Unverified50% confidenceFactExact time
The OOD task risk upper bound given in the paper is controlled by the sum of the ID subtask risks, with a scaling coefficient on the order of O(k) that is independent of the number of reasoning steps T, where k is the number of output tokens per step.
1
Sources
50%
Confidence
Long-term
Relevance
11/5/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThrough a parity-coupling construction, the paper proves that for the subtask empirical risk minimizer, when T≤n, the expected final task risk is at least the optimal in-class risk plus Ω(√(T/n)), showing that the upper bound is tight in order.76% similarUnverifiedThe limitations the authors list include: the 9% saving is not an order-of-magnitude improvement, the 7 tasks are skewed toward English, there are only three mid-tier judge models, each configuration was run only once, ratio-type estimators such as F1 are not yet supported, and active routing requires the workflow to log probabilities while some production systems lack such logs.66% similarUnverifiedThe first version of the task set was 100% trick questions, with a single-attempt baseline of 0/12, making it impossible to measure the harm of ungated correction; the harm only became measurable after adding clean tasks.65% similarUnverifiedData breadth is critical for the instruction-following task, while breadth is nearly redundant for the high-risk task.64% similarUnverifiedThe combination of structured answers plus probabilities allows the caller to do threshold control: when the probability is below a certain value, the task is handed off to a human, and when it is above the threshold, it is executed automatically.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1138321API
curl https://kongchang.com/api/v1/knowledge/claims/1138321MCP
get_claim(id=1138321)