arXiv cs.AI: Where Should the Code Go? Locating Foundation-Model Errors in Bitemporal Inventory Simulation

arXiv cs.AI: Where Should the Code Go? Locating Foundation-Model Errors in Bitemporal Inventory Simulation

Authors Durga Sandeep Saluru, Lanchen Song and Aditya Kumaran study where code can help a foundation-model component maintain inventory state from reports and roll a supplied scenario forward. The reports have two timestamps: when an event happened and when its report arrived; later reports can revise earlier ones. In an exactly solvable lost-sales inventory environment, they compare code that prepares inputs, continues the simulation or operates inside the model call, alongside requests for more reasoning. They report that weak models make separable report-selection and rollout errors: canonical input removes the first, while exact continuation removes the second, and neither fix repairs the other. Which error dominates varies by model; strong models reach ceiling on raw reports. More reasoning is not a general substitute: its effects vary by model and interface, including cases where forced calls or output limits impair performance. The authors frame the work as a controlled component diagnostic and caution that, for some model and provider pairs, the tool-calling interface confounds reasoning evaluation. The evidence comes from one exactly solvable inventory environment, so whether the reported error patterns carry over to other settings remains open.

Publication: 5eaeee73-ad07-4677-84ef-c61e6c454ee8

attribution — pass: The draft correctly credits Saluru, Song and Kumaran and keeps the findings as the authors' reported results. The title prefix uses the feed label 'arXiv cs.AI', although the paper's primary listing is cs.SE. This is minor and does not misattribute the work.

Authors:Durga Sandeep Saluru, Lanchen Song, Aditya Kumaran

summary — uncertain: The draft is faithful but hard to follow without specialist background. It never defines central terms: 'foundation-model component', 'lost-sales inventory', 'canonical input', 'exact continuation', 'rollout', 'at ceiling', 'forced calls' and 'tool-calling interface'. It also gives little plain account of why this matters. 'Not a general substitute' softens the source's flat 'is no substitute'. The draft omits the named collapses: GPT-5.4 mini and DeepSeek V4 Pro under a forced answer call. Smallest repair: restore 'no substitute', name those cases, and add tooltips or plain glosses for the central terms.

Asking for more reasoning is no substitute and can depend on the interface

date — pass: The originalPublished date of 2026-11-10 matches the arXiv v1 submission date. It is not the later feed date of 2026-11-13.

[Submitted on 10 Nov 2026]

support — pass: Each claim matches the abstract. This covers the remedies, the separable errors, the variation in which error dominates, strong models at ceiling, interface-dependent reasoning effects and the single test environment. The open question and the limitations about generalization are properly framed as unknowns, not as conclusions.

Canonical input removes report-selection error, exact continuation removes rollout error, neither repairs the other, and which error dominates differs by model

media — pass: The selected media is the verified arXiv abstract page for this paper. It is the direct original source.