According to StanfordAILab, an ICML 2026 workshop oral shows internal data repetition degrades language models, earning runner-up recognition.