CORTEXA
← Browse
openalexInstitutional Repositories DataBase (IRDB)Cited by 0

Multi-dimensional hierarchical temporal alignment for improved temporal commonsense reasoning in large language models

Ge Yan, Hai-Tao Yu, Lei Chao

Benefiting from recent advances in generative AI and large language models (LLMs), current LLM-driven AI systems are rapidly reshaping how we reason, plan, and make decisions. Yet endowing these systems with human-like temporal intelligence (e.g., understanding and comparing events across time) remains largely unresolved. To address this gap, we introduce a novel Multi-dimensional Hierarchical Temporal Alignment framework (denoted as MHTA), of which the core idea is to effectively ground the temporal alignment of an LLM in both event-specific and cross-event real-world temporal commonsense knowledge. Specifically, we first frame the temporal space as a multi-dimensional hierarchical space, where each dimension comprises its own set of temporal units. Building on this formulation, we curate an initial burn-in dataset in which annotations are constructed to reflect real-world human temporal commonsense in accordance with our defined temporal space. Based on the initial burn-in dataset, we propose two complementary fine-tuning objectives: (i) Grounding the temporal alignment in event-specific temporal commonsense knowledge, which is achieved through a novel hierarchy-preserving Wasserstein loss, and (ii) Grounding the temporal alignment in cross-event temporal commonsense knowledge, which is achieved via a margin-adaptive triplet loss. To evaluate the effectiveness of MHTA, we conduct extensive experiments on three representative benchmarks (MCTACO, DurationQA, and TimeDial). For example, on the MCTACO dataset, MHTA improves Mistral-7B by 6.0% in accuracy and 10.3% in exact match, and improves DeepSeek-LLM-7B by 5.4% in accuracy and 9.8% in exact match. Combined with different backbones, MHTA surpasses nearly all baselines across the three benchmarks and exceeds the performance of advanced models such as GPT-4o on MCTACO and DurationQA. These results demonstrate that MHTA provides consistent gains in temporal commonsense reasoning across diverse LLM backbones.

View free PDFSource page

Related papers

openalexInstitutional Repositories DataBase (IRDB)2026-08-16

『通信教育部論集』第29号全文

With Dewey I hold that knowledge, mind, and meaning are part of the same world that they have to do with, and that they are to be studied in the same empirical spirit that animates natural science.There is no place for a priori phi-

View free PDFSource page
openalexInstitutional Repositories DataBase (IRDB)2026-08-01

Season-robust localization for autonomous robot in orchard using two-stage LiDAR points segmentation with sensor fusion

PAN SIYU, Hu Yaohua, Ohya Akihisa, Ayanori Yorozu

A robust cross-season, multi-route localization framework is proposed for agricultural environments with significant seasonal variations. Based on intensity-calibrated maps constructed across four seasons, a two-stage point segmentation method is employed to extract geometrically…

View free PDFSource page
openalexInstitutional Repositories DataBase (IRDB)

HCl-based halide vapor phase epitaxy and selective-area HCl gas etching of (−112) β-Ga <sub>2</sub> O <sub>3</sub>

Takayoshi Oshima, Yuichi Oshima

We propose (−112) and crystallographically equivalent (1−1−2), (−1−12), and (11−2) planes as fundamental crystal orientations for β-Ga2O3 studies. In homoepitaxy, the epilayer exhibited single crystallinity with tilt and twist dispersions comparable to those of the substrate and…

View free PDFSource page
openalexInstitutional Repositories DataBase (IRDB)

Catalyst poisoning influences from various functional groups of energy carriers towards electrochemical oxidation reactions on non-noble high-entropy alloy anodes in acidic media

Tahawy Rafat, Muflihah Salma Aridha, Hara Kosuke, Ohto Tatsuhiko, Tanimoto Hisanori, Li Tianshu, et al.

Electrolytic synthesis of energy carriers using renewable energy and fuel cells that use energy carriers for regeneration are important technologies for achieving our carbon-neutral society. However, electrochemical reactions in electrolyte containing organic molecules significan…

View free PDFSource page