Time-R1: Towards Comprehensive Temporal Reasoning in LLMs

TL;DR

Time-R1 addresses the gap in temporal intelligence for LLMs by endowing a compact 3B model with unified abilities to understand, predict, and creatively generate time-anchored content. It achieves this through a novel three-stage reinforcement learning curriculum with a dynamic reward system that progressively builds temporal comprehension from historical data to extrapolative future tasks, culminating in generation without fine-tuning. The approach yields strong results, with Time-R1 outperforming models hundreds of times larger on challenging future-prediction and creative-generation benchmarks, and it releases Time-Bench and model checkpoints to foster ongoing research. This work demonstrates a practical, scalable path toward truly time-aware AI with substantial implications for time-sensitive reasoning applications.

Abstract

Large Language Models (LLMs) demonstrate impressive capabilities but lack robust temporal intelligence, struggling to integrate reasoning about the past with predictions and plausible generations of the future. Meanwhile, existing methods typically target isolated temporal skills, such as question answering about past events or basic forecasting, and exhibit poor generalization, particularly when dealing with events beyond their knowledge cutoff or requiring creative foresight. To address these limitations, we introduce \textit{Time-R1}, the first framework to endow a moderate-sized (3B-parameter) LLM with comprehensive temporal abilities: understanding, prediction, and creative generation. Our approach features a novel three-stage development path; the first two constitute a \textit{reinforcement learning (RL) curriculum} driven by a meticulously designed dynamic rule-based reward system. This framework progressively builds (1) foundational temporal understanding and logical event-time mappings from historical data, (2) future event prediction skills for events beyond its knowledge cutoff, and finally (3) enables remarkable generalization to creative future scenario generation without any fine-tuning. Strikingly, experiments demonstrate that Time-R1 outperforms models over 200 times larger, including the state-of-the-art 671B DeepSeek-R1, on highly challenging future event prediction and creative scenario generation benchmarks. This work provides strong evidence that thoughtfully engineered, progressive RL fine-tuning allows smaller, efficient models to achieve superior temporal performance, offering a practical and scalable path towards truly time-aware AI. To foster further research, we also release \textit{Time-Bench}, a large-scale multi-task temporal reasoning dataset derived from 10 years of news data, and our series of \textit{Time-R1} checkpoints.

Figures (13)

• Figure 1: Generated outputs from Time-R1 showcasing its capabilities. (Left) Future Event Time Prediction (Stage 2). (Right) Creative Scenario Generation (Stage 3), with output compared to a real-world headline.
• Figure 2: Overview of the Time-R1 framework. The process consists of three stages: (a) Stage 1 establishes foundational understanding by fine-tuning a base LLM on historical data across four temporal subtasks, driven by reinforcement learning (GRPO) and a dynamic reward system, resulting in model $\theta_1$. (b) Stage 2 trains $\theta_1$ for future event time prediction using post-cutoff data and a rule-based reward, producing $\theta_2$. (c) Stage 3 leverages $\theta_2$ for inference-based creative future scenario generation, followed by evaluation, without further RL.
• Figure 3: Stage 1 Training Performance vs.$\,$Baselines. Training curves for Time-R1 ($\theta_1$) and its ablation variant, Time-R1-Fixed-Reward ($\theta_1'$), evaluated against baseline models (indicated by horizontal dashed lines). Plot (a) shows the Overall Total Score across all subtasks, while plot (b) presents the Masked Time Entity Completion subtask. The solid lines demonstrate our models' scores improving throughout the training process, ultimately surpassing the performance levels of most baseline models, including those with significantly larger scales.
• Figure 4: Monthly Average Total Score $R(x, y)$ for Stage 2 Future Event Prediction (August 2024 - Feb 2025). Compares Time-R1 variants ($\theta_2$ and $\theta_2'$) against baselines. Evaluated with $\alpha=0.1$.
• Figure 5: Impact of Dynamic Reward on Response Length. The average response length (in tokens) across all Stage 1 tasks during training. The model trained with our full dynamic reward mechanism ("Dynamic Reward") produces consistently and significantly more concise outputs compared to the ablation model trained with a static, fixed reward ("Fixed Reward").