Geometry of Divergence: Tracking Hidden-State Trajectories for Adaptive Multi-Turn Reasoning
Not assessedPlan blockedFindingclaim-overall-headline

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.

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

Reported and observed measurements

average_task_reward

meas-headline-reward-never-think

Reported 0.241 fraction

Observed — fraction

average_task_reward

meas-headline-reward-geometry-conditioned

Reported 0.396 fraction

Observed — fraction

mean_token_cost_thousands

meas-headline-token-cost-never-think

Reported 104.8 thousands_tokens

Observed — thousands_tokens

mean_token_cost_thousands

meas-headline-token-cost-geometry-conditioned

Reported 93 thousands_tokens

Observed — thousands_tokens

token_cost_reduction

meas-headline-token-cost-reduction

Reported 11.2 percentage_points

Observed — percentage_points

Assessments (0)

No immutable Assessment has been published for this Claim yet.

Runs (0)

No execution has been linked to this Claim yet.