Research map / Large language models / Retrieval and knowledge
Long-form factuality in large language models
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.
Introduces LongFact (2,280 GPT-4-generated fact-seeking prompts over 38 topics), SAFE (an LLM agent that splits responses into facts and verifies them with multi-step Google Search), and F1@K (F1 with a recall term based on a human-preferred length K), then benchmarks 13 LLMs.
Problem
Long-form factuality of LLM responses to open-domain fact-seeking prompts lacks comprehensive benchmarks, scalable automated evaluation, and a metric that accounts for both factual precision and recall.
Approach
Use an LLM agent with multi-step Google Search to rate each decomposed, self-contained fact as supported or not supported, and aggregate per-fact labels into F1@K, where recall is measured against a user-preferred number of supported facts K.
Key results
- SAFE agrees with crowdsourced human annotations on 72.0% of 16,011 individual facts, and wins 76% of 100 disagreement cases against human labels.
- SAFE costs $0.19 per model response versus $4.00 for crowdsourced human annotation (more than 20x cheaper).
- GPT-4-Turbo achieves the best F1@64 (95.0) and F1@178 (66.4) among the 13 models benchmarked; larger models generally score higher.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPT-4-Turbo instead of GPT-4 | GPT model family · LongFact-Objects | F1@64: 88.0 -> 95.0 | Helped |
| Use GPT-4 instead of GPT-3.5-Turbo-0125 | GPT model family · LongFact-Objects | F1@64: 79.6 -> 88.0 | Helped |
| Use Gemini Ultra instead of Gemini Pro | Gemini model family · LongFact-Objects | F1@64: 83.7 -> 91.7 | Helped |
| Use ULM-Lit-RLHF instead of ULM-Lit | ULM-Lit model family · LongFact-Objects | F1@64: 31.1 -> 91.0 | Helped |
| Use Claude 3 Sonnet instead of Claude 3 Opus | Claude model family · LongFact-Objects | F1@64: 89.3 -> 89.4 | No clear effect |
| Add a postamble asking for as many specific details as possible | Models answering LongFact-Objects prompts · LongFact-Objects | F1@1 | Hurt |
| Add a postamble asking for as many specific details as possible | Models answering LongFact-Objects prompts · LongFact-Objects | F1@K at higher K | Helped |
| Evaluate on LongFact-Concepts rather than LongFact-Objects | GPT-4-Turbo, Claude 3 Opus, and ULM-Lit-RLHF · LongFact | F1@K | Helped |
| Include recall by using F1@64 instead of precision alone | Model-level factuality evaluation · LongFact and Chatbot Arena | Pearson correlation with Arena ELO: r = 0.502 (p = 0.205) -> r = 0.754 (p = 0.031) | Helped |
Surprising findings
- Claude-3-Sonnet matches Claude-3-Opus on long-form factuality despite being a smaller model.
- Using precision alone did not significantly correlate with Chatbot Arena ELO (r=0.502, p=0.205), whereas F1@64 did (r=0.754, p=0.031).
Limitations
- SAFE relies on GPT-3.5-Turbo, so reasoning errors occur; a stronger model could reduce them.
- Google Search as ground truth may fail for corner cases such as expert-level law and medicine.
- F1@K assumes responses contain no repeated facts, which can be gamed by repetition.
Experimental setup
Datasets: LongFact-Concepts LongFact-Objects FActScore-style human annotations (496 prompt-response pairs, 16,011 facts)
Benchmarks: LongFact-Objects LongFact-Concepts
Metrics: factual precision factual recall R_K F1@K supported/not-supported/irrelevant fact counts agreement with human annotations Pearson correlation with Chatbot Arena ELO
Baselines: FActScore crowdsourced human annotations
Builds on: FActScore chain-of-thought
Closest papers by meaning
- Fine-Tuning Language Models for Factuality (ICLR 2024 · similarity 0.93)
- The FACTS Leaderboard: A Comprehensive Benchmark for Large Language Model Factuality (NeurIPS 2026 · similarity 0.91)
- FactTest: Factuality Testing in Large Language Models with Finite-Sample and Distribution-Free Guarantees (ICML 2025 · similarity 0.89)
- Is Factuality Enhancement a Free Lunch For LLMs? Better Factuality Can Lead to Worse Context-Faithfulness (ICLR 2025 · similarity 0.89)
- Improving Factuality and Reasoning in Language Models through Multiagent Debate (ICML 2024 · similarity 0.89)
- Characterizing Truthfulness in Large Language Model Generations with Local Intrinsic Dimension (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.