TCAR-Gen: Temporal Graph Retrieval with Evidence Fusion for Knowledge-Grounded Generation

Researchers introduce TCAR-Gen, a retrieval-augmented generation framework that improves temporal reasoning and evidence fusion for answering complex questions over historical narratives. The system outperforms existing RAG approaches on the Victorian Crime Diaries benchmark by combining graph neural networks with temporal modeling and chain-of-trees reasoning.

TCAR-Gen addresses a fundamental limitation in retrieval-augmented generation systems: the inability to reason effectively across temporal sequences and synthesize multiple evidence sources coherently. This research tackles a genuinely difficult problem in knowledge-grounded AI—moving beyond simple document retrieval toward contextual understanding of how events and facts relate across time. The Victorian Crime Diaries benchmark provides a rigorous testing ground where questions span multiple event types including multi-hop reasoning and counterfactual scenarios, forcing systems to understand not just what happened, but when and why.

The framework's technical innovation centers on query-conditioned graph neural networks that adapt retrieval based on question semantics, combined with explicit temporal penalty mechanisms that prevent anachronistic evidence fusion. This represents a maturation of RAG architectures beyond keyword matching toward genuine semantic understanding. The comparison against GraphRAG variants shows that temporal modeling isn't a peripheral enhancement but a core requirement for complex reasoning tasks.

The cross-model evaluation revealing degradation at smaller model scales carries practical implications for deployment. While larger models maintain robust retrieval coverage even with TCAR-Gen's sophisticated architecture, smaller models struggle with generation quality—suggesting that this approach works best within well-resourced inference environments. This scalability limitation matters for democratizing access to advanced reasoning capabilities.

Future work should focus on whether these temporal reasoning improvements transfer to other domains beyond criminal case narratives, and whether the framework can operate efficiently at edge scales. The research validates that explicit temporal modeling and multi-branch fusion are necessary for reasoning-intensive QA systems, advancing the field beyond naive retrieval approaches.

• →TCAR-Gen achieves 37.38% Recall@5 on temporal reasoning tasks by combining graph neural networks with explicit temporal modeling mechanisms.
• →Ablation studies confirm that context graphs, temporal penalties, and query conditioning are individually critical components for performance gains.
• →Performance degrades significantly on smaller language models (TinyLlama 1.1B), limiting practical deployment in resource-constrained environments.
• →The framework outperforms Vanilla RAG, Temporal RAG, and GraphRAG variants across seven diverse query types including multi-hop and counterfactual reasoning.
• →Research validates that temporal reasoning requires architectural innovations beyond standard retrieval augmentation for knowledge-grounded generation tasks.