In September 2026, the TRANS Research Group published a study on temporal causal reasoning in large language models. Inspired by document layout reordering, this work proposes a paradigm shift from discrete narratives to causal topologies and introduces two architectures: the training-free neuro-symbolic TCR-Agent and the end-to-end TCR-Former, which internalizes causal-topological attention masks and temporal span biases into Transformer attention. On the TempoCausal benchmark, TCR-Former achieves 16.5% higher accuracy than SOTA methods and, with only about 8.6B parameters, outperforms GPT-5.4 and DeepSeek-V3.2, while revealing and bridging the structure-reasoning disconnection in existing methods.
CITATION: Yang, Z., Luo, W., Chen, Y., & Ou, S. (2026). From document layout to causal topology: An end-to-end architecture for temporal causal reasoning. arXiv preprint arXiv:2610.34094. https://doi.org/10.48550/arXiv.2610.34094.