IC-485LLMs show a significant performance gap between Wikipedia-based factual multi-hop QA and counterfactual multi-hop QA, indicating reliance on memorized knowledge rather than reasoning from context
The paper evaluates nine LLMs on both standard Wikipedia-based factual multi-hop QA datasets (HotpotQA, 2WikiMultiHopQA, MuSiQue) and the newly constructed COFCA counterfactual benchmark, where key entities and facts are replaced so the answers cannot be recalled from memory. All models show a substantial drop in EM and F1 scores on the counterfactual data. For example, GPT-4 drops from EM 69.9 on HotpotQA to 53.1 on COFCA 2-hop, and GPT-3.5 drops from 58.6 to 40.6. A progressive annotation experiment (original passage, stage-1 keyword replacement, stage-2 full paraphrase) confirms the drop is due to data contamination: GPT-4's EM falls from 71.35 to 65.62 to 63.49 across the three stages.
Evidence
correlational
Key metric
GPT-4: HotpotQA EM 69.9±1.5, F1 82.3±1.3 vs COFCA 2-hop EM 53.1±1.5, F1 62.8±1.3; GPT-3.5: HotpotQA EM 58.6±0.9 vs COFCA 2-hop EM 40.6±0.7; GPT-4 annotation stages: EM 71.35 → 65.62 → 63.49
Caveat
The COFCA dataset contains 900 counterfactual QA pairs, which the authors note is still limited in size. EM scoring may undercount correct answers due to alias/abbreviation issues, though a GPT-4-turbo partial-match metric confirms the same gap.