Large language models generally lack deep causal reasoning ability in temporal reasoning. Inspired by document layout reordering, this work proposes a paradigm shift that converts discrete narratives into causal topologies.
Two architectures are built. The first is a training‑free neuro‑symbolic framework called Temporal Causal Reasoning Agent (TCR‑Agent). By instantiating an external causal graph and using a symbolic intervention engine, it achieves interpretable causal deduction and counterfactual truncation.
The second is an end‑to‑end model named Temporal Causal Reasoning Former (TCR‑Former). It internalizes the same mechanisms within the Transformer attention space, featuring a Causal‑Topological Attention Mask and temporal span biases.
Experiments on the TempoCausal benchmark show that TCR‑Agent outperforms all training‑free baselines comprehensively. TCR‑Former improves accuracy by 16.5% over state‑of‑the‑art methods while using only about 8.6 B parameters, surpassing large models such as GPT‑5.4 and DeepSeek‑V3.2.
Further analysis reveals a structure‑reasoning disconnection in existing approaches: external symbolic constraints and chain‑of‑thought reasoning can increase the number of reasoning steps but fail to enhance deep causal deductive capability. By internalizing temporal causal reasoning into the model, TCR‑Former effectively bridges this gap, offering an efficient and reliable foundation for complex reasoning scenarios.
Review