Preprint reports stronger time-series reasoning from paired visual representations

Researchers report that combining two visual representations of time-series data improved a language model's overall TSRBench accuracy from 44.5% with line plots alone to 53.2%. Their preprint builds on Qwen3-VL-8B and reports 36.7% accuracy for the strongest evaluated dedicated time-series language model baseline. Line plots preserve original measurements, while normalized delay embeddings expose temporal structure while removing positive amplitude rescaling and offsets. Dedicated visual encoders feed both views into the language model. The results establish benchmark gains rather than deployment performance.