Research map / Large language models / Retrieval and knowledge
Statistical Knowledge Assessment for Large Language Models
Qingxiu Dong, Jingjing Xu, Lingpeng Kong, Zhifang Sui, Lei Li
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 |
|---|---|---|---|
| instruction-tune LLaMA-13B on Alpaca data | LLaMA-13B · KaRR T-REx | KaRR score: 13.86 -> 8.24 | Hurt |
| instruction-tune LLaMA-13B on Vicuna data from a stronger model | LLaMA-13B · KaRR T-REx | KaRR score: 13.86 -> 19.50 | Improved |
| instruction-tune LLaMA-13B on WizardLM data | LLaMA-13B · KaRR T-REx | KaRR score: 13.86 -> 16.90 | Improved |
| scale LLaMA from 7B to 13B to 65B | LLaMA · KaRR T-REx | KaRR score: 12.37 -> 13.86 -> 14.56 | Improved |
| replace T5-3B with GPT-NEO of similar size | pretrained small LMs · KaRR T-REx | KaRR score: 9.52 -> 13.44 | Improved |
| average probability over k paraphrased prompts (K-Prompts) instead of latent-variable ratio (ablation) | KaRR · Human eval GPT2-XL | Kendall tau: 0.43 -> 0.32 | Hurt |
| use KaRR instead of LAMA@1 for detecting unknown facts | knowledge assessment · Human eval GPT2-XL | recall: 83.25% -> 95.18% | Improved |
| use KaRR instead of K-Prompts under paraphrased templates (ablation) | knowledge assessment · Paraphrased templates | variance: 2.34 -> 0.67 | Improved |
| use KaRR instead of LAMA@10 on false synthetic facts | knowledge assessment · Spurious correlation | SP: 64.29 -> 1.94 | Improved |
| use KaRR instead of K-Prompts on spurious positives | knowledge assessment · Spurious correlation | SP: 0.00 vs 1.94 | No clear effect |
| increase sampling number K (ablation) | KaRR GPT2-XL · Variance and human correlation | variance and correlation | Improved |
| scale OPT from 350M to 2.7B to 175B | OPT · Best/worst-known relations | KaRR score | Improved |
Closest papers by meaning
- Long-form factuality in large language models (NeurIPS 2024 · similarity 0.91)
- KoLA: Carefully Benchmarking World Knowledge of Large Language Models (ICLR 2024 · similarity 0.89)
- RepLiQA: A Question-Answering Dataset for Benchmarking LLMs on Unseen Reference Content (NeurIPS 2024 · similarity 0.89)
- Rethinking Generative Large Language Model Evaluation for Semantic Comprehension (ICML 2024 · similarity 0.89)
- SuRe: Summarizing Retrievals using Answer Candidates for Open-domain QA of LLMs (ICLR 2024 · similarity 0.89)
- Improving Factuality and Reasoning in Language Models through Multiagent Debate (ICML 2024 · similarity 0.89)
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.
