TIME: Explicit and Time-Aware Reasoning for LLMs

A new study introduces TIME, a "Temporally Intelligent Meta-reasoning Engine" designed to enhance the reasoning capabilities of large language models (LLMs) in time-sensitive contexts. The traditional approach of reasoning-oriented LLMs often relies on long traces of explicit "thinking" at the beginning of each response, a costly and inefficient method.

TIME addresses these limitations by treating explicit reasoning as a context-sensitive resource, driven by temporal and discourse cues. The framework enriches dialogues with <time> tags in ISO 8601 format, "tick" turns representing silences, and short reasoning blocks that can appear anywhere in a response.

A Four-Phase Training Curriculum

The training of TIME involves a four-phase curriculum, culminating in a full-batch alignment step that trains Qwen3 models to invoke brief in-place reasoning, keeping user-facing text compact. The effectiveness of TIME was evaluated with TIMEBench, a benchmark for dialogues with temporal foundations that analyzes chronology, commonsense in the presence of temporal gaps, anomaly detection, and continuity.

The results show that TIME improves TIMEBench scores compared to baseline Qwen3 models, while reducing reasoning tokens by about an order of magnitude. Training data and code are available on GitHub.