acceptodds
Withdrawn from ICLR 2027

Constrained Triple-Head Price Optimization (C3PO) Network for Discrete-Choice Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We propose a Constrained Triple-Head Price Optimization (C3PO) network for real-time bi-level decision making in discrete-choice environments. In these settings, a seller selects product or service offerings to achieve business objectives, while heterogeneous users make personalized acceptance or rejection decisions based on their preferences, ultimately determining seller outcomes. Existing approaches that fine-tune pre-trained tabular foundation models on synthetic data are limited by their imitation-learning paradigm and inability to enforce common pricing rules, including cross-product and global constraints. C3PO addresses these limitations through a novel transformer-based architecture that jointly integrates price imitation learning, multi-task revenue prediction, and in-context elasticity learning to generate robust pricing recommendations across domains without retraining. A lightweight differentiable constraint layer enables generated prices to satisfy business constraints while optimizing the revenue objective. During inference, frontier-model prompting elicits elasticity priors for previously unseen products informed by behavioral economics literature, improving out-of-distribution pricing effectiveness. We demonstrate strong out-of-distribution performance on simulated, synthetic, and real-world datasets. C3PO is trained on data generated from multiple classical discrete-choice models, comprising simulated customer segments and counterfactual action-outcome pairs, and is evaluated on randomly generated choice environments without observing the underlying preference structure. Across simulated benchmarks where the calibrated MNL model is misspecified relative to the ground-truth choice model, C3PO improves expected revenue by 1.4 to 7.9 percentage points over a calibrated MNL baseline. The constraint layer reduces average absolute constraint violations by more than 50%. We further deploy the model in real-world applications, including healthcare tender pricing and airline ancillary pricing, where C3PO improves directional pricing effectiveness, measured by price-increase and price-decrease recall, by 18–34 percentage points depending on the baselines considered, including calibrated MNL, classical optimization, and off-the-shelf tabular foundation-models.

Primary area:
optimization
Keywords:
in-context learning, discrete-choice models, price optimization, tabular foundation models, meta-learning, constrained neural networks

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_3U2DOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a new network, called Constrained Triple-Head Price Optimization (C3PO), that aims to provide prices of services to customers. The input features of the network is the customer features as well as the historical prices. Given the inputs, the network consists of three sequential stages: 1) encoding the features through transformers, 2) taking elasticity prior as new inputs to enable in-context learning, 3) two heads that output the price and revenue. The training of the network consists of three learning paradigms: 1) imitation learning that makes the output price to approximate the labeled optimal price; 2) learning the revenue head such that the output revenue predicts the true revenue caused by the output price; 3) In-context learning with elasticity priors. Experiments show that the proposed network outperforms a calibrated MNL baseline.

Critical strengths

None

Critical weaknesses
  • The paper suffers from severe presentation issues, making me hard to understand the core contributions of the paper, which I also believe is not solely caused by my unfamiliarity with the topic. Therefore, I can not recommend acceptance at this stage. Here are some examples:

    • The paper lacks mathematical notations to formally formulate the problem of this paper. There're only very few notations, e.g., paragraphs around equation (1) and (3). Many concepts are described only informally (for few examples, the elasticity priors, the discrete choices model etc.) I believe that the presentation requires a major revision before acceptance.
    • For another example, in line 140-142, the paper formulates the goal as "Instead, a desirable theoretical guarantee is that, for any discrete choice model, the probabilities it generates can be approximated by those produced by the model, or equivalently, that the induced utility functions are themselves universally approximable". I can not understand this sentence. I also do not how this goal connects to the other part of this paper.
    • Even in places where mathematical notations are used, they are not complete enough neither. For example, the paper does not describe the expression of each term in loss function in equation (2).
    • Some sentences are overclaimed, e.g., in line 72-73, the paper says "the proposed approach is applicable beyond pricing to many other problem areas such as optimal product design, assortment optimization, transportation planning, and marketing, among others". However, no experiments support this claim, so this is better framed as future works.
  • In addition to the presentation issue, I also do not see too many contributions from the paper. The network architecture largely uses transformers as building block. The training paradigms are also similar to supervised learning. Theoretical results (e.g., Theorem 4.4) seem very straightforward. The significance and difficulty of the studied problem are also not satisfactorily justified. The main contribution seems to be a direct application of machine learning to a new problem, which I believe has limited contribution, and given the competitiveness of ICLR, does not warrants an acceptance either.

Questions for authors

None

Reviewer_52NpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes C3PO, a transformer trained on simulated discrete-choice environments to recommend prices in new domains through in-context learning. The method combines optimal-price imitation, auxiliary revenue prediction, elasticity-based regularization, and an approximate constraint layer. Experiments report revenue improvements over calibrated MNL on three benchmarks with heterogeneous demand, alongside directional pricing metrics on two private datasets.

However, the current submission contains incorrect theoretical claims and substantial weaknesses in evaluation. In particular, the strongest directional-proxy results start from oracle-optimal prices, the private-data metrics do not establish revenue improvement, and the constraint layer leaves substantial violations. These issues prevent the evidence from supporting the paper’s central claims.

Critical strengths

Amortizing price optimization across markets and product categories could reduce the cost of repeatedly fitting and optimizing domain-specific models.

Critical weaknesses

Lemma 4.6 claims that conditioning the target distribution on additional observed labels cannot increase its divergence from the source distribution. This is false without additional assumptions. For example, let source and target initially be identical, with a binary label Y∼Bernoulli(1/2)Y\sim\rm{Bernoulli}(1/2), and consider a hypothesis whose loss is YY. The initial divergence is zero. Conditioning the target on Y=1Y=1 changes its expected loss to one while the source expectation remains 1/21/2, increasing the divergence to 1/21/2. Consequently, the claimed improvement in Corollary 4.7 is not established.

Theorem B.5 also gives an incorrect multiproduct MNL optimal-price formula. Under its stated assumptions—common sensitivity β\beta, zero costs, and an interior optimum—the first-order conditions yield

\qquad R=\sum_j p_jq_j.$$ Thus all optimal prices satisfy $p_k^*=R^*+1/\beta$, or equivalently $$p_k^*=\frac{1}{\beta q_0^*},$$ where \$q_0^* \$ is the outside-option probability. The stated expression $1/[\beta(1-q_k^*)]$ generally applies only in the single-product case. The final algebraic step in the proof incorrectly removes the contribution of the other products. Established multiproduct pricing results retain these substitution effects. These errors do not automatically invalidate the implementation, but the authors must reconcile the mathematics with the label-generation code and elasticity anchors.
Questions for authors
  1. Can the theoretical errors be resolved, with affected results audited?

  2. How does C3PO perform from nonoptimal starting prices against stronger demand-model baselines?

Reviewer_bwpHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method called C3PO for multi-product pricing under discrete choice models. It integrates price imitation, revenue forecasting, contextual learning, and elastic priors, while adding an additional layer to minimize constraint violations through price adjustments. The model is trained in simulated demand environments and evaluated in two aspects: comparing revenue on synthetic datasets and assessing the correctness of pricing direction using real-world B2B and airline historical data.

Critical strengths
  1. The chosen question is indeed quite practical. Every time we enter a new market or create a new category, we have to build a pricing model from scratch, which is too costly. Learning a universal model and quickly adapting it to new scenarios based on examples in context is a worthwhile direction; Especially now, pricing has to meet a bunch of business rules at the same time, it's not just about calculating the best price.

  2. The design of learning objectives also makes sense. Just letting the model imitate pricing only learns the mapping from "features → price"; Adding profit forecasts also allows the model to learn more. Additionally, using elastic priors to incorporate domain experience when data is scarce is also a viable approach.

  3. The experimental scene is open. Five demand models were used for training, and the evaluation also covered different applications and product quantities. Especially during training, I only saw 2-6 products, but during testing, I directly tested 15. This generalization test was done well.

Critical weaknesses
  1. The theoretical part needs to be revised. The conclusion of the optimal price for multiple products in Theorem B.5 cannot be derived by its own deduction. Lemma 4.6 also assumes one thing: if a few more labels are given as conditions, the target distribution will definitely be closer to the source distribution - but this generally does not hold. The author needs to correct these two results and clarify whether these errors have affected the actual price anchor points used in the code.

  2. On the experimental side, the persuasiveness cannot support the conclusion of the paper. C3PO Dir started directly with the known optimal price in Group A, but after a slight adjustment, it is still not far from the optimal price, which does not indicate that it will optimize on its own. As for the real data, increasing prices when there are transactions and decreasing prices when there are no transactions in history may not necessarily result in more profits. If a random baseline comparison could be added, these indicators would be better explained; Especially for the aviation team, both scores are stuck around 0.5, which has little reference value.

  3. As for the constraint layer, it cannot really guarantee the feasibility of the solution. From Tables 15-16, it can be seen that there are still many constraint violations, and even new violations have been added: in aviation data, the average relative ranking violation rate has increased from 0% to 18.6%. The paper should calculate the proportion of all constraints satisfied simultaneously in the final result and compare it with the precise projection method.

  4. The input of real data is not clear enough. In the appendix, it is stated that the contextual input includes the optimal price tag and the counterfactual price return combination. But in B2B and aviation scenarios, the demand model itself is unknown. How did these things come about? Do all baselines receive the same amount of information? I didn't explain this clearly.

Questions for authors

1.How do you plan to modify Theorem B.5 and Lemma 4.6? After the modification, will the price anchor, training label, and reported results in the code be affected?

2.How are the optimal price tags and counterfactual price revenue context pairs obtained when the underlying demand models of B2B and Airline are unknown? What corresponding information can each baseline obtain?

3.What would happen if C3PO Dir were adjusted from historical real prices instead of starting from the optimal price? How much is the difference between being unable to keep up with the price or randomly adjusting the direction?

4.What is the proportion of cases where all constraints are met simultaneously in the final prediction? What would be the final benefit if an accurate feasibility projection were done first?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.