Post
3,233
S²-MAS update: YN decision models, a 500M-persona resource, and household consumption prediction.
YN-mini, YN-Lite-4B and YN-master-9B map household states and complete choice menus to option probabilities. The paper now uses these names throughout.
Completed population application: 12M distinct households, 24M one-shot choices and 288M monthly ledger states in 3h59m17s on 14 GPUs. YN-mini-MLX achieves 74.55% accuracy and 38.37% question-macro F1 on the 240,458-record consumer-choice test.
New 1,060-household results bring standalone 4B/9B forecasts together with fully rational and advised policies. Standalone 9B's annual MAE is 3.42% below the deterministic rational benchmark and 2.11% below carry-forward; its population mean remains 19.7% below observed consumption. The paper reports income/wealth group MAE and explains the consumption–cash paths behind these differences. The household 4B uses the head-only 200k variant.
Relabeling the policy year from 2018 to 2031 leaves selected answers unchanged in 12 paired tests per model; separate outcome-note interventions reveal task-dependent information sensitivity. Paired policy results subtract separately simulated baseline paths from tax-cut paths.
The 56-page paper presents learned decision policies as another route to large-scale heterogeneous-household modeling.
Article: https://huggingface.co/blog/ya9919/s2-mas
Updated PDF: https://github.com/KataDavidXD/s2-mas/blob/main/paper/S2_MAS_Paper.pdf
Paper + source release: https://github.com/KataDavidXD/s2-mas/releases/tag/v67
Models: https://huggingface.co/collections/ya9919/s-mas-models-sharp
Updated 28 October 2026. Research contact: yl9919@connect.hku.hk.
YN-mini, YN-Lite-4B and YN-master-9B map household states and complete choice menus to option probabilities. The paper now uses these names throughout.
Completed population application: 12M distinct households, 24M one-shot choices and 288M monthly ledger states in 3h59m17s on 14 GPUs. YN-mini-MLX achieves 74.55% accuracy and 38.37% question-macro F1 on the 240,458-record consumer-choice test.
New 1,060-household results bring standalone 4B/9B forecasts together with fully rational and advised policies. Standalone 9B's annual MAE is 3.42% below the deterministic rational benchmark and 2.11% below carry-forward; its population mean remains 19.7% below observed consumption. The paper reports income/wealth group MAE and explains the consumption–cash paths behind these differences. The household 4B uses the head-only 200k variant.
Relabeling the policy year from 2018 to 2031 leaves selected answers unchanged in 12 paired tests per model; separate outcome-note interventions reveal task-dependent information sensitivity. Paired policy results subtract separately simulated baseline paths from tax-cut paths.
The 56-page paper presents learned decision policies as another route to large-scale heterogeneous-household modeling.
Article: https://huggingface.co/blog/ya9919/s2-mas
Updated PDF: https://github.com/KataDavidXD/s2-mas/blob/main/paper/S2_MAS_Paper.pdf
Paper + source release: https://github.com/KataDavidXD/s2-mas/releases/tag/v67
Models: https://huggingface.co/collections/ya9919/s-mas-models-sharp
Updated 28 October 2026. Research contact: yl9919@connect.hku.hk.