Research map / Large language models / Retrieval and knowledge
LFQA-E: Carefully Benchmarking Long-form QA Evaluation
Yuchen Fan, Chen Ling, Xin Zhong, Shuo Zhang, Heng Zhou, Yuchen Zhang, Mingyu Liang, Chengxing Xie, Ermo Hua, Zhizhou He, Cheng Huang, Gang Chen and 2 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| scale from Qwen2.5-32B to Qwen2.5-72B as evaluator | Qwen2.5 evaluator · LFQA-E avg | accuracy (%): 60.1 -> 57.1 | Hurt |
| use a reasoning model (o1-mini) instead of GPT-4o | LLM evaluator · LFQA-E avg | accuracy (%): 57.5 -> 60.9 | Improved |
| set evaluator temperature to 0 for o1-mini (ablation) | o1-mini · LFQA-E-zh | accuracy (%): 58.9 -> 5.8 | Hurt |
| set evaluator temperature to 0 for GPT-4o (ablation) | GPT-4o · LFQA-E avg | macro-F1 (%): 44.5 -> 46.7 | Improved |
| remove CoT, ask for final answer only (ablation) | Qwen2.5-32B-Instruct · LFQA-E-en | accuracy (%): 63.5 -> 57.3 | Hurt |
| remove CoT, ask for final answer only (ablation) | DeepSeek-R1 · LFQA-E-zh | accuracy (%): 57.8 -> 60.0 | Mixed |
| drop tie-labeled comparisons (ablation) | GPT-4o · LFQA-E avg | accuracy (%): 57.5 -> 63.9 | Improved |
| add TTRL test-time RL training | Qwen2.5-7B-Instruct · LFQA-E-en | accuracy (%): 60.6 -> 68.2 | Improved |
| add Clip Higher to TTRL (ablation) | Qwen2.5-3B-Instruct · LFQA-E-en | accuracy (%): 63.9 -> 66.5 | Improved |
| add Clip Higher to TTRL (ablation) | Qwen2.5-7B-Instruct · LFQA-E-en | accuracy (%): 68.2 -> 68.6 | No clear effect |
| use structured answer-tag prompt instead of CoT (ablation) | Qwen2.5-3B-Instruct · LFQA-E-en | accuracy (%): 49.6 -> 59.7 | Improved |
| use length as metric instead of reference-based judging | Length · LFQA-E avg | accuracy (%): 60.1 -> 47.7 | Hurt |
Closest papers by meaning
- A Semantic-Sampling Framework for Evaluating Calibration in Open-Ended Question Answering (NeurIPS 2026 · similarity 0.88)
- Improving Automatic VQA Evaluation Using Large Language Models (AAAI 2024 · similarity 0.88)
- RepLiQA: A Question-Answering Dataset for Benchmarking LLMs on Unseen Reference Content (NeurIPS 2024 · similarity 0.88)
- SuRe: Summarizing Retrievals using Answer Candidates for Open-domain QA of LLMs (ICLR 2024 · similarity 0.87)
- CofCA: A STEP-WISE Counterfactual Multi-hop QA benchmark (ICLR 2025 · similarity 0.87)
- Rethinking Generative Large Language Model Evaluation for Semantic Comprehension (ICML 2024 · similarity 0.87)
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.
