IC-487Including sub-questions in the prompt improves LLM performance on multi-hop QA tasks

Jian Wu, Linyi Yang, Zhen Wang, Manabu Okumura, Yue Zhang

SourceCofCA: A STEP-WISE Counterfactual Multi-hop QA benchmark

An ablation study removes sub-questions from the prompt and asks LLMs to produce only the final answer, under a chain-of-thought setting. All models show lower EM and F1 without sub-questions. For GPT-4 on 2-hop, EM drops from 53.2 to 43.8 and F1 from 67.7 to 65.2; for GPT-3.5, EM drops from 40.4 to 34.3 and F1 from 56.9 to 51.3. The effect is consistent across all nine models and all hop counts, with the largest absolute gains seen in weaker models such as Gemini-Pro (2-hop EM: 25.2 without vs 34.6 with sub-questions).

Evidence
correlational
Key metric
GPT-4 2-hop: EM 43.8±0.2 (w/o) vs 53.2±0.5 (w/), F1 65.2±0.3 vs 67.7±0.6; GPT-3.5 2-hop: EM 34.3±0.2 vs 40.4±0.5, F1 51.3±0.1 vs 56.9±0.4; Gemini-Pro 2-hop: EM 25.2±0.5 vs 34.6±0.5
Caveat
The ablation is conducted under a chain-of-thought prompting setting. The sub-questions are provided as part of the prompt, so the improvement may reflect the additional information rather than a change in the model's internal reasoning.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, GPT-3.5 / ChatGPT-3.5, Gemini Pro, GPT-3 / GPT base text-davinci-003, O1 / OpenAI-o1-preview o1-preview
Methods
Chain-of-Thought prompting / Chain-of-Thought (CoT) prompting / CoT prompting / Few-shot Chain of Thought / Few-shot CoT / Chain-of-Thought (CoT-S) / CoT-bag prompting / Few-shot CoT prompting / Rationale prompting / Zero-shot CoT prompting / Wei et al. (2022) chain-of-thought prompting [primary]
Related work
Zero-shot Chain-of-Thought / Wei et al. 2022 (Chain of Thought) / Zero-shot chain-of-thought prompting [builds-on]
Related findings
IC-485, IC-486, IC-488
Extraction
automatic-extraction