Reproduce “RADAR: Reasoning as Discrimination with Aligned Representations for LLM-based Knowledge Graph Reasoning” on OpenResearch

About this paper

Bo Xue, Yuan Jin, Luoyi Fu, Jiaxin Ding, Xinbing Wang

RADAR, a framework developed at Shanghai Jiao Tong University, reformulates LLM-based Knowledge Graph Reasoning (KGR) as a discriminative entity selection task to address co-occurrence shortcuts and hallucinations. This approach, which aligns supervision, optimization, and representation-based inference, achieved a 5.3% relative improvement in MRR for link prediction on FB15K-237N and an average 6.1% relative accuracy gain in triple classification over LLM-only baselines.

Key insights

• Shifting KGR from generative pattern matching to discriminative entity selection is critical for overcoming co-occurrence shortcuts and achieving robust generalization.
• Aligning the entire KGR pipeline—supervision, optimization, and inference—around discrete entity separation induces relational semantics as discriminative structures in the LLM's representations.
• Representation-based inference, distinct from autoregressive decoding, more effectively utilizes learned internal relational information and mitigates generation-induced hallucinations.

Summary by alphaXiv.