acceptodds
Under review as a conference paper at ICLR 2027

The Orthogonality Frontier: Reliability-Adaptive Causal Decisions with Many Treatments

Abstract

Large treatment libraries make dense prediction scalable while causal verification remains scarce. Even with a preliminary predictor and a doubly robust score in hand, a decision maker must determine how strongly causal discrepancies should revise the final normalized target. To formalize this choice, we introduce reliability-adaptive target orthogonalization (RATO), which represents the correction as an operator on the treatment-contrast quotient and distinguishes two criteria for choosing it: first-order orthogonality and local target risk. Full correction, , uniquely removes first-order preliminary-predictor sensitivity within the target family. In contrast, minimum local target risk selects under a zero cross-second-moment condition, where and capture both variance and persistent local bias. Under these local-risk conditions, whenever is not a scalar multiple of , the directional oracle has strictly lower local risk than every scalar correction. Because the required total error moments are not generically identified from routine deployment data, independent validation of the scalar family provides the default deployable route. Controlled experiments examine the frontier against exact nonlinear target error. In a retrospective participant-randomized study with 72 actions per context, increasing correction-and-tuning budget with preliminary fitting and test data fixed moves mean selected correction from at to at . A second, weak-signal study shows stronger correction without higher held-out value. The frontier allocates available causal information; it does not create treatment signal.

Primary area:
causal reasoning
Keywords:
causal inference, generated text treatments, orthogonal learning, policy learning, doubly robust estimation, semantic treatment effects, decision-focused learning, regret calibration

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_B5iFOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies policy learning with many treatments, given an ML predictor and a doubly robust score. The paper proposes a family of first-order corrections to a softmax policy target and derives the optimal weight matrix for a local quadratic target and orthogonality. They show that $A=I$ uniquely achieves first-order orthogonality with respect to the preliminary predictor, whereas minimizing local quadratic target risk yields A⋆=B(B+S)−1A^\star=B(B+S)^{-1}, which balances prediction error against causal-score noise.

Critical strengths
  1. The problem is well-motivated. When there is multiple treatments, using a double robust estimator alone can have large variance, and using an ML predictor alone can have large bias.

  2. This paper derives the closed-form characterization of the weight matrix A for orthogonality and local quadratic loss.

Critical weaknesses
  1. This paper only considers the softmax of the policy value, which is only an approximation of the optimal policy value. I wonder if it is possible to generalize the results to the optimal policy value.

  2. Similar techniques have already appeared in the literature. For example, [1] uses softmax smoothing to develop an inference method for the optimal policy value. This paper does not discuss the connection and differences with this paper. At a rough glance, [1] seems to have a similar orthogonal estimator and has developed an inference theorem for their estimator.

  3. This paper discusses the optimal weight for the local MSE weight, but under that weight (or even for the proposed weight A=αIA=\alpha I), the estimator is not orthogonal, and it is hard to do inference. I wonder if there is a way to do inference while still achieving a good MSE loss.

[1] Whitehouse, J., Chen, Q., Austern, M., and Syrgkanis, V. (2025). “Inference on Optimal Policy Values and Other Irregular Functionals via Softmax Smoothing.” arXiv preprint, arXiv:2507.11780.

Questions for authors

Please see the critical weakness section.

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

The paper combines preliminary treatment outcome predictions with causal estimates that are noisy because of limited samples. It studies how much the causal estimates should correct the preliminary predictions. Full correction removes first order sensitivity to prediction error, but partial correction can have lower error.

Critical strengths
  1. a useful problem, especially when a small randomized sample is divided among many treatments.
  2. Clear explanation of why full correction may introduce too much noise.
  3. Included the cases where stronger correction does not improve the outcome.
Critical weaknesses
  1. The optimal correction requires errors relative to the unknown true treatment values. Difficult to see how these quantities would be obtained in practice.
  2. The theory allows different correction strengths across treatment comparisons, but the real-data method only selects one common strength. A gap between the main theoretical result and the method actually used.
  3. No consistent improvement over the simpler methods in the real data experiments. Both estimates also come from randomized data, so the main issue seems to be finite sample error rather than causal identification.
Questions for authors
  1. Why correct the treatment probabilities instead of combining the treatment outcome estimates first, then applying softmax?
  2. How could the error quantities needed for different correction strengths be estimated in a real application?
Reviewer_wsCGOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a correction operator A on the (K−1)-dimensional treatment-contrast quotient, applied to a softmax target as T_A = q_λ(v̂) + J_λ(v̂)UAUᵀ(Φ−v̂), where v̂ is a dense preliminary predictor and Φ a doubly robust score. Two criteria are separated: first-order orthogonality uniquely selects A = I (Prop. 2), while local target risk selects the oracle A* = B(B+S)^{-1} under zero cross second moment (Thm. 4), with scalar optimum a* = tr(WB)/(tr(WB)+tr(WS)) and strict directional dominance when A* is non-scalar (Cor. 5–6). Since B, S are not identified from routine data, the deployable procedure restricts to A = aI and selects (a, λ) on validation (Alg. 1, Thm. 7). Experiments are Gaussian synthetic perturbations of the paper's own local model plus two retrospective reanalyses of randomized LLM-message studies.

Critical strengths

S1. The orthogonality-vs-risk distinction is clean and worth stating: A = I removes first-order preliminary sensitivity but is generally not risk-optimal. The contrast-quotient geometry (J_λ1 = 0) is correct, and Prop. 8's cycle condition (Σ logit r = 0 around triples) is a nice reason why K ≥ 3 couples comparisons.

S2. The math checks out. I verified Prop. 2's derivative, Thm. 4's excess-risk identity R(A)−R(A*) = tr{W(A−A*)(B+S)(A−A*)ᵀ}, Cor. 5, Cor. 6 (spectrum of BM^{-1} in [0,1]), Cor. 11, Prop. 14, and Thm. 13's two-point argument.

S3. App. E (stratified-cluster CLT + frozen-policy bootstrap, Thm. 15/Prop. 16) is the most valuable content in the paper and is under-sold.

S4. The candour is exemplary: App. I and F.9 state plainly that the oracle is not identified and that the real-data results do not support superiority.

Critical weaknesses

W1. The headline object is unidentifiable and the fallback is not a contribution. B, S are total second moments relative to unknown v, so A* is an oracle (the paper says so). Algorithm 1 is then a 1-D grid search over (a, λ) — generic hyperparameter selection — and App. I concedes Thm. 7 does not certify the selector actually implemented (context-standardized ratios). The only guarantee covers a procedure that is not the one run. W2. The synthetic experiments instantiate the theorem's own assumptions. The DGP (Eq. 35: v̂ = v+tUξ, Φ = v+tUζ, Gaussian with known moments) is literally the local experiment of Thm. 4, with oracles from population truth. This verifies algebra, not applicability: misspecification, overlap, nuisance error, and K are not varied in ways that could break the local approximation. W3. The real-data evidence points the wrong way. At full budget â = 1, so RATO = Full by construction. Roleplay: .5456 vs plug-in .5499. Scaling: .5525 with CI [.5081,.5955] spanning everything. Paired differences vs plug-in are +.00632 [−.03488,.04681] and −.00429 [−.01446,.00586]. On the decoded endpoint (Table 14) RATO loses clearly on both (.5399 vs .5744; .5422 vs .5564). Across-thinning bands are an order of magnitude wider than any method gap. No setting shows a benefit.

Questions for authors

Q1. What minimal calibration design identifies B and S, concretely? Absent one, what should a reader take from contribution (i) beyond "the Gauss–Markov gain is optimal in contrast coordinates"? Q2. What guarantee, if any, covers Algorithm 1 as implemented (ratio selector with realized denominators)? If the answer is post-selection bootstrap, please say so in the main text.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.