Technical Summary: Controllable Logical Hypothesis Generation for Abductive Reasoning in Knowledge Graphs

Problem Definition

Abductive reasoning in Knowledge Graphs (KGs) aims to generate the most plausible logical hypotheses that explain a set of observed entities. While existing methods (e.g., AbductiveKGR) have demonstrated the capability to generate hypotheses, they suffer from a critical lack of controllability. In large-scale knowledge graphs, a single observation often leads to numerous plausible but redundant, irrelevant, or overly complex hypotheses. This limits practical utility in domains such as clinical diagnosis or scientific discovery, where users require explanations tailored to specific semantic aspects (e.g., pathology vs. treatment) or specific levels of structural complexity.

The work identifies two main challenges in implementing controllable generation for long, complex logical hypotheses:

1. Collapse of the Hypothesis Space: As the length and complexity of logical hypotheses increase, the number of valid, plausible candidates drops sharply. This makes it difficult for models to learn complex structural patterns and to effectively apply controls over structural complexity.
2. Over-sensitivity of Hypothesis Reward: Previous approaches relied heavily on the Jaccard score as a reward signal. However, this metric is highly sensitive; minor structural errors in a generated hypothesis can lead to a drastic drop in the reward value, causing training instability and steering the model in incorrect directions during Reinforcement Learning.

Methodology: CtrlHGen

The authors propose CtrlHGen (Controllable Logical Hypothesis Generation), a framework trained in a two-stage paradigm (supervised learning followed by Reinforcement Learning) to address these challenges.

1. Data Augmentation via Sub-logical Decomposition

To mitigate the collapse of the hypothesis space, the authors introduce a data augmentation strategy based on sub-logical decomposition.

• Process: Complex hypotheses are recursively decomposed into simpler sub-hypotheses based on identifiable sub-logical patterns.
• Mechanism: For a complex hypothesis $H$ and an observation $O$, the system generates sub-hypotheses $H_{sub}$ and derives corresponding sub-observations $O_{sub}$ by executing them in the knowledge graph.
• Advantage: This creates a dense set of observation-hypothesis pairs where complex structures are built from simpler, semantically aligned components. This enables the model to learn complex logical structures step-by-step by leveraging patterns found in simpler sub-components, thereby effectively expanding the available training space for long logical chains.

2. Framework for Controllable Generation

The generative model is a Transformer-based decoder trained to generate hypothesis sequences guided by an observation $O$ and a control condition $C$.

• Control Conditions:
  • Semantic Focus: Restricting generation to specific entities or relations (e.g., "Focus on treatment").
  • Structural Constraints: Enforcing specific logical patterns (e.g., intersection, union, negation), limiting the number of involved entities, or restricting the number of relations used.

3. Reinforcement Learning with Smoothed Rewards

To address the over-sensitivity of hypothesis rewards, the authors refine the reward function used in the Reinforcement Learning (RL) step (specifically using Group Relative Policy Optimization, GRPO).

• Semantic Alignment: Instead of relying solely on the Jaccard score, the semantic reward $R_{sem}$ is a weighted combination of Jaccard, Dice, and Overlap coefficients. The inclusion of Dice and Overlap ensures smoother gradients and greater tolerance for minor deviations, stabilizing training.
• Condition Adherence: A binary reward $R_{cond}$ is introduced to explicitly encourage the model to satisfy user-specified control constraints (e.g., correct logical pattern or entity count).
• Objective: The total reward balances semantic similarity with constraint adherence, ensuring that generated hypotheses are both accurate and controllable.

Main Contributions

1. Introduction of the Task: This work is the first to formally introduce the task of controllable abductive reasoning in knowledge graphs, defining the capability to control both semantic content and structural complexity.
2. Sub-logical Decomposition: A novel data augmentation strategy that addresses the hypothesis space collapse problem, enabling models to learn complex logical structures by leveraging simpler sub-components.
3. Robust Reward Design: A refined reward function integrating Dice and Overlap coefficients to mitigate reward over-sensitivity, alongside a condition-adherence reward to ensure strict compliance with control constraints.
4. Empirical Validation: Comprehensive experiments on three benchmark datasets (DBpedia50, WN18RR, FB15k-237) demonstrating superior performance over baselines.

Experimental Results

The model was evaluated on three datasets under five control conditions (logical pattern, number of relations, number of entities, specific entity, specific relation).

• Semantic Similarity: CtrlHGen significantly outperformed the uncontrolled baseline (AbductiveKGR) and various Large Language Models (LLMs), including GPT-4o, Kimi K2, Grok-3, and DeepSeek-V3. For example, on FB15k-237 under the "specific entity" condition, CtrlHGen achieved a Jaccard score of 64.3, while the best LLM (GPT5 in thinking mode) reached only 24.1.
• Condition Adherence: The model achieved high adherence rates (often >90%) for structural and semantic constraints, whereas LLMs faced significant difficulties with structural constraints (e.g., specific logical patterns), often failing to adhere to the required syntax or logical depth.
• Ablation Studies:
  • Removing sub-logical decomposition led to a significant drop in Jaccard scores, particularly for complex patterns involving disjunctions and negations.
  • Removing Dice/Overlap rewards resulted in performance degradation, confirming the necessity of smoothed rewards for stable RL training.
  • Removing the condition-adherence reward slightly improved semantic similarity but drastically reduced the model's ability to follow control constraints.

Significance and Claims

The work claims that CtrlHGen successfully bridges the gap between the theoretical capability of abductive reasoning and its practical application. By introducing controllability, the framework enables users to filter the vast hypothesis space to generate explanations that are not only plausible but also relevant to specific user intents (e.g., focusing on a specific aspect of a disease or limiting complexity).

The authors emphasize that while LLMs possess strong semantic capabilities, they struggle with the structured reasoning required for abductive tasks over knowledge graphs, especially when strict structural constraints are imposed. CtrlHGen demonstrates that a specialized neuro-symbolic approach, combining supervised learning, data augmentation, and tailored Reinforcement Learning, is necessary to achieve high-quality, controllable generation of logical hypotheses. The work positions controllable abductive reasoning as a crucial step toward making KG-based inference systems practical for real-world, high-stakes domains.