Research map / Large language models / Retrieval and knowledge
Explore What LLM Does Not Know in Complex Question Answering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes KEQA, a retrieval-augmented QA framework that uses quiz-based consistency checks to detect what the LLM does not know and a utility-guided picker to retrieve and keep only helpful external passages.
Problem
Retrieval-augmented QA for complex questions needs the LLM to know which required knowledge it lacks (its knowledge boundary) and which retrieved passages actually help reasoning; existing confidence- or probability-based adaptive retrieval is inaccurate or unavailable for black-box LLMs.
Approach
Decompose the complex question into simple quizzes, judge the LLM's knowledge by the semantic consistency of multiple sampled answers (self-consistency), retrieve only for failed quizzes, and evaluate retrieved passages' utility by whether they raise F1 in reasoning, using a utility-labeled demonstration set from training data.
Key results
- On NQ, KEQA reaches F1 0.483 and EM 0.352, versus 0.455 F1 for the strongest baseline FLARE (EM 0.318).
- On 2WikiMultihopQA, KEQA reaches F1 0.405 and EM 0.326, versus 0.368 F1 for SlimPLM.
- On StrategyQA, KEQA reaches 0.680 accuracy, versus 0.666 for Rowen.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use KEQA instead of FLARE | GPT-3.5 on NQ · NQ | F1: 0.455 -> 0.483* | Helped |
| use KEQA instead of FLARE | GPT-3.5 on NQ · NQ | EM: 0.318 -> 0.352* | Helped |
| use KEQA instead of Rowen | GPT-3.5 on StrategyQA · StrategyQA | ACC: 0.666 -> 0.680* | Helped |
| use KEQA instead of SlimPLM | GPT-3.5 on StrategyQA · StrategyQA | F1: 0.393 -> 0.400* | Helped |
| use KEQA instead of ReAct | GPT-3.5 on StrategyQA · StrategyQA | EM: 0.270 -> 0.278* | Helped |
| use KEQA instead of SlimPLM | GPT-3.5 on HotpotQA · HotpotQA | F1: 0.368 -> 0.405* | Helped |
| use KEQA instead of FLARE | GPT-3.5 on HotpotQA · HotpotQA | EM: 0.268 -> 0.278* | Helped |
| remove quiz-based knowledge evaluation (ablation) | KEQA on QA benchmarks · NQ, StrategyQA, HotpotQA, 2WikiMultihopQA | F1/ACC: 0.483 -> 0.409; 0.680 -> 0.644; 0.405 -> 0.396; 0.326 -> 0.258 | Hurt |
| use all retrieved results instead of utility-guided knowledge picking (ablation) | KEQA on QA benchmarks · NQ, StrategyQA, HotpotQA, 2WikiMultihopQA | F1/ACC: 0.483 -> 0.453; 0.680 -> 0.678; 0.405 -> 0.350; 0.326 -> 0.314 | Hurt |
| remove utility reference demonstrations (ablation) | KEQA on QA benchmarks · NQ, StrategyQA, HotpotQA, 2WikiMultihopQA | F1/ACC: 0.483 -> 0.456; 0.680 -> 0.676; 0.405 -> 0.398; 0.326 -> 0.302 | Hurt |
| randomly sample utility demonstrations (ablation) | KEQA on QA benchmarks · NQ, StrategyQA, HotpotQA, 2WikiMultihopQA | F1/ACC: 0.483 -> 0.474; 0.680 -> 0.666; 0.405 -> 0.385; 0.326 -> 0.288 | Hurt |
| add key-sentence extraction (ablation) | KEQA on QA benchmarks · NQ, StrategyQA, HotpotQA, 2WikiMultihopQA | F1/ACC: 0.483 -> 0.475; 0.680 -> 0.678; 0.405 -> 0.397; 0.326 -> 0.316 | Mixed |
Surprising findings
- Vanilla RAG does not consistently beat non-RAG baselines, especially on simple tasks, which the authors attribute to retrieval noise.
- Removing the quiz-based knowledge evaluation causes the largest performance drop among ablations.
Limitations
- Evaluation uses 500 sampled instances per dataset, and the KEQA latency is higher than IRCoT and FLARE.
- The semantic discriminator (Llama-2-7b) is less accurate than GPT-3.5 as an ideal discriminator, at 0.611 vs. the ideal on NQ.
Experimental setup
Datasets: NaturalQuestions StrategyQA HotpotQA 2WikiMultihopQA Wikipedia dump (Dec 2018)
Benchmarks: NaturalQuestions StrategyQA HotpotQA 2WikiMultihopQA
Metrics: F1 exact match accuracy
Baselines: Vanilla GPT-3.5 Zero-shot CoT Few-shot CoT Vanilla RAG ReAct IRCoT FLARE Self-Rag SearChain Rowen SlimPLM
Builds on: Self-consistency FLARE Least-to-most prompting Chain-of-thought
Closest papers by meaning
- SuRe: Summarizing Retrievals using Answer Candidates for Open-domain QA of LLMs (ICLR 2024 · similarity 0.89)
- Promoting Knowledge Base Question Answering by Directing LLMs to Generate Task-relevant Logical Forms (AAAI 2025 · similarity 0.88)
- Evaluating LLM Reasoning in the Operations Research Domain with ORQA (AAAI 2025 · similarity 0.88)
- QuestBench: Can LLMs ask the right question to acquire information in reasoning tasks? (NeurIPS 2025 · similarity 0.88)
- Cooperative Retrieval-Augmented Generation for Question Answering: Mutual Information Exchange and Ranking by Contrasting Layers (NeurIPS 2025 · similarity 0.88)
- Perception of Knowledge Boundary for Large Language Models through Semi-open-ended Question Answering (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.