ResearchBenchmarksQuestion Answering on NQ, TriviaQA, PopQA, HotpotQA, 2Wiki, MuSiQue, Bamboogle (held-out val)Question Answering on NQ, TriviaQA, PopQA, HotpotQA, 2Wiki, MuSiQue, Bamboogle (held-out val)Follow48.4Exact Match (EM)TIPS33.00837.0044144.996Mar 11, 2026Evaluation resultsMethodMethodLinksExact Match (EM)F1 ScoreFLOPs Overhead (%)TIPSModel=Qwen3-4B-Instruc...Model=Qwen3-4B-Instruct-25072026.0348.457.111.846TIPSModel=Qwen2.5-14B-Inst...Model=Qwen2.5-14B-Instruct2026.0345.453.111.813TIPSModel=Qwen2.5-7B-InstructModel=Qwen2.5-7B-Instruct2026.0341.751.211.81TIPSModel=Llama3.1-8BModel=Llama3.1-8B2026.0340.34911.659TIPSModel=Qwen2.5-3B-InstructModel=Qwen2.5-3B-Instruct2026.0333.641.111.761