Title: TCAR-Gen: Leveraging Evidence Fusion in Temporal Graph Retrieval for Knowledge-Grounded Generation

Abstract:

Answering complex questions regarding historical criminal case narratives presents significant challenges for retrieval-augmented generation (RAG) systems, particularly in the areas of temporal reasoning and evidence fusion. Current methodologies often fall short by either retrieving information without regard for query semantics or by failing to coherently integrate multiple sources of evidence. To address these limitations, we introduce Temporal Context Augmented Retrieval Generation (TCAR-Gen). This framework grounds answer generation in retrieved evidence by integrating query-conditioned graph neural networks, temporal evidence fusion, and chain-of-trees reasoning.

When evaluated on the Victorian Crime Diaries benchmark, TCAR-Gen attained a Recall@5 score of 0.3738. This performance surpasses that of Vanilla RAG, Temporal RAG, GraphRAG-C, and GraphRAG-T across seven distinct query categories, which include multi-hop reasoning and counterfactual inquiries. Our ablation studies highlight the importance of three specific elements: the context graph, the temporal penalty mechanism, and query conditioning.

Furthermore, cross-model evaluations spanning five language models—from GPT-OSS 20B down to TinyLlama 1.1B—indicate that while TCAR-Gen preserves robust retrieval coverage at smaller model scales, the quality of the generated output declines significantly as model capacity decreases. Ultimately, our findings suggest that explicit temporal modeling and multi-branch evidence fusion are vital for conducting faithful, reasoning-intensive question answering over knowledge-grounded corpora.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC