Geometry of Divergence: Tracking Hidden-State Trajectories for Adaptive Multi-Turn Reasoning
Not assessedPlan blockedFindingclaim-action-conditioned-triggers
Action-conditioned trigger policies exhibit domain-dependent performance: on Retail with Qwen3-14B, triggering on Read (0.393) and Write (0.381) outperforms Respond (0.290), whereas on Airline with Qwen3-14B, the ordering inverts, with Respond achieving 0.372 while Read achieves 0.352 and Write achieves 0.340.
Source: paper:Page 6, Section 6, Paragraph 'The effectiveness of action-conditioned triggers varies across tasks'
Reported and observed measurements
task_reward
meas-action-retail-read
Reported 0.393 fraction
Observed — fraction
task_reward
meas-action-retail-write
Reported 0.381 fraction
Observed — fraction
task_reward
meas-action-retail-respond
Reported 0.29 fraction
Observed — fraction
task_reward
meas-action-airline-respond
Reported 0.372 fraction
Observed — fraction
task_reward
meas-action-airline-read
Reported 0.352 fraction
Observed — fraction
task_reward
meas-action-airline-write
Reported 0.34 fraction
Observed — fraction
Assessments (0)
No immutable Assessment has been published for this Claim yet.
Runs (0)
No execution has been linked to this Claim yet.