NEURIPS / EMNLP / LOG 2026 Learning from self-play. Looking inside models. Questioning the tests. SCOPE / Cyrus Kwan et al. NeurIPS SCOPE LOCOS Composed image retrieval AI psychophysics EMNLP Multimodal retrieval heads Hidden VLM instability Composing steering vectors LoG Graph-to-table interfaces PREPRINTS SPSD / FACE-Eval / VLM frequency filtering