPT-2 1.5B, Source=Open2026.02 | 22 | -58.7 | |
| GPT-3 175BBackbone=GPT-3 175B, Source=Closed2026.02 | 21 | -59.7 | |
| GPT-J 6BBackbone=GPT-J 6B, Source=Open2026.02 | 20 | -60.7 | |
| UnifiedQA 3BBackbone=UnifiedQA 3B, Source=Open2026.02 | 19 | -61.7 |