Research map / Large language models / Retrieval and knowledge
GIVE: Structured Reasoning of Large Language Models with Knowledge Graph Inspired Veracity Extrapolation
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 GIVE, a training-free KG-guided reasoning method where an LLM links query entities via expert KG neighborhoods and "veracity extrapolation" to build affirmative and counterfactual knowledge before answering.
Problem
LLM reasoning on scientific and niche queries fails when parametric knowledge is insufficient and comprehensive knowledge bases are unavailable; retrieval methods retrieve irrelevant or incomplete triplets that cause hallucination.
Approach
Build entity groups around query entities using KG-similar concepts, induce inner-group and inter-group relations with the LLM's internal knowledge (yes/no/maybe verification), keep affirmative and counterfactual knowledge sets, and generate answers progressively from them, so limited expert triplets inspire divergent reasoning.
Key results
- GPT3.5-turbo + GIVE reaches 53.6% on PubmedQA, 88.2% on BioASQ and 73.4% on ProcessBank, exceeding GPT-4 baselines in several settings.
- On CommonsenseQA with full ConceptNet, GIVE a+c+e reaches 74.7% vs 71.3% for RAG and 69.8% for ToG (GPT3.5-turbo).
- GIVE wins 50.3% of TruthfulQA questions in GPT4o-judged win rate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GIVE structured reasoning with UMLS to GPT3.5-turbo | GPT3.5-turbo · BioASQ | accuracy: 43.5 -> 88.2 | Helped |
| Add GIVE structured reasoning with UMLS to GPT3.5-turbo | GPT3.5-turbo · PubMedQA | accuracy: 48.6 -> 53.6 | Helped |
| Add GIVE structured reasoning to GPT3.5-turbo | GPT3.5-turbo · ProcessBank | accuracy: 70.9 -> 73.9 | Helped |
| Add GIVE with ConceptNet to GPT3.5-turbo | GPT3.5-turbo · CommonsenseQA (100% ConceptNet) | accuracy: 72.2 -> 74.7 | Helped |
| Add affirmative, counterfactual, and expert knowledge to the answer-generation context (ablation) | GIVE a+c · BioASQ | accuracy: 86.1 -> 88.2 | Mixed |
| Add counterfactual knowledge to affirmative knowledge (ablation) | GIVE a · TruthfulQA | win rate | Hurt |
| Add inter-group connections to inner-group connections (ablation) | GIVE · PubmedQA (100 questions) | accuracy: 14 -> 56 | Helped |
| Add inter-group connections to inner-group connections (ablation) | GIVE · BioASQ (100 questions) | accuracy: 50 -> 88 | Helped |
| Add inner-group connections to inter-group connections (ablation) | GIVE · ProcessBank (100 questions) | accuracy: 70 -> 76 | Helped |
| Add inner-group connections to inter-group connections (ablation) | GIVE · CommonsenseQA (100 questions) | accuracy: 74 -> 76 | Helped |
| Increase additional KG concepts per group from 2 to 3 (ablation) | GIVE · Biomedical reasoning | accuracy | Hurt |
| Use triplet-form knowledge prompting instead of text-form prompting (ablation) | GIVE · Biomedical and commonsense QA | accuracy | Mixed |
Surprising findings
- Counterfactual knowledge and expert KG knowledge do not always improve results; on TruthfulQA the affirmative-only variant was most accurate.
- Increasing additional entities per group from 2 to 3 declined performance.
Limitations
- Relies on LLM yes/no/maybe judgments per candidate edge, causing many LLM calls and higher token use than ToG in some settings.
- Performance depends on the presence of expert KG relations; sparse KGs provide little inspiration.
- Evaluated mainly on GPT-family models with small-scale 50-100 question subsets for ablations.
Experimental setup
Datasets: PubmedQA BioASQ ProcessBank CommonsenseQA ConceptNet UMLS TruthfulQA
Benchmarks: PubmedQA BioASQ ProcessBank CommonsenseQA TruthfulQA
Metrics: accuracy win rate
Baselines: I/O prompting CoT prompting RAG ToG GraphRAG
Builds on: chain-of-thought prompting retrieval-augmented generation Think-on-Graph
Closest papers by meaning
- Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph (ICLR 2024 · similarity 0.90)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.89)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.89)
- LightPROF: A Lightweight Reasoning Framework for Large Language Model on Knowledge Graph (AAAI 2025 · similarity 0.88)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.88)
- Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning (ICML 2025 · similarity 0.88)
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.