Hwang and Won study retaining repair alternatives for cost-aware execution

Hwang and Won study retaining repair alternatives for cost-aware execution

Jae Hun Hwang and Hong-In Won study a way to make execution decisions without committing to just one repair of an executable model. Several repairs can fit the same observations but predict different outcomes; their interface keeps compatible finite guard repairs, updates their weights using diagnostic and ordinary-execution observations, and uses their predictions to choose routes, priced tests, or a backup. In experiments on synthetic two-stage devices, an equal mixture of a language-model-generated repair and a uniform belief over compatible repairs lowered no-test cost relative to committing to the generated repair. But diagnosis was not uniformly cheaper: access charges could erase its benefit on isolated tasks, while repeated work could recover it. The comparison also depended on how repairs were weighted and whether the system learned from normal execution. Under uniform beliefs, information-gain selection had the lowest total among the compared selectors; workload-valued testing cost more than normal learning. The paper’s central point is that retaining alternatives is not enough by itself: diagnostic selection needs to be evaluated alongside repair weighting and learning from execution. Its results concern the tested synthetic devices and comparisons, so they do not establish how the approach performs on other systems or workloads.

Publication: 64961b7e-a6a3-414a-8b08-b6b5fead4e65

attribution — pass: The study is correctly attributed to the listed authors, Jae Hun Hwang and Hong-In Won. Results are framed as the paper's findings rather than as established fact.

Authors:Jae Hun Hwang, Hong-In Won

summary — uncertain: Central jargon goes unexplained for non-specialists: executable model, finite guard repairs, two-stage devices, information-gain selection and workload-valued testing. The reading also blurs a key finding into 'depended on how repairs were weighted': uniform normal learning beat AI-weighted (language-model-informed) learning by 540.75, and testing did not close that gap. It omits the small magnitudes (2.14 percent; 3.33 percent versus random). Smallest repair: state the uniform-beats-AI-weighted result plainly, give the percentages, and add plain-language glosses or tooltip suggestions for the core terms.

uniform normal learning costs 540.75 less than AI-weighted normal learning on this cohort, with a device-paired interval excluding zero

date — pass: originalPublished is 2026-10-29, which matches the v1 submission date. The feed date of 2026-10-31 is later and is correctly not used.

[Submitted on 29 Oct 2026]

support — uncertain: The listed claims match the abstract verbatim. However, the summary calls 'the paper's central point' that 'retaining alternatives is not enough by itself.' The abstract supports both retaining alternatives and evaluating selection together with weighting and learning; the 'not enough by itself' wording is an interpretive reframing presented as the authors' thesis. Smallest repair: use the abstract's own conclusion wording.

These results support retaining alternatives to a generated repair and evaluating diagnostic selection together with repair weighting and learning from execution.

media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct way to access the source.