Research map / Large language models / Retrieval and knowledge
Controllable Logical Hypothesis Generation for Abductive Reasoning in Knowledge Graphs
Yisen Gao, Jiaxin Bai, Tianshi Zheng, Ziwei Zhang, Qingyun Sun, Xingcheng Fu, Jianxin Li, Yangqiu Song
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add reinforcement learning on top of supervised training (ablation) | CtrlHGen · WN18RR (pattern condition) | condition adherence accuracy: 81.5 -> 93.5 | Improved |
| add reinforcement learning on top of supervised training (ablation) | CtrlHGen · WN18RR (pattern condition) | Jaccard: 71.5 -> 77.0 | Improved |
| add Dice and Overlap rewards alongside Jaccard (ablation) | CtrlHGen · WN18RR (pattern condition) | condition adherence accuracy: 90.3 -> 93.5 | Improved |
| add Dice and Overlap rewards alongside Jaccard (ablation) | CtrlHGen · WN18RR (pattern condition) | Jaccard: 74.8 -> 77.0 | Improved |
| add condition-adherence reward (ablation) | CtrlHGen · WN18RR (pattern condition) | condition adherence accuracy: 68.3 -> 93.5 | Improved |
| add condition-adherence reward (ablation) | CtrlHGen · WN18RR (pattern condition) | Jaccard: 77.5 -> 77.0 | Hurt |
| add sub-logical decomposition data augmentation (ablation) | CtrlHGen · DBpedia50 (13 logical patterns) | Jaccard | Improved |
| add sub-logical decomposition data augmentation (ablation) | CtrlHGen · DBpedia50 (13 logical patterns) | condition adherence accuracy | No clear effect |
| use sub-logic decomposition instead of Logic-Gen augmentation | CtrlHGen · DBpedia50 (pattern condition, SFT only) | condition adherence accuracy: 65.9 -> 73.1 | Improved |
| use sub-logic decomposition instead of Logic-Gen augmentation | CtrlHGen · DBpedia50 (pattern condition, SFT only) | Jaccard: 69.5 -> 70.1 | Improved |
| add control conditions (pattern) to AbductiveKGR | CtrlHGen · FB15k-237 | Jaccard: 61.4 -> 65.5 | Improved |
| replace GPT5(Thinking) with CtrlHGen | CtrlHGen · FB15k-237 (avg of five conditions) | Jaccard: 18.7 -> 64.3 | Improved |
Closest papers by meaning
- Hypothesis Search: Inductive Reasoning with Language Models (ICLR 2024 · similarity 0.86)
- Curriculum Abductive Learning (NeurIPS 2025 · similarity 0.86)
- Abductive Reasoning in Logical Credal Networks (NeurIPS 2024 · similarity 0.86)
- Stability and Generalization Capability of Subgraph Reasoning Models for Inductive Knowledge Graph Completion (ICML 2025 · similarity 0.85)
- Inference-Time Conformal Reasoning with Valid Factuality Control for Large Language Models (ICML 2026 · similarity 0.85)
- SciReason: A Controllable Benchmark for Scientific Reasoning in LLMs (NeurIPS 2026 · 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.
