Research map / Large language models / Reasoning and chain of thought
Chain-of-Thought Improves Text Generation with Citations 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.
Applies chain-of-thought prompting modeled on a human reasoning process, plus a Citation Insurance Mechanism that adds missing citations, to improve LLM text generation with citations on the ALCE benchmark.
Problem
LLMs given retrieved documents often synthesize answers incorrectly and fail to cite documents or leave statements uncited, giving poor correctness and citation quality.
Approach
Use a 1-shot CoT prompt that mirrors a human thought process (find relevant documents, write per-document cited statements, merge, rank and concatenate), then apply a post-hoc Citation Insurance Mechanism that detects uncited statements and adds citations via the same LLM or a GTR retriever.
Key results
- CoT prompting improves citation quality and correctness across six open-source LLMs on ALCE; e.g., LLaMA-2-70B-Chat ORACLE ASQA citation recall 73.5 vs. 69.8-ish baseline (ablation: removing CoT drops LLaMA-2-70B ASQA citation recall from 70.2 to 63.3).
- The Citation Insurance Mechanism adds about +2.4 citation recall and +2.2 precision on LLaMA-2-13B on ASQA, with smaller gains (+1.1 recall, +1.9 precision) on LLaMA-2-70B.
- LLaMA-2-70B-Chat with CoT and CIM achieves correctness and citation quality comparable to ChatGPT-based ALCE baselines, e.g., ASQA ORACLE correctness 46.2 vs. 48.9 for ChatGPT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CoT prompting (ablation) | LLaMA-2-13B VANILLA on ASQA · ASQA | Fluency: 76.3 -> 82.0 | Helped |
| Add CoT prompting (ablation) | LLaMA-2-13B VANILLA on ASQA · ASQA | Correctness: 33.2 -> 37.4 | Helped |
| Add CoT prompting (ablation) | LLaMA-2-13B VANILLA on ASQA · ASQA | Citation Recall: 62.8 -> 66.5 | Helped |
| Add CoT prompting (ablation) | LLaMA-2-13B VANILLA on ASQA · ASQA | Citation Precision: 54.4 -> 59.2 | Helped |
| Add LLM-based Citation Insurance Mechanism (ablation) | LLaMA-2-13B VANILLA on ASQA · ASQA | Fluency: 80.4 -> 82.0 | Helped |
| Add LLM-based Citation Insurance Mechanism (ablation) | LLaMA-2-13B VANILLA on ASQA · ASQA | Correctness: 36.8 -> 37.4 | Helped |
| Add LLM-based Citation Insurance Mechanism (ablation) | LLaMA-2-13B VANILLA on ASQA · ASQA | Citation Recall: 64.1 -> 66.5 | Helped |
| Add LLM-based Citation Insurance Mechanism (ablation) | LLaMA-2-13B VANILLA on ASQA · ASQA | Citation Precision: 57.0 -> 59.2 | Helped |
| Add CoT prompting (ablation) | LLaMA-2-70B VANILLA on ASQA · ASQA | Citation Recall: 63.3 -> 70.2 | Helped |
| Add CoT prompting (ablation) | LLaMA-2-70B VANILLA on ASQA · ASQA | Citation Precision: 58.9 -> 69.8 | Helped |
| Add LLM-based Citation Insurance Mechanism (ablation) | LLaMA-2-70B VANILLA on ASQA · ASQA | Citation Recall: 69.1 -> 70.2 | Helped |
| Add LLM-based Citation Insurance Mechanism (ablation) | LLaMA-2-70B VANILLA on ASQA · ASQA | Citation Precision: 67.9 -> 69.8 | Helped |
Surprising findings
- The LLM-based CIM can change fluency and correctness, since it sometimes completes statements that were not ended normally.
- Results vary substantially across semantically similar descriptions of the thought process, indicating prompt sensitivity.
Limitations
- CIM may add incorrect citations, for example when all retrieved documents are irrelevant.
- Results are sensitive to prompt descriptions and temperature.
- Evaluated only on the ALCE benchmark with three datasets.
Experimental setup
Datasets: ASQA QAMPARI ELI5
Benchmarks: ALCE ASQA ALCE QAMPARI ALCE ELI5
Metrics: fluency correctness citation recall citation precision
Baselines: ALCE LLM-based prompting baselines ChatGPT VANILLA ChatGPT SUMM ChatGPT SNIPPET ChatGPT ORACLE
Builds on: chain-of-thought prompting ALCE zero-shot CoT self-consistency
Closest papers by meaning
- Chain-of-Knowledge: Grounding Large Language Models via Dynamic Knowledge Adapting over Heterogeneous Sources (ICLR 2024 · similarity 0.87)
- WikiContradict: A Benchmark for Evaluating LLMs on Real-World Knowledge Conflicts from Wikipedia (NeurIPS 2024 · similarity 0.86)
- Citations and Trust in LLM Generated Responses (AAAI 2025 · similarity 0.86)
- Generating Chain-of-Thoughts with a Pairwise-Comparison Approach to Searching for the Most Promising Intermediate Thought (ICML 2024 · similarity 0.85)
- Algorithm of Thoughts: Enhancing Exploration of Ideas in Large Language Models (ICML 2024 · similarity 0.85)
- Towards Verifiable Text Generation with Generative Agent (AAAI 2025 · similarity 0.85)
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.