Exponential decay-augmented transformers for interpretable temporal reasoning

Jothi Prakash

doi:10.1016/j.knosys.2026.117108

Exponential decay-augmented transformers for interpretable temporal reasoning

2026-09-28

0

PRE

AI

J

Jothi Prakash

S

S. Arul Antran Vijay *

DOI：10.1016/j.knosys.2026.117108

Original

Original request for help

Save

Abstract

En 中文

• Proposed EDAT for interpretable temporal reasoning with controlled forgetting. • Integrated exponential decay into self-attention with certified guarantees. • Introduced EDAT-Multi for adaptive multi-scale temporal memory control. • Achieved consistent gains on TORQUE, WikiTimeline, and MATRES. • Verified theory via certified bounds, ablations, and robustness tests.

Keywords:

Temporal reasoning

Transformer attention

Interpretable machine learning

Memory decay modeling

Long-context learning

Exponential decay-augmented transformers for interpretable temporal reasoning

Abstract

Journal

Organization

Cited Papers