[Research Update] Unifying Logic Reasoning in Knowledge Graphs with Diffusion Models

I am thrilled to share our latest work, DARK, which presents a unified framework for Deductive and Abductive Reasoning in Knowledge Graphs using Masked Diffusion Models. Congradulation to Yisen Gao

🔍 The Problem: Traditionally, deductive reasoning (finding answers to complex queries) and abductive reasoning (generating hypotheses to explain observations) are treated as isolated tasks. However, they are inherently complementary: deduction validates hypotheses, while abduction uncovers logical patterns.

💡 Our Solution - DARK: We treat logical queries and their conclusions as a unified sequence, leveraging the bidirectional generation capability of Masked Diffusion Models.

• Self-Reflective Denoising: Mirrors human reasoning by generating candidate hypotheses and using deductive verification to select the best one at each step.
• Logic-Exploration RL: A reinforcement learning approach that masks both queries and conclusions to explore novel logical compositions.

📈 Impact: Our extensive experiments show that DARK achieves State-of-the-Art performance on both deductive and abductive tasks across benchmark datasets (FB15k-237, WN18RR, DBpedia50), significantly outperforming existing methods and general LLMs like GPT-4o in complex logical scenarios.

This work bridges the gap between generative diffusion models and structured logical reasoning.

🔗 Read the full paper: https://arxiv.org/abs/2510.11462 🔗 Use our implementation: https://github.com/HKUST-KnowComp/DARK

#ArtificialIntelligence #KnowledgeGraph #Reasoning #DiffusionModels #DeepLearning #AcademicResearch

DARK：当扩散模型遇上知识图谱推理，演绎与溯因终于统一了

大家好，很高兴分享我们最新的工作 DARK (Unifying Deductive and Abductive Reasoning in Knowledge Graphs with Masked Diffusion Model)。这篇论文已被 WWW-2026 接收。向一作同学高意森表示祝贺。

核心痛点： 在知识图谱（KG）推理中，演绎推理（Deductive，给查询找答案）和溯因推理（Abductive，给结果找解释/假设）通常是被割裂研究的。但其实它们互为表里：溯因生成的假设需要演绎来验证，演绎推理也能利用溯因发现深层逻辑。现有的方法（包括很多 LLM）往往难以兼顾两者，或者在处理复杂逻辑（如否定、析取）时表现不佳。

我们的方法 (DARK)： 我们提出了一种基于 Masked Diffusion Model 的统一框架。核心洞察是：将逻辑查询（Query）和结论（Conclusion）视为一个联合序列。

1. 双向建模： 利用扩散模型的双向生成能力，既能从 Query 生成 Conclusion（演绎），也能从 Conclusion 反推 Query（溯因）。
2. 自反思去噪 (Self-reflective Denoising)： 在溯因推理过程中，模型会生成多个候选假设，并利用“演绎推理”来验证哪个假设最符合观测结果，以此指导下一步去噪。这就像人类思考时不断“假设-验证”的过程。
3. 逻辑探索强化学习 (Logic-exploration RL)： 我们设计了一种 RL 策略，同时 Mask 掉查询和结论的一部分，鼓励模型探索更多样、合理的逻辑组合，避免陷入局部最优。

实验结果： 在 FB15k-237, WN18RR, DBpedia50 三大基准数据集上，DARK 在演绎和溯因任务上均取得了 SOTA 的效果。特别是在溯因推理任务中，相比于 AbductiveKGR 和 GPT-4o/DeepSeek-V3 等大模型，我们在处理复杂逻辑结构（如包含否定操作）时优势非常明显。

欢迎大家阅读论文并交流！ 📄 论文链接： https://arxiv.org/abs/2510.11462 💻 代码开源： https://github.com/HKUST-KnowComp/DARK

#知识图谱 #逻辑推理 #扩散模型 #人工智能 #深度学习