LLaTiSA: The AI That Learns to Read Charts Like Humans — Easy to Hard

Why do even the smartest AI models stumble when reading a simple graph? A new paper from Alibaba Group reveals a surprising truth: most AI systems try to run before they can walk.

LLaTiSA (Large Language and Time Series Assistant) introduces a radical approach to time series analysis — teaching AI through a difficulty-stratified curriculum, just like how humans learn mathematics from arithmetic before calculus.

The system breaks time series reasoning into four cognitive levels: numerical read-out (finding exact values), pattern perception (spotting trends and anomalies), semantic reasoning (interpreting data with domain knowledge), and predictive inference. Then it trains the model sequentially through each level.

What makes LLaTiSA unique is its "dual-view" input architecture. The model simultaneously receives a visual plot for intuitive pattern recognition and a precision-calibrated numerical table for exact value verification. This combination bridges the gap between qualitative intuition and quantitative accuracy that has plagued previous approaches.

The results are striking. LLaTiSA outperforms GPT-4o on numerical read-out tasks (86.8% vs 54.2%) and semantic reasoning (67.0% vs 48.0%), despite being built on a much smaller 8B parameter backbone. In ECG interpretation tests, it achieved 84% lead assessment coverage using only 2.5% of the training data required by specialized models.

The accompanying HiTSR dataset provides 83,000+ verified samples with chain-of-thought annotations, establishing a new benchmark for the field.

Ablation studies confirmed the curriculum approach is essential — joint training on all difficulty levels simultaneously led to a 14.93% drop in out-of-distribution performance compared to the staged approach.

This research signals a paradigm shift: building smarter AI isn't just about scale, it's about structured learning.