Research map / Large language models / Retrieval and knowledge
Deliberation on Priors: Trustworthy Reasoning of Large Language Models on Knowledge Graphs
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 Deliberation over Priors (DP), a KG-based RAG framework that distills structural relation-path priors into LLaMA3.1-8B via SFT plus KTO, then uses constraint-based introspection with backtracking to verify reasoning paths before answer generation.
Problem
LLM reasoning over knowledge graphs for KGQA is unfaithful and unreliable because existing methods underuse KG structural priors (relation paths) and explicit/implicit constraints.
Approach
Distill structural priors via weak supervision from shortest KG paths (SFT + KTO with perturbed negative paths) into a path generator, and verify instantiated reasoning paths against five predefined question constraint types with feedback-driven backtracking.
Key results
- DP with GPT-4.1 reaches 87.2 H, 75.8 H@1, 69.4 F1 on CWQ, versus 56.0/53.0/48.9 for GPT-4.1 without KG augmentation.
- DP surpasses LightPROF by 16.5% H@1 on CWQ, and claims a 13% H@1 improvement on ComplexWebQuestions.
- Removing introspection causes the largest drop among ablations (H@1 82.0 vs 86.7 on WebQSP with GPT-4.1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add KTO preference optimization to the path generator (ablation) | DP with GPT-4.1 · WebQSP | H@1: 84.7 -> 86.7 | Helped |
| Add path truncation perturbations to preference data (ablation) | DP with GPT-4.1 · WebQSP | H@1: 84.8 -> 86.7 | Helped |
| Add entity-path swapping perturbations to preference data (ablation) | DP with GPT-4.1 · WebQSP | H@1: 85.8 -> 86.7 | Helped |
| Add relation deletion perturbations to preference data (ablation) | DP with GPT-4.1 · WebQSP | H@1: 84.8 -> 86.7 | Helped |
| Add reasoning introspection (ablation) | DP with GPT-4.1 · WebQSP | H@1: 82.0 -> 86.7 | Helped |
| Use predefined constraints instead of LLM-induced constraints (ablation) | DP with GPT-4.1 · WebQSP | H: 87.8 -> 90.6 | Helped |
| Add verification feedback for constraint violations (ablation) | DP with GPT-4.1 · WebQSP | H@1: 83.0 -> 86.7 | Helped |
| Add DP deliberation and knowledge augmentation to GPT-4.1 | GPT-4.1 · WebQSP | F1: 54.6 -> 80.1 | Helped |
| Add DP deliberation and knowledge augmentation to GPT-4.1 | GPT-4.1 · CWQ | F1: 48.9 -> 69.4 | Helped |
| Increase the number of exemplars | DP · CWQ and WebQSP | H@1 | Hurt |
| Use one-to-many question-to-relation mappings and leverage priors for path generation | DP path generation · WebQSP | F1: +29.3% | Helped |
Surprising findings
- Constraint verification with LLM-generated constraints instead of predefined ones lowers WebQSP H by 2.8%, showing manual constraint definitions matter.
- DP uses only about 2.5-2.9 LLM calls per question, far fewer than ToG or PoG.
Limitations
- Constraint predefinition still requires human effort for vertical domains.
- Path generation errors and reasoning errors remain, with LLMs sometimes relying on parametric knowledge.
- Evaluation uses sampled subsets (500 WebQSP/CWQ questions, 600 MetaQA) due to computational constraints.
Experimental setup
Datasets: WebQSP ComplexWebQuestions MetaQA
Benchmarks: WebQSP ComplexWebQuestions MetaQA
Metrics: Hit Hits@1 F1
Baselines: EmbedKGQA TransferNet UniKGQA RoG AMAR GNN-RAG PoG ToG Readi DoG Interactive-KBQA LightPROF
Builds on: KTO RoG ToG DoG LoRA
Closest papers by meaning
- Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning (ICLR 2024 · similarity 0.90)
- Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.89)
- Debate on Graph: A Flexible and Reliable Reasoning Framework for Large Language Models (AAAI 2025 · similarity 0.89)
- Towards Trustworthy Knowledge Graph Reasoning: An Uncertainty Aware Perspective (AAAI 2025 · similarity 0.89)
- Reasoning of Large Language Models over Knowledge Graphs with Super-Relations (ICLR 2025 · similarity 0.89)
- Harnessing Large Language Models for Knowledge Graph Question Answering via Adaptive Multi-Aspect Retrieval-Augmentation (AAAI 2025 · 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.