acceptodds
Under review as a conference paper at ICLR 2027

Tail-Risk Cross-Entropy for Robust Generative Decision-Focused Learning

14 views

Abstract

Robust generative *decision-focused learning* (*DFL*) seeks to learn predictive distributions that induce reliable decisions under adverse outcome uncertainty. Existing *Conditional Value-at-Risk* (*CVaR*)-based methods, however, rely on estimated CVaR values or CVaR-optimal decisions as supervision, which depend on the conditional tail distribution and are not directly available from standard single-outcome data. We propose *Tail-Risk Cross-Entropy* (*TRCE*), a decision-focused surrogate that bypasses robust-supervision reconstruction and enables CVaR-aware generative learning directly from single-outcome observations. Our key insight is that, although CVaR is a distribution-level risk measure, the ideal CVaR-weighted objective admits an equivalent sample-wise representation through the Rockafellar–Uryasev (RU) formulation. TRCE learns the corresponding conditional Value-at-Risk (VaR) threshold from observed input–outcome pairs and uses it to construct sample-wise tail-risk signals, eliminating the need for explicit ground-truth CVaR estimates or CVaR-optimal decision labels. We establish finite-sample guarantees connecting model approximation, threshold-estimation, and generalization errors to downstream CVaR decision regret. Experiments under both light- and heavy-tailed uncertainty demonstrate consistent improvements over deterministic and generative DFL baselines.

Primary area:
optimization
Keywords:
Robust Decision-Focused Learning, Conditional Value-at-Risk, Generative Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_91kqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Tail-Risk Cross-Entropy for generative decision-focused learning. It uses a sample-wise objective derived from the CVaR formulation, allowing risk-aware training from single-outcome observations.

Critical strengths

The paper addresses an important problem in robust decision-making. The connection between CVaR optimization and sample-wise generative training is interesting and potentially useful.

Critical weaknesses

The theoretical analysis uses an oracle KL regularizer involving the true conditional distribution, while the practical algorithm uses a diffusion training loss. It is unclear how directly the regret guarantees apply to the implemented method.

TRCE still requires learning conditional VaR thresholds, which may be difficult when tail observations are scarce. The experiments do not sufficiently establish robustness to threshold estimation errors in more challenging low-data settings.

The practical method relies on a restricted candidate decision set and periodically solves sample-based CVaR problems to expand it. The impact of candidate coverage and the associated computational cost are not clear.

Questions for authors

Can the authors clarify the relationship between the oracle KL regularizer in Theorem 4.1 and the practical diffusion objective in Equation (7)?

How sensitive is downstream CVaR regret to errors in the learned VaR thresholds, especially for smaller tail probabilities?

How does candidate-set size affect decision quality and training cost?

Reviewer_Hx84OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Tail-Risk Cross-Entropy (TRCE), which can enable CVaR-aware generative learning directly from single-outcome observations. Specifically, TRCE learns the corresponding conditional Value-atRisk (VaR) threshold from observed input–outcome pairs and uses it to construct sample-wise tail-risk signals. The authors provide both empirical and theoretical analsis.

Critical strengths
  1. The problem is important and the supervision gap is clearly stated.
  2. The authors provide the finite-sample guarantees.
  3. Experimental results show that the proposed approach is effective.
Critical weaknesses
  1. The Theorem 4.1 does not include the TRCE term. This means the theoretical results are not obtained by TRCE.
  2. The paper lacks the related work section.
  3. Test sets used in the experiment are small.
  4. TRCE only removes the reconstructed target, but the tail estimation from the model still exists.
Questions for authors

Please see the Weaknesses.

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

The paper addresses robust generative decision-focused learning when only one outcome is observed per training input, while the downstream objective is conditional value-at-risk (CVaR). It proposes Tail-Risk Cross-Entropy (TRCE), which normalizes model-induced CVaR costs across a candidate decision set and trains this distribution against true CVaR-weighted targets. The Rockafellar–Uryasev representation replaces unavailable target CVaR costs with sample-wise losses using a learned conditional VaR threshold. A diffusion model is then fine-tuned with a combination of generative and TRCE losses. The paper gives population optimality and finite-sample regret results and reports lower CVaR decision regret than several DFL baselines on portfolio, power-scheduling, and web-advertising tasks with Gaussian or Student-t perturbations.

Critical strengths
  • This paper provides a principled reduction from distribution-level CVaR to a sample-wise training objective.
  • The paper establishes finite-sample bounds connecting approximation, threshold estimation, and generalization errors to downstream CVaR regret.
Critical weaknesses
  • The regret guarantee relies on an oracle KL regularizer, while the implemented algorithm uses a diffusion loss and adaptive gradient balancing.
  • All experiments use synthetically injected Gaussian or Student-t noise, leaving effectiveness under naturally occurring tail risks unclear.
  • With only 80 evaluation scenarios per input, CVaR estimates at α=0.1\alpha = 0.1 rely on roughly eight tail samples. This raises concerns about reference-regret accuracy, particularly under heavy tails, which also fall outside the theory's bounded-cost assumptions.
Questions for authors
  • Can the authors provide an argument connecting the implemented diffusion objective in Eq. (7), to the oracle-KL regularized learner analyzed in Theorem 4.1? If this is not possible, which guarantees apply to the actual training procedure?
  • How stable are the Table 1–3 regret rankings when the reference CVaR is estimated with substantially more than 80 scenarios per test input (e.g., 1,000 or more)? Please report the Monte Carlo uncertainty of the reference regret estimates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.