The computational cost of large language models (LLMs) is a primary obstacle to sustainable deployment. Static resource allocation is inefficient, as not all inputs require the same depth of processing. We propose a framework for adaptive, compute-efficient learning via conceptual criticality, which dynamically tailors computation to the assessed difficulty of an input. A lightweight criticality prediction module es- timates conceptual complexity on a continuous scale, and this score governs the LLM’s inference pathway, selectively activating token pruning, layer skipping, and quantization. Simple inputs are processed with minimal FLOPs and la- tency, while complex inputs use the model’s full capacity to preserve accuracy. We benchmark our framework and in- troduce metrics to quantify sensitivity to input criticality and per-sample computational savings. Results demonstrate an improved accuracy-efficiency trade-off, paving the way for more resource-aware systems.

Deep learning models are emerging as strong alternatives to numerical weather prediction, yet their internal representations remain poorly understood. We analyze the latent space of Microsoft’s Aurora model to test whether its embed- dings align with known physical processes. First, we show that land–sea distinctions are strongly captured, with errors mainly at coastlines. Second, we examine extreme surface temperatures using percentile-based thresholds, finding that embeddings reveal a gradient from moderate to severe events, though recall degrades at the rarest percentiles. These results suggest that Aurora’s encoder encodes physically consistent features but underestimates rare extremes. Our study combines deep learning forecasting, interpretable representation learning, and classical ML probing, illustrating how cross-disciplinary AI methods can yield insight into foundation models