- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper investigates the gap between predictive accuracy and sequential decision-making in perishable inventory control using Decision-TabPFN. Through controlled experiments, the authors analyze predictor quality, decision interfaces, and context selection strategies. The paper further derives an inventory-specific value bound and establishes a regret guarantee that separates demand-prediction error from planning error.
Overall, the paper is primarily an empirical analysis of existing challenges, without providing a convincing solution to address the identified limitations. Moreover, the empirical and theoretical contributions are not sufficiently connected, making the overall contribution of the paper less coherent.
Critical strengths
-
The experiments cover multiple inventory settings, decision interfaces, and context selection strategies.
-
The theoretical analysis provides an interesting perspective on how prediction and planning errors affect long-term policy performance.
Critical weaknesses
-
Limited methodological novelty. The first two contributions primarily provide empirical analyses of the discrepancies between predictive accuracy, decision-interface performance, and online policy returns. However, these issues have been extensively studied in sequential decision-making. The paper does not sufficiently clarify what new insights its empirical findings provide beyond existing knowledge, particularly since the investigated alternatives largely rely on established decision-making principles.
-
Limited theoretical novelty and practical relevance. As acknowledged in the paper, Gong and Simchi-Levi [1] have already established square-root-order regret guarantees for inventory control with cyclic demand. Although this paper derives an inventory-specific value bound and a different regret analysis, the significance of these theoretical advances over existing results is not sufficiently established. Moreover, the guarantee relies on an explicit demand model and complete-horizon planning, whose computational scalability remains limited. This raises concerns about the novelty and practical applicability of the theoretical contribution.
-
Weak connections among the contributions. The empirical findings and theoretical analysis are not sufficiently integrated. In particular, the regret guarantee is established for a separate planning extension and does not rely on any TabPFN-specific properties. Furthermore, the experiments do not demonstrate a clear advantage from using TabPFN proposals within the planning extension. Consequently, the relevance of the theoretical analysis to the proposed Decision-TabPFN framework remains unclear.
Reference
[1] Gong, X.-Y., and Simchi-Levi, D. Bandits atop Reinforcement Learning: Tackling Online Inventory Models with Cyclic Demands. Management Science, 70(9): 6139–6157, 2024.
Questions for authors
-
In the AC–FO comparison, both the prediction target and decision rule change. Can the authors provide further ablations to isolate the contribution of each component?
-
For coverage-based context selection, can the authors separately evaluate state-similarity retrieval and action balancing?
-
What is the computational overhead of the planning extension at larger action spaces? Can the authors clarify whether TabPFN proposals provide any measurable efficiency advantage over simpler alternatives?
2 comments
Q1. How far do the comparisons separate the prediction target from the ordering rule?
When AC also selects greedily, AC and cost-optimizing FO apply the same decision rule: each picks the order with the highest estimated expected profit for the coming day. They differ only in what the frozen TabPFN predicts, either the realized profit of a candidate order (AC) or next-day demand, which the known cost formula converts into expected profit over 129 forecast quantiles (FO). The right-hand column below therefore compares the two prediction targets under a common decision rule. Entries are mean daily profit differences on ten paired 10,000-day demand paths per setting (Table 3).
| Inventory setting | Top-ten AC minus FO | Greedy AC minus FO |
|---|---|---|
| Lost sales, oldest first, no fee | ||
| Lost sales, newest first, no fee | ||
| Backlogging, newest first, fee of 20 |
Margins are paired 95% confidence interval half-widths. Only the backlogging difference passed Holm correction, in both columns. The sampler itself mattered little: greedy minus top-ten AC was , and on these paths, none significant after Holm correction, and on the Table 1 paths greedy selection earned slightly less than the default (Table 12). In the backlogging setting the FO advantage therefore comes from predicting demand and applying the cost formula. These two cannot be separated further, because a demand forecast needs some rule, such as the cost formula or an order-up-to target, to become an order. Table 11 shows how much that rule matters: changing the safety-stock multiplier from 0.2 to 1.0 reversed the ranking against AC in both settings tested.
Q2. Which component of coverage-based retrieval helped?
We withdraw the claim that coverage-based retrieval improves online profit over uniform sampling in the two newest-first settings. Selecting similar inventories and balancing historical order sizes changed together, so the original comparison could not assign the gain to either component. The deployed AC uses every completed record. Table 12 now addresses the practical question for that policy by varying memory length and recorded features directly.
Q3. How much does planning cost, and do TabPFN proposals speed it up?
Direct scoring. At the largest grid tested, scoring roughly 800 candidate orders took about three seconds per decision (Figure 4).
Planning cost. For shelf life and maximum order , the capped demand has possible values. With demand-mixture components, exact two-day planning takes operations and working memory beyond the period counts (Appendix A.8). For fixed and , this gives a quadratic upper bound in .
For a horizon of days and a period upper bound , exact full-horizon planning requires at most state updates, where (Appendix A.12). On the main grid, , and , giving and . This count is prohibitive at that scale, so on the main grid the full-horizon theorem is a statistical guarantee rather than a practical planner (Appendix A.6). We ran exact full-horizon checks on small instances with binary orders over 15 or 31 days.
Do TabPFN proposals help? Two-day planning shares its conditional demand calculations across orders. A TabPFN proposal adds an inference call and gave no detectable profit gain over ranking orders by expected one-day profit. In the full-horizon checks, the two rules chose identical orders throughout 79 of 80 paired runs and differed on one day within tolerance in the remaining run (Appendix F.1). Whether a proposal saves enough search time to offset its inference cost on a large grid remains unmeasured.
What inventory history tells a frozen scorer, and what one-day scoring misses
Thank you for pressing us on what is specific to perishable inventory and how the theory bears on Decision-TabPFN. Our direct action-scoring policy (AC) uses frozen TabPFN to predict the profit of candidate orders from operating records. The revision studies what those records contribute, then measures what even an accurate one-day profit score can leave out.
For a perishable product, two inventories of the same size can have different amounts left to sell tomorrow. Under backlogging, newest-first issuance and a fixed order fee of 20, AC earned 515.74 per day using the full operating record, 462.11 when the age-specific inventory features were removed and 347.76 when only orders and profits were retained. All three figures average days 1 through 10,000, including initialization (Table 12).
The timing of past records also mattered. Across the ten positions in the hidden demand cycle, the correlation between average order size and average demand in the final third of the runs ranged from 0.88 to 0.98 for AC, compared with −0.97 to −0.92 for the conventional stock rules (Figure 6). At fixed inventories under lost sales and oldest-first issuance, removing past records from the same cycle position increased the chosen order's expected one-day profit loss by 27.6%, compared with 3.9% after removing the same number at random (Appendix G.3). Only the evaluator saw cycle labels.
In lost sales, AC beat median forecast-to-order (FO) in all four settings. Neither cost-optimizing FO contrast passed Holm correction (Tables 1, 3, 15).
Gong and Simchi-Levi (2024) give a square-root-order regret guarantee for cyclic nonperishable inventory with a known period and discuss unknown length with a known upper bound in Remark 2. Our planning theorem also assumes a known upper bound. The new step is to control how the stock left by today's order affects future value when units have different expiration dates. Under lost sales, Lemma 6.2 bounds the difference in optimal remaining profit between two starting inventories after the same demand history, independently of the days remaining. Without expiration or disposal, a single excess unit under zero demand can incur a holding charge every remaining day (Appendix A.13).
Past demand remains informative after the original inventory expires. Subtracting the value of empty stock after the same demand history lets the history-only contribution cancel in expectation between policies, because orders do not affect demand (Appendix A.5). This gives prediction error a coefficient independent of the remaining horizon. Theorem 6.3 uses that bound to separate demand estimation from the error allowed in full-horizon planning when demand is observed or recoverable.
At inventories visited by AC, Proposition A.3 bounds immediate scoring and sampling losses and the effect of the resulting inventory on future value, without requiring a gap between orders. Table 14 measures the corresponding economic quantities in eight small lost-sales instances. With three-day shelf life, oldest-first issuance and a fee of 20, AC's executed order gave up 1.866 in immediate profit per day relative to the best one-day order. Even that best one-day order gave up 3.752 relative to the full-horizon choice when both were followed by optimal orders. The table also reports the signed difference in continuation value between AC's order and the best one-day order, which is needed to recover AC's total loss. In both three-day instances with this fee the myopic loss had the larger mean. The immediate loss had the larger mean in the other six.
Corollary A.4 gives sufficient conditions for sublinear AC regret. It requires a uniform positive gap between the two best immediate orders and sublinear cumulative contributions from score error, sampling and the future value forgone by the best immediate order. For the default ten retained orders, let st be the standard deviation of their scores and τt the sampling temperature. The effective temperature in profit units is λt=τt(st+10−6). Since τt≥0.1, the proof's daily sampling upper bound satisfies λtlog10≥10−7log10 after initialization. Summing that bound cannot certify sublinear regret for the default sampler. Actual loss may be smaller. Proposition A.5 gives an alternative certificate using valid bounds on score error and on the probability that two orders leave different inventories, without computing the optimal continuation. The deployed checkpoint has not been shown to satisfy these sufficient conditions. Theorem 6.3 gives a separate guarantee for the planner that checks each proposal against its full-horizon alternative.