List your company
ResearchFormerly WizWand

Question Answering on PIQA (B-1, R-L)

0.5Accuracy DeltaGAIN
-0.1864-0.00820.170.3482Dec 9, 2025Dec 28, 2025Jan 17, 2026Feb 6, 2026Feb 25, 2026Mar 17, 2026Apr 6, 2026
Updated 6mo ago

Evaluation results

MethodLinks
GAINModel=Mistral-7B, Params=131K2026.040.5———
LoRAModel=Mistral-7B, Params=262K, Learning Rate=10−32026.040.3———
LoRAModel=Mistral-7B, Params=262K, Learning Rate=3×10−42026.040.3———
LoRAModel=Llama-2-13B, Params=410K, Learning Rate=10−32026.040.2———
GAINModel=Llama-2-70B, Params=655K2026.040.2———
GAINModel=Llama-2-13B, Params=205K2026.040———
LoRAModel=Llama-2-13B, Params=410K, Learning Rate=3×10−42026.040———
LoRAModel=Llama-2-70B, Params=1.3M, Learning Rate=10−32026.040———
LoRAModel=Llama-2-70B, Params=1.3M, Learning Rate=3×10−42026.040———
AutoPromptTarget Model=Qwen2-1.5B, Suffix length (K)=42025.12-0.04——0.1
Proposed methodTarget Model=Qwen2-1.5B, Suffix length (K)=42025.12-0.04——0.45
Universal TriggerTarget Model=Qwen2-1.5B, Suffix length (K)=42025.12-0.16——0.3
Diverse DecodingTraining Strategy=Inference-time optimization (LM FT / -)2022.04—11.315.5—
FiD-T5Model Scale=base, Backbone=T5, Training Strategy=LM FT2022.04—12.911.9—
FiD-T5Model Scale=large, Backbone=T5, Training Strategy=LM FT2022.04—14.817.4—
KIDGPT2Backbone=GPT2, Implementation=Knowledge trie, Training Strategy=Inference-time optimization2022.04—17.718.4—
Little KIDImplementation=naive (plain list), Training Strategy=Inference-time optimization2022.04—12.213.5—
PPLMTraining Strategy=Inference-time optimization (LM FT / -)2022.04—10.311.9—