Research map / Large language models / Retrieval and knowledge
Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning
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.
Proposes RoG, a planning-retrieval-reasoning framework where an instruction-tuned LLaMA2-Chat-7B generates KG-grounded relation paths, retrieves reasoning paths from Freebase, and reasons over them for KGQA.
Problem
LLMs lack up-to-date knowledge and hallucinate during reasoning, and existing KG-based LLM methods use KGs only as fact stores, ignoring their structure, while semantic parsing yields non-executable queries.
Approach
Treat relation paths grounded in the KG as faithful plans: the LLM generates relation paths, a constrained BFS retrieves matching reasoning paths from the KG, and the LLM reasons over them; the planning module can also be plugged into other LLMs at inference.
Key results
- RoG achieves 85.7 Hits@1 on WebQSP (vs. DECAF 82.1) and 62.6 Hits@1 / 56.2 F1 on CWQ (vs. UniKGQA 51.2 / 49.1).
- Adding the RoG planning module to LLaMA2-Chat-7B raises WebQSP Hits@1 from 64.37 to 74.20 and CWQ Hits@1 from 34.60 to 56.41 without retraining.
- Ablation: removing the planning module drops WebQSP F1 from 70.81 to 49.69; random plans drop it to 35.24.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add KG-grounded relation-path planning (ablation) | RoG · WebQSP | F1: 49.69 -> 70.81 | Helped |
| add KG-grounded relation-path planning (ablation) | RoG · CWQ | F1: 33.76 -> 56.17 | Helped |
| add LLM reasoning over retrieved paths (ablation) | RoG · WebQSP | F1: 49.56 -> 70.81 | Helped |
| add LLM reasoning over retrieved paths (ablation) | RoG · CWQ | F1: 22.26 -> 56.17 | Helped |
| replace generated plans with random plans (ablation) | RoG · WebQSP | F1: 70.81 -> 35.24 | Hurt |
| replace generated plans with random plans (ablation) | RoG · CWQ | F1: 56.17 -> 37.64 | Hurt |
| replace LLM reasoning with majority voting over retrieved paths (ablation) | RoG · WebQSP | F1: 70.81 -> 47.96 | Hurt |
| replace LLM reasoning with majority voting over retrieved paths (ablation) | RoG · CWQ | F1: 56.17 -> 26.52 | Hurt |
| integrate RoG planning with ChatGPT | ChatGPT · WebQSP | Hits@1: 66.77 -> 81.51 | Helped |
| integrate RoG planning with Alpaca-7B | Alpaca-7B · CWQ | Hits@1: 27.44 -> 44.04 | Helped |
| transfer RoG training from Freebase to Wiki-Movies rather than train from scratch | RoG · MetaQA-3hop | F1: 41.32 -> 50.68 | Helped |
Surprising findings
- On WebQSP, RoG's F1 (70.8) is below DECAF's (78.8) despite higher Hits@1, so it does not win on every metric.
- Random relation-path plans perform worse than removing the planning module entirely (F1 35.24 vs 49.69 on WebQSP).
Limitations
- Ablation and comparison on WebQSP show F1 lower than DECAF, indicating precision/recall trade-offs with noisy retrieved paths.
- Relies on entity linking of question and answer entities to the KG and on a Freebase subgraph, limiting applicability to unlinked questions.
- Several baseline numbers are copied from prior papers rather than re-run, and ChatGPT results depend on a closed model version.
Experimental setup
Datasets: WebQSP CWQ Freebase MetaQA-3hop
Benchmarks: WebQSP CWQ MetaQA-3hop
Metrics: Hits@1 F1 Precision Recall
Baselines: KV-Mem EmbedKGQA NSM TransferNet KGT5 GraftNet PullNet SR+NSM SR+NSM+E2E SPARQL QGG ArcaneQA
Builds on: plan-and-solve prompting chain-of-thought Fusion-in-Decoder evidence lower bound (variational inference)
Closest papers by meaning
- Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models (ICML 2025 · similarity 0.94)
- Plan-on-Graph: Self-Correcting Adaptive Planning of Large Language Model on Knowledge Graphs (NeurIPS 2024 · similarity 0.94)
- Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph (ICLR 2024 · similarity 0.93)
- Debate on Graph: A Flexible and Reliable Reasoning Framework for Large Language Models (AAAI 2025 · similarity 0.92)
- Talk like a Graph: Encoding Graphs for Large Language Models (ICLR 2024 · similarity 0.91)
- Topology of Reasoning: Understanding Large Reasoning Models through Reasoning Graph Properties (NeurIPS 2025 · similarity 0.91)
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.