Across four domain-model settings on tau-Bench, geometry-conditioned adaptive reasoning triggers raise average task reward from 0.241 (24.1%) under Never-thinking to 0.396 (39.6%) while reducing mean token cost from 104.8k to 93.0k, an 11.2% reduction.

来源：paper:Page 6, Section 6, Paragraph 'Geometry-conditioned triggers combine token efficiency with competitive reward'

报告指标与观测值

average_task_reward

meas-headline-reward-never-think

论文报告 0.241 fraction

实际观测 — fraction

average_task_reward

meas-headline-reward-geometry-conditioned

论文报告 0.396 fraction

实际观测 — fraction

mean_token_cost_thousands

meas-headline-token-cost-never-think

论文报告 104.8 thousands_tokens

实际观测 — thousands_tokens

mean_token_cost_thousands

meas-headline-token-cost-geometry-conditioned

论文报告 93 thousands_tokens

实际观测 — thousands_tokens

token_cost_reduction

meas-headline-token-cost-reduction

论文报告 11.2 percentage_points

实际观测 — percentage_points

这条 Claim 暂无已发布的不可变 Assessment。

这条 Claim 暂无关联执行。