cs.SENov 10, 2026

Where Should the Code Go? Locating Foundation-Model Errors in Bitemporal Inventory Simulation

Authors: Durga Sandeep Saluru, Lanchen Song, Aditya Kumaran

Organizations: Distyl AI

Abstract

Operational reports carry two clocks: when an event happened and when its report arrived; later reports can revise earlier ones. A foundation-model component that maintains the state supported by the reports visible at a cutoff, then rolls a supplied scenario forward, can be helped by exact code before it, after it or inside the call, or asked to reason harder. We ask which remedy removes which error, for which models, on an exactly solvable lost-sales inventory environment: 36,672 independently rescored requests over twelve configurations of closed and open-weight models, with a forced-answer-call interface as a control and protocols frozen before the calls. Weak models fail in separable ways. Canonical input removes report-selection error, exact continuation removes rollout error, neither repairs the other, and which error dominates differs by model; strong models are at ceiling on raw reports. Asking for more reasoning is no substitute and can depend on the interface: under a forced answer call high effort collapses GPT-5.4 mini and DeepSeek V4 Pro, which gain or hold under automatic choice; on its provider the forced call disables reasoning for DeepSeek V4 Flash; and at high effort gpt-oss exhausts its output budget. The contribution is a controlled component diagnostic, released with every request, and evidence that for some model and provider pairs the tool-calling interface confounds reasoning evaluation.

Figures & tables

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Nov 11, 2026

From Forecasting to World Modeling: Can Temporal Foundation Models Recover Causal Response?

A model that forecasts sales accurately can still mislead a retailer about how sales respond to a price change. We evaluate released tabular and time-series foundation models, as well as a foundation model pretrained specifically to estimate causal effects. We report three findings. First, forecast accuracy does not indicate the accuracy of the predicted effect of price, both on simulated data in which the true effect is known and on real retail sales records. Second, for the tabular foundation models, additional context improves forecasts but not the predicted effect of price. Third, instrumental variables, the standard remedy in econometrics, reduce this error only when supplied in a specific form and increase it for every model under confounding when supplied as raw instruments.
Nov 16, 2026

Procedure without Judgment: What Small Language Models Learn under an Exact Verifier

When an authorization policy blocks an agent's tool call, the audit log should name the facts that forced the decision and nothing more. For an executable policy over an enumerable context space this minimal sufficient reason can be computed exactly, so an explanation is scored against ground truth rather than a proxy for it. We use that exactness to ask what a model can learn from a verifier that has no error in it. The explainer sees one request and one decision, never the policy, and has thirty turns in which it may call the policy on altered requests and ask whether a candidate set of facts is sufficient. Frontier models find the exact reason about half the time with the tools alone, and one sentence telling them to shrink the answer lifts Haiku 4.5 from 58% to 75%. A 4B model fine-tuned on 2000 traces of the deletion algorithm reaches 92% on the policy it was trained on. Teacher kind is what moves this: at matched data size the algorithm-taught student leads the frontier-taught one in all three training seeds, and the two arms never overlap. What the student learns is a fixed schedule of sufficiency checks, and the schedule does not travel. On three public policies it never saw, its answers stay sufficient (94%) and stop being minimal (48%), while Haiku under the same sentence is exact on all 47 episodes. Reinforcement learning from the exact reward leaves held-out accuracy where it was and inverts the strategy, from probing the policy to spending the entire budget on sufficiency checks. The oracle made the effect of training legible, but it did not make training better.
Feb 16, 2026

Broken Chains: The Cost of Incomplete Reasoning in LLMs

Reasoning-specialized models like OpenAI's 5.1 and DeepSeek-V3.2 allocate substantial inference compute to extended chain-of-thought (CoT) traces, yet reasoning tokens incur significant costs. How do different reasoning modalities of code, natural language, hybrid, or none do perform under token constraints? We introduce a framework that constrains models to reason exclusively through code, comments, both, or neither, then systematically ablates token budgets to 10%, 30%, 50%, and 70% of optimal. We evaluate four frontier models (GPT-5.1, Gemini 3 Flash, DeepSeek-V3.2, Grok 4.1) across mathematical benchmarks (AIME, GSM8K, HMMT). Our findings reveal: (1) truncated reasoning can hurt as DeepSeek-V3.2 achieves 53% with no reasoning but only 17% with truncated CoT at 50% budget; (2) code degrades gracefully as Gemini's comments collapse to 0% while code maintains 43-47%; (3) hybrid reasoning underperforms single modalities; (4) robustness is model-dependent as Grok maintains 80-90% at 30% budget where OpenAI and DeepSeek collapse to 7-27%. These results suggest incomplete reasoning chains actively mislead models, with implications for deploying reasoning-specialized systems under resource constraints.