dougalldeepmind/2026-08-25-table2-9284-difficult-advice-verbose-token-matched-train-mixture

Token-matched verbose difficult-advice arm. Holds difficult advice's share of the TRAINABLE TOKENS at the control's value while the traces are ~3x longer, by keeping only a subset of the expanded rows. Its sibling arm holds the ROW share instead; together they separate more deliberation from more difficult-advice signal. field value experiment Token-matched verbose difficult-advice arm. Holds difficult advice's share of the TRAINABLE TOKENS at the control's value… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-25-table2-9284-difficult-advice-verbose-token-matched-train-mixture.