Climate research is pivotal for mitigating global environmental crises, yet the accelerating volume of multi-scale datasets and the complexity of analytical tools have created significant bottlenecks, constraining scientific discovery to fragmented and labor-intensive workflows. While the emergence Large Language Models (LLMs) offers a transformative paradigm to scale scientific expertise, existing explorations remain largely confined to simple Question-Answering (Q A) tasks. These approaches often oversimplify real-world challenges, neglecting the intricate physical constraints and the data-driven nature required in professional climate science. To bridge this gap, we introduce ClimAgent, a general-purpose autonomous framework designed to execute a wide spectrum of research tasks across diverse climate sub-fields. By integrating a unified tool-use environment with rigorous reasoning protocols, ClimAgent transcends simple retrieval to perform end-to-end modeling and analysis. To foster systematic evaluation, we propose ClimaBench, the first comprehensive benchmark for real-world climate discovery. It encompasses challenging problems spanning 5 distinct task categories derived from professional scenarios between 2000 and 2025. Experiments on ClimaBench demonstrate that ClimAgent significantly outperforms state-of-the-art baselines, achieving a 40.21% improvement over original LLM solutions in solution rigorousness and practicality. Our code are available at https://github.com/usail-hkust/ClimAgent.

Multimodal large language models have recently shown strong potential for time series reasoning, where models jointly process textual context and long numerical sequences. However, dense time series representations introduce a large number of tokens, leading to substantial computational and memory overhead during inference. Existing token pruning methods are mainly designed for visual tokens and typically rely on global attention or representation geometry, overlooking the fact that token redundancy varies across layers and modalities. This makes pruning-induced perturbations difficult to control, especially when tokens are pruned across multiple transformer layers. In this paper, we propose StratPrune, a cascade stratum pruning framework for efficient time series reasoning. StratPrune first identifies an early-layer pruning stratum where time series tokens remain less diverse than text tokens, indicating high redundancy suitable for stable pruning. We then derive a hidden-state stability bound that decomposes pruning-induced perturbation into a layer-dependent stability term and a token-level attention saliency term. Guided by this analysis, StratPrune allocates layer-wise pruning budgets according to time series diversity and removes low-saliency time series tokens through a context-aware cascade process. Experiments on 7 tasks show that StratPrune outperforms strong pruning baselines under matched token budgets, achieving up to 30% relative improvement.

The recent advancement of time series has witnessed Vision Language Models (VLMs) as a powerful tool for a variety of tasks of multimodal time series analysis, including forecasting, classification, question answering, etc. However, with time series as an explicit modality, inputs with long time series horizons yield excessive numeric tokens for VLMs, resulting in high inference latency. Moreover, such long contexts often contain redundant or noisy segments that might distract the model and thus degrade downstream performance. Although previous token pruning methods can reduce visual or textual tokens, they are mostly unimodal and do not explicitly explore multimodal semantics for time series. To address this, we propose TimePrune, a plug-and-play temporal pruning framework for VLMs in multimodal time series analysis, which can accelerate VLM inference without fine-tuning the VLM backbone. Specifically, TimePrune formulates time series pruning as a budgeted temporal selection problem and introduces a Trimodal Budgeted Network to score temporal patches via jointly conditioning on linguistic and visual semantics. Then, under a fixed budget, this framework is optimized via the proposed Narrative-Calibrated Binned Grounding constraints, to further facilitate context-aware and temporally-coherent pruning. Extensive experiments in four time series tasks across application domains on multiple VLM backbones show that TimePrune could consistently achieve better performances. For example, on Qwen3VL-8B, TimePrune improves average forecasting performance by 33.91%.