XiaomanZhang finds +13% time‑series boost

- Researchers found models swing strongly by modality: systems that excel on imaging can fail on spectral or time‑series tasks unless retrained for that input type. - Adding raw time‑series as model input improved accuracy by about 13% on the tasks XiaomanZhang tested, showing data modality can be a bigger lever than model size. - Teams are now pairing modality‑aware benchmarks with production telemetry to predict real‑world performance gaps rather than trusting single‑run evals. (x.com) (x.com)