acceptodds
Under review as a conference paper at ICLR 2027

Functional Martingale Coherence and Amortized Uncertainty for Frozen Forecasters

Abstract

Predictive resampling treats unobserved future data as a source of uncertainty, recursively imputing observations from a forecaster to induce a distribution over a target functional of interest. However, a generic pretrained forecaster need not be coherent under its own imputations. Sampled continuations can systematically shift the target, conflating future uncertainty with self-induced drift. We first develop a calibration method that constructs functionally coherent continuations while keeping the forecaster fixed. Exact one-step responses to candidate imputations enable a KL-minimal reweighting of the imputation distribution to remove self-induced drift when feasible. Same construction naturally extends to local covariance calibration, but local control alone does not determine uncertainty over an entire continuation. We derive an exact finite-horizon decomposition of target covariance into local variation and residual drift. Building on this, we introduce Amortized Martingale Coherence (AMC), which learns to estimate the resulting covariance target from current-state information without repeated future rollouts. Across tabular and language forecasters, functional repair reduces self-induced drift while preserving the frozen predictor. The resulting finite-horizon uncertainty captures future belief instability beyond local covariance, and AMC can predict much of this signal from the current state without future rollouts.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
martingale posterior, predictive resampling, predictive coherence, amortized inference, uncertainty quantification

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jbKcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for repairing functional coherence and estimating uncertainty in frozen pretrained forecasters. The authors formulate coherence repair as a KL-minimal reweighting problem, enforcing zero expected changes in a chosen target functional, plus extensions to covariance constraints and approximate repair. The authors further introduce Amortized Covariance Estimation (ACE), which predicts finite-horizon uncertainty from current candidate responses without requiring future sequential rollouts at inference time.

The approach is evaluated with TabPFN classifiers and small reasoning LLMs. The reported results demonstrate substantial reductions in defined self-induced drift and gains in predicting future uncertainty and computational efficiency relative to explicit resampling.

Critical strengths
  1. Meaningful problem.\textbf{Meaningful problem.} The paper identifies a motivated issue in predictive resampling with pretrained models, where repeated self-generated imputations may systematically shift the quantity of interest. The distinction between the forecaster's original predictive distribution and the induced continuation law provides a clear motivation for coherence repair.

  2. Connection between coherence and uncertainty estimation.\textbf{Connection between coherence and uncertainty estimation.} he analysis connects martingale coherence with accumulated finite-horizon covariance, motivating the estimation of future uncertainty beyond the current local covariance. ACE provides a practical approach to amortizing this otherwise expensive computation.

  3. Promising potentials.\textbf{Promising potentials.} The evaluations include tabular classifiers and reasoning LLMs, with reported improvements in drift reduction, future-variation estimation, and inference efficiency support.

Critical weaknesses
  1. Limited differentiation from established optimization and probability theory.\textbf{Limited differentiation from established optimization and probability theory.} The main coherence-repair formulation is essentially a classical KL information projection under moment constraints, whose exponential-tilting solution and convex-hull feasibility conditions are well established. Similarly, the finite-horizon covariance decomposition and Gaussian limit largely follow standard martingale theory under the stated assumptions.

  2. Functional coherence does not essentially imply improved uncertainty quality.\textbf{Functional coherence does not essentially imply improved uncertainty quality.} The proposed repair enforces a martingale condition for a selected readout under the modified continuation law. However, this property does not establish predictive coherence for the full distribution or calibration with respect to the underlying data-generating process. Although the paper acknowledges this distinction, the nonlinear synthetic comparisons in Appendix F.2 do not show uniformly improved ground-truth coverage or prediction error after repair.

Questions for authors
  1. As stated in weakness point 1, the reviewer is more familiar with optimization; the principal appears to lie in the problem formulation and integration with amortized uncertainty estimation rather than in the underlying optimization machinery. Personally, I think the most interesting takeaway is rather the problem definition: frozen forecasters can systematically alter a selected prediction when repeatedly conditioned on self-generated observations.

  2. Empirical generality remains somewhat limited; e.g., the LLM uncertainty experiments consider relatively short continuation horizons. Since repairability and computational cost depend on candidate support and augmented-context evaluations, it remains unclear how well the proposed framework extends to other model families or more complex continuation settings.

  3. How would the authors explain the gap between asymptotic guarantees and finite-horizon uncertainty estimation? The Gaussian approximation relies on martingale limit conditions and sufficiently accurate covariance estimation. However, these assumptions are not directly established for the evaluated finite-horizon settings, particularly when approximate repair introduces residual drift. And perhaps undercoverage issues in some setups?

6 comments

[W1, Q1] Contribution beyond Classical Information Projection

Thank you for evaluating our work from an optimization perspective and for highlighting the importance of the problem formulation. While our approach builds on classical information projection, its central contribution is to formulate how predictive resampling can induce uncertainty without systematically shifting the selected readout of a frozen forecaster, and to connect this repair to efficient future uncertainty estimation.

Specifically, we reweight the imputation law while preserving the forecaster's responses to augmented contexts. Exact repair enforces functional coherence when feasible, while our approximate repair formulation provides a means to control residual drift when exact constraints are infeasible. Using in-context influences computed from exact augmented-context responses, we construct the required moment constraints without retraining the forecaster. Our conditional repair formulation further allows these constraints to be enforced while preserving the prescribed covariate marginal.

Importantly, enforcing functional coherence does not make future uncertainty directly available from the current state. Even when the current local covariance Γ0\Gamma_0 is known, the finite-horizon centered covariance Sn,HS_{n,H} depends on how the forecaster responds to subsequent imputations. This motivates ACE, which learns to estimate future centered variation from current candidate responses without requiring sequential future rollouts at inference time. Together, coherence repair and ACE provide a unified approach to constructing and efficiently estimating uncertainty under a frozen forecaster.

Our additional experiments further support these contributions. Direct evaluation against Sn,HS_{n,H} yields a Spearman correlation of 0.901 for ACE, compared with 0.826 for Local covariance and 0.854 for Hidden + Local. Moreover, controlled comparisons with direct endpoint regression show improved CRPS across the evaluated horizons, including an 8.2% reduction at H=8H=8 under limited rollout supervision, confirmed on independent evaluation states (Holm-adjusted p=0.0004p=0.0004). These results demonstrate the practical value of the proposed covariance-based uncertainty estimation beyond the underlying information-projection formulation.

We will clarify these contributions in the final manuscript, emphasizing the formulation of functional coherence for frozen forecasters, the response-preserving repair mechanism, and its integration with amortized future uncertainty estimation.

[W2] Functional Coherence and External Uncertainty Quality

Thank you for highlighting the distinction between functional coherence and calibration to the underlying data-generating process. Our approach enforces coherence for a selected readout while preserving the frozen forecaster, not imposing coherence on its entire predictive distribution. Under exact repair, the readout has zero conditional mean drift, providing a martingale-based uncertainty construction without guaranteeing external calibration.

We acknowledge that the nonlinear comparisons in Table 7 do not establish a statistically significant improvement in ground-truth coverage or probability error (Repaired − Raw: coverage −0.006 [−0.081, +0.069], interval width +0.028 [−0.010, +0.050], and probability error +0.025 [−0.004, +0.054]). This is consistent with the scope of functional coherence. Under exact repair, the expected future readout remains equal to the current prediction, so any initial bias relative to the true probability is not automatically corrected by self-imputation.

Nevertheless, our additional experiments provide evidence of inferential benefits in specific settings. On three logistic data-generating tasks with known ground-truth probabilities, repair improves coverage from 0.722 to 0.778 while reducing interval width by 0.090 and CRPS by 0.012, with paired confidence intervals excluding zero, as reported in Reviewer Cf7r [W1, Q1]. In a separate evaluation on nonlinear tasks, repair also reduces conditional drift and achieves lower CRPS than both raw continuation and post-hoc re-centering, as shown in Table R.3. On LLM reasoning tasks, repaired continuations also achieve lower NLL and ECE for continuation-averaged answer probabilities across both models at H=2H=2 and H=4H=4, with paired confidence intervals excluding zero and no statistically detectable change in final-answer accuracy (Table R.6, Reviewer RABn [W2, Q2]).

These findings demonstrate that functional coherence can lead to improved uncertainty quality, while not implying universal calibration guarantees. We will clarify this distinction and incorporate the additional evaluations into the final manuscript.

[Q2] Longer Horizons, Candidate Support, and Computational Scope

Thank you for raising the questions about longer continuation horizons, candidate support, and computational scalability. To extend our LLM evaluation, we conducted an additional H=8H=8 evaluation, comparing native and repaired continuations.

Table R.7. Additional LLM evaluation at H=8H=8.

ModelSquared drift reductionShare of S8S_8 from steps 5–8Accuracy Q−PQ-P (pp) [95% CI]
Llama54%7.4%+0.3 [−2.9, +3.5]
OLMo24%15.2%−2.2 [−5.2, +0.7]

At H=8H=8, repair reduces squared conditional mean shifts without a statistically significant accuracy difference. We also found out that the H=2H=2 ACE predictor transfers to H=8H=8 via the horizon rescaling in Appendix G.2 without retraining. On independent repaired continuations, ACE outperforms the calibrated local-covariance baseline for one-pool Llama (squared-error difference −3.8×10−4-3.8\times10^{-4}, 95% CI [−7.1,−0.6]×10−4[-7.1,-0.6]\times10^{-4}), with significant gains also in the evaluated OLMo comparisons. These results extend the empirical evidence for drift reduction and amortized uncertainty estimation over different horizons, without retraining the ACE predictor.

Candidate support and computational cost.

We further evaluated the feasibility of exact mean repair at fixed reasoning prefixes while varying the candidate pool size from 8 to 96 blocks. At initial prefixes, feasibility increases from 78% to 98% for Llama and from 65% to 91% for OLMo. After two reasoning blocks, the corresponding rates increase from 26% to 49% and from 27% to 46%, respectively. These results show that larger candidate pools improve repairability.

As analyzed in Appendix G.4, ACE requires O(M)O(M) current augmented-context readouts, whereas explicit estimation through RR repaired rollout paths of horizon HH requires up to O(RHM)O(RHM) candidate generations and readouts. Thus, increasing candidate support incurs additional probing cost, while ACE avoids future sequential rollouts at inference without eliminating current probing or offline supervision. We will incorporate these additional findings and clarify the computational trade-offs in the final manuscript.

[Q3] Finite-Horizon Uncertainty, Gaussian Approximation, and Undercoverage

Thank you for raising the important questions. We clarify that our finite-horizon covariance analysis and Gaussian approximation rely on different levels of assumptions.

Finite-horizon covariance and Gaussian limits.

As established in Proposition D.3, the decomposition under approximate repair remains valid without invoking a CLT.

\mathbb{E}[M_{n,H}\mid\mathcal F_0]=0, \qquad \operatorname{Var}(M_{n,H}\mid\mathcal F_0)=S_{n,H}.$$ Our additional diagnostics support this finite-horizon identity, with ratios of realized centered second moments to accumulated covariance ranging from 0.963 to 1.032 across the evaluated horizons up to $H=128$. Thus, $S_{n,H}$ characterizes the accumulated centered variation even when approximate repair introduces residual drift. In contrast, Theorem 4.1 and Proposition D.1 require exact mean repair and additional martingale CLT conditions. Since exact repair is **not always feasible under finite candidate supports**, practical continuations may require approximate repair. Consequently, these asymptotic conditions cannot be assumed to hold. Therefore, we further assess the endpoint Gaussian directly as an empirical finite-horizon approximation. **Understanding the observed undercoverage.** To investigate the 84.1% coverage of nominal 90% intervals, we separately replaced the learned endpoint mean and variance with independently estimated reference moments. **Table R.8.** Coverage of nominal 90% intervals under endpoint-moment substitutions. | Gaussian approximation | Coverage | |-|-| | Learned mean and variance | 0.841 | | Reference mean, learned variance | 0.852 | | Learned mean, reference variance | 0.850 | | Reference mean and variance | **0.903** | Replacing both learned moments with reference estimates increases coverage from 84.1% to 90.3%, close to the nominal level. The individual substitutions also improve coverage, indicating that **errors in estimating both the endpoint mean and variance contribute to the observed undercoverage**. Further analysis of independent rollouts finds 90.3% and 94.1% coverage for nominal 90% and 95% intervals with reference moments, comparable to Gaussian finite-sample benchmarks of 88.5% and 93.7% using 32 paths for moment estimation and 128 for evaluation. These findings support the **finite-horizon interval adequacy of a moment-matched Gaussian**. We will clarify how the finite-horizon covariance identities and conditional Gaussian limits provide a theoretical foundation for the empirically supported finite-horizon approximation.
Reviewer_jbKcOpenReview ↗

Re: Official Comment by Authors

Thank you for the response and additional experiments.

The additional H=8 experiments provide useful evidence of longer-horizon applicability, although the gains remain relatively modest. Two limitations nevertheless remain: (1) exact repairability degrades substantially along LLM continuations, even with larger candidate pools, making the frequency and impact of fallback steps under the actual 16-candidate setting important to clarify; and (2) the learned endpoint Gaussian still achieves 84.1% coverage against a nominal 90%, despite the useful diagnosis provided by the reference-moment experiments.

So, I am therefore inclined to maintain my rating.

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

This paper studies coherence in predictive resampling with frozen forecasters. It proposes a KL-based repair that reweights imputations to remove expected drift in a chosen readout. It also introduces ACE to predict future uncertainty without additional rollouts. Experiments are conducted on TabPFN and reasoning LLMs.

Critical strengths
  1. This paper presents the feasibility conditions, exact and approximate solutions, and a finite-time covariance analysis for coherence repair. The distinction between functional coherence and predictive calibration is clearly established.
  2. On TabPFN, repair reduced the normalized drift from 0.38 to 0.08. Compared to local covariance, ACE increased the Spearman correlation coefficient from 0.789 to 0.844 and raised the AUROC from 0.892 to 0.931.
Critical weaknesses
  1. In Table 7, coverage on nonlinear tasks remains low (45.6% before repair v.s. 45.0% after). Median error also increases from 0.18 to 0.21. Moreover, joint repair preserving the original covariance is feasible on only 71.6% of evaluated steps. Covariance projection increases feasibility, but the resulting distortion is not quantified.
  2. ACE is trained to predict centered covariance but evaluated mainly against endpoint variance. These can differ under approximate repair (Appendix F.3). The endpoint Gaussian also achieves only 84.1% coverage for nominal 90% intervals (Table 4). For LLMs, the moment-ratio results mainly verify a theoretical identity. Direct drift comparisons and repair feasibility rates are missing. ACE's LLM gains are also modest (3.0-4.4%).
Questions for authors
  1. Can the authors demonstrate that consistency repair yields improvements in uncertainty quality beyond simple re-centering? To what extent does covariance projection introduce distortion?
  2. Can the authors directly evaluate ACE by comparing it to its centered covariance objective? For LLMs, compared to native sampling, what are the repair success rate, regression frequency, and reduction in drift?
2 comments

[W1, Q1] Uncertainty Quality Beyond Re-centering and Covariance-Projection Distortion

Thank you for raising the questions of whether coherence repair provides benefits beyond simple re-centering and how much covariance projection changes the original uncertainty. We conducted additional comparisons and quantified the covariance distortion using the evaluated continuation steps.

Repair versus re-centering.

In a separate H=32H=32 evaluation from Table 7, we compared raw continuation, post-hoc endpoint re-centering, and sequential repair on matched initial states using the same predictive-probability readout and covariate marginal. The re-centering baseline shifts the completed raw endpoint samples so that their mean matches the initial prediction, without modifying the sampled trajectories or their intermediate conditional drift.

Table R.3. Comparison with endpoint re-centering on nonlinear tasks (H=32H=32).

MethodConditional step drift ↓Coverage ↑Δ\DeltaCRPS vs. Raw ↓
Raw continuation0.01000.494—
Re-centering0.00990.512−0.005 [−0.013, +0.003]
Repaired continuation0.00300.550−0.012 [−0.019, −0.005]

Sequential repair reduces conditional step drift, defined as the average magnitude of the expected one-step readout change at each visited state, by approximately 70%, whereas re-centering leaves it essentially unchanged. Repair also increases observed coverage from 0.494 to 0.550 and achieves lower CRPS than both raw continuation (−0.012 [−0.019, −0.005]) and re-centering (−0.007 [−0.012, −0.003]). Unlike post-hoc re-centering, which only adjusts the endpoint distribution, sequential repair modifies the imputation law at each visited context to control conditional drift. These results demonstrate that repair provides not only a distinct conditional-coherence property but also improved uncertainty quality over re-centering in the evaluated setting. We will add this comparison to the updated manuscript.

Covariance-projection distortion.

We agree that satisfying the selected covariance constraint does not necessarily preserve the original covariance. We therefore quantified the discrepancy between the reference covariance Γtref=Cov⁡qt(ψt)\Gamma_t^{\mathrm{ref}}=\operatorname{Cov}_{q_t}(\psi_t) and the covariance target selected by repair, as well as the covariance actually achieved under QtQ_t. In these experiments, the readout is a scalar predictive probability at a fixed query, so the covariance targets are scalar variances.

Across evaluated steps, 71.6% admit exact mean-covariance repair while preserving the original covariance. An additional 21.9% achieve exact joint repair with a projected covariance target, while the remaining 6.5% require approximate repair. Among projected steps, the selected variance is typically smaller than the reference variance (median ratio 0.706). Along the same visited continuation paths, the accumulated achieved covariance is 8.3% lower than the corresponding accumulated reference covariance, with a Spearman rank correlation of 0.996. Thus, although projection can modify individual covariance targets, the relative ordering of accumulated uncertainty is largely preserved. We will include these distortion diagnostics and clarify their implications in the final manuscript.

[W2, Q2] Direct Evaluation against Centered Covariance

We agree that evaluating primarily against endpoint variance did not directly answer how accurately ACE predicts its training target. We now evaluate ACE against independent 128-path estimates of the centered covariance Sn,HS_{n,H} at H=32H=32 across the 30 datasets.

Table R.4. Prediction of the centered covariance Sn,32S_{n,32}.

MethodSpearman with SS ↑High-SS AUROC ↑
Margin0.6600.793
Hidden-only0.7700.855
Local covariance0.8260.910
Hidden + Local0.8540.916
ACE0.9010.955

ACE improves Spearman over Local by +0.075 [0.052, 0.099], with improvement on 27/30 datasets, and over Hidden + Local by +0.046 [0.025, 0.070], with improvement on 23/30 datasets. These results directly demonstrate ACE's effectiveness in predicting its intended centered covariance target Sn,HS_{n,H}. We will include this target-aligned evaluation in the final manuscript, while retaining the endpoint-variance analysis as a complementary evaluation of future readout instability. Importantly, as shown in Proposition D.3 (Eq. (96)), endpoint variance and Sn,HS_{n,H} need not coincide under approximate repair due to accumulated residual drift and its cross-covariance with the centered martingale component.

[W2, Q2] LLM Repair Success, Regression Frequency, and Drift Reduction

Thank you for pointing out the need to evaluate coherence repair directly against native LLM continuations, beyond verifying the centered-variance identity. We compared native (PP) and repaired (QQ) continuations on the same initial reasoning prefixes, expanding the H=4H=4 evaluation.

Repair feasibility, drift reduction, and answer quality.

For each initial reasoning prefix D0D_0, we computed the conditional mean shift ER[θH−θ0∣D0]\mathbb{E}_R[\theta_H-\theta_0\mid D_0] under R∈{P,Q}R\in\{P,Q\} and averaged its squared magnitude across prefixes. This measures accumulated drift directly, beyond the centered-variance identity. We also evaluated NLL and ECE (expected calibration error) using answer probabilities averaged over continuations.

Table R.5. Native versus repaired LLM continuations.

ModelHHExact repair at final stepPaths with ≥1\geq 1 fallbackSquared drift reductionAccuracy Q−PQ-P (pp) [95% CI]
Llama249%56%50%−0.3 [−1.5, +0.9]
Llama446%69%44%−0.6 [−1.8, +0.5]
OLMo252%60%33%−0.5 [−1.6, +0.5]
OLMo437%71%25%−0.8 [−1.6, +0.1]

Table R.6. Probabilistic answer quality under native and repaired continuations. Lower values are better, and negative Q−PQ-P differences favor repair.

ModelHHNLL (Current / PP / QQ)NLL Q−PQ-P [95% CI]ECE (Current / PP / QQ)ECE Q−PQ-P [95% CI]
Llama21.17 / 1.05 / 0.91−0.13 [−0.16, −0.10]0.150 / 0.108 / 0.075−0.033 [−0.040, −0.022]
Llama41.14 / 1.09 / 0.95−0.14 [−0.20, −0.09]0.144 / 0.093 / 0.062−0.031 [−0.046, −0.022]
OLMo22.04 / 1.20 / 1.03−0.18 [−0.23, −0.13]0.245 / 0.128 / 0.085−0.042 [−0.053, −0.030]
OLMo42.07 / 1.39 / 1.14−0.26 [−0.40, −0.12]0.259 / 0.115 / 0.085−0.030 [−0.053, −0.020]

Repair reduces squared conditional mean shifts by 25–50%, although 56–71% of paths encounter at least one fallback. While final-answer accuracy differences are not statistically significant, repaired continuations consistently achieve lower NLL and ECE than native continuations, with all paired confidence intervals excluding zero. These statistically significant improvements in probabilistic prediction quality and calibration provide direct empirical evidence that repair improves uncertainty quality, beyond simply reducing self-induced drift.

To quantify answer regressions, we measured initially correct top answers becoming incorrect after continuation. At H=2H=2, these transitions decrease from 13.8% to 10.9% for Llama and from 15.8% to 11.7% for OLMo. Incorrect-to-correct transitions also decrease, indicating fewer answer changes in both directions. Thus, repair allows uncertainty to be estimated through resampling while reducing systematic shifts in the current belief.

ACE uncertainty estimation.

We also evaluated ACE against the calibrated local covariance baseline under the same current probing budget. At H=2H=2, ACE reduces squared prediction error by 3.8% for Llama and 3.4% for OLMo. On the enlarged H=4H=4 evaluation, the H=2H=2 predictor transfers using the horizon rescaling in Eq. (120) without retraining. For one candidate pool, the paired ACE − Local squared-error differences are −7.06×10−4-7.06\times10^{-4} (95% CI [−8.32,−5.78]×10−4[-8.32,-5.78]\times10^{-4}) for Llama and −5.85×10−4-5.85\times10^{-4} (95% CI [−8.21,−3.52]×10−4[-8.21,-3.52]\times10^{-4}) for OLMo. These results support ACE's ability to estimate future centered variation beyond the current local covariance.

Regarding the endpoint Gaussian's 84.1% coverage for nominal 90% intervals, replacing the learned endpoint moments with independent reference estimates increases coverage to 90.3%, indicating that moment-estimation errors contribute to undercoverage (see Reviewer jbKc [Q3]).

Together, these results provide direct evidence that coherence repair reduces self-induced drift while improving probabilistic answer quality, and that ACE improves future centered-variation estimation. We will incorporate these additional findings and clarify the scope of the evaluated settings in the final manuscript.

Reviewer_Cf7rOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies uncertainty induced by predictive resampling from frozen forecasters, where self-generated observations can systematically shift a chosen readout. It proposes KL-minimal reweighting of candidate imputations to eliminate conditional mean drift, with extensions to covariance constraints and approximate repair. It then introduces Amortized Covariance Estimation (ACE), which predicts finite-horizon accumulated centered variation from current candidate responses, avoiding future sequential rollouts at inference. Experiments with TabPFN and two LLM families demonstrate drift reduction and improved prediction of future variation; an additional endpoint-moment predictor provides a fast approximation to finite-horizon readout distributions.S1. A precise and interpretable intervention for controlling functional coherence. The formulation separates the forecaster’s responses from the continuation law: it preserves responses to augmented contexts while adjusting candidate probabilities through a KL projection. The explicit feasibility conditions and extension preserving the covariate marginal make the scope of the intervention clear. Table 2 provides concrete evidence of its effectiveness, reducing average normalized endpoint drift from 0.38 to 0.08 across eight TabMGP-aligned settings. This is a useful contribution for predictive-resampling methods that require a centered readout process. S2. A useful distinction between local and future uncertainty, supported by empirical controls. The paper distinguishes current local covariance, accumulated centered variation, and endpoint variance under residual drift. This clarification is consequential for interpreting uncertainty from predictive resampling. Across 30 TabPFN datasets, ACE improves macro Spearman correlation with future endpoint variance from 0.789 to 0.844 and AUROC from 0.892 to 0.931 relative to local covariance. Independent rollout references and hidden-state controls strengthen the evidence that current responses contain predictive information beyond their local covariance.

Critical strengths

S1. The formulation separates the forecaster’s responses from the continuation law: it preserves responses to augmented contexts while adjusting candidate probabilities through a KL projection. The explicit feasibility conditions and extension preserving the covariate marginal make the scope of the intervention clear. Table 2 provides concrete evidence of its effectiveness, reducing average normalized endpoint drift from 0.38 to 0.08 across eight TabMGP-aligned settings. This is a useful contribution for predictive-resampling methods that require a centered readout process. S2. The paper distinguishes current local covariance, accumulated centered variation, and endpoint variance under residual drift. This clarification is consequential for interpreting uncertainty from predictive resampling. Across 30 TabPFN datasets, ACE improves macro Spearman correlation with future endpoint variance from 0.789 to 0.844 and AUROC from 0.892 to 0.931 relative to local covariance. Independent rollout references and hidden-state controls strengthen the evidence that current responses contain predictive information beyond their local covariance.

Critical weaknesses

W1. The inferential benefit of the repaired continuation remains insufficiently established. Eliminating drift ensures internal consistency for a selected readout, but does not establish that the resulting uncertainty distribution is preferable for inference. Although the paper acknowledges this distinction, the experiments provide limited evidence resolving it. In Table 7, nonlinear-task macro coverage of the true probability changes from 0.456 under Direct qq to 0.450 under Direct QQ, while probability error increases from 0.18 to 0.21. These results do not invalidate coherence repair, but leave its benefit beyond satisfying the imposed constraint unclear. The accuracy improvements in Appendix G.3 concern prediction of the native continuation’s future mean and therefore do not establish a benefit of repair itself. This limits the practical significance of the repair component, particularly for reasoning, where systematic belief changes may reflect useful computation.

W2. The practical advantage of the proposed covariance structure over direct supervised prediction is not convincingly isolated. Table 9 reports CRPS 0.0441 for the adapted ACE endpoint predictor versus 0.0446 for direct endpoint-moment regression. This small point-estimate difference does not establish a reliable advantage of the proposed parameterization. Furthermore, the endpoint predictor’s nominal 90 percent intervals achieve 84.1 percent coverage, and the stated noninferiority criterion against the rollout reference is not met. The inference-time savings are valuable, but are also compatible with generic amortization of future moments. Consequently, the most application-oriented results leave unclear how much the proposed structure contributes beyond providing an interpretation and training target.

Questions for authors
  • Beyond satisfying the selected moment constraint, what criterion makes the repaired continuation preferable to the original continuation for the intended uncertainty task? Is there evidence in the submitted experiments that isolates a benefit of repair itself from the benefits of changing the readout or predicting native future answers? This would directly affect my assessment of W1.

  • In Table 9, do direct endpoint-moment regression and the adapted ACE endpoint model use the same current-state features, comparable encoder capacity, rollout supervision budget, and model-selection protocol? What is the paired uncertainty of their CRPS difference, including variation across training seeds? A consistent advantage under matched conditions would materially affect my assessment of W2.

2 comments

[W1, Q1] Inferential Benefits of Coherence Repair

Thank you for asking how repair benefits uncertainty estimation. Predictive resampling quantifies uncertainty through self-imputed observations, but these imputations can systematically shift the forecaster's current readout despite providing no new external evidence. Our goal is to estimate uncertainty around the current readout while controlling self-induced drift, not necessarily to improve predictive accuracy.

Our experiments isolate the effect of repair by retaining the same target readout while changing only the continuation law from qtq_t to QtQ_t. When exact repair holds throughout the continuation, the readout forms a martingale, so its future conditional mean equals the current prediction and its finite-horizon variance equals the accumulated conditional covariance Sn,HS_{n,H}. This enables uncertainty to be induced through resampling without systematic self-induced drift.

From a martingale-posterior perspective, this distinction matters because resampled variation should characterize uncertainty around the current readout, not systematic shifts induced by self-imputation. On eight real-world datasets using TabMGP's loss-defined functional, Table 2 shows that repair reduces normalized endpoint drift from 0.38 to 0.08, while maintaining similarly high 95% joint coverage and comparable covariance traces. LLM experiments also show fewer correct-to-incorrect transitions at H=2H=2 (Llama: 13.8% to 10.9%; OLMo: 15.8% to 11.7%), without a statistically significant change in overall accuracy (see Reviewer RABn [W2, Q2]). These results support uncertainty estimation through predictive resampling while limiting unwanted shifts in the current readout.

Predictive performance and inferential benefits.

We evaluated inferential benefits on three synthetic logistic classification tasks with known ground truth. In each task, covariates XX are sampled and binary labels are generated according to Y∣X=x∼Bernoulli(p⋆(x))Y\mid X=x\sim\mathrm{Bernoulli}(p^\star(x)), where p⋆(x)=σ(β0+β⊤x)p^\star(x)=\sigma(\beta_0+\beta^\top x) with task-specific logistic coefficients. This provides known ground-truth conditional probabilities for evaluating the coverage and efficiency of predictive-resampling intervals. We compared raw and repaired continuations using the same frozen forecaster and predictive-probability readout at H=32H=32.

Table R.1. Uncertainty quality on logistic data-generating tasks (H=32H=32).

MetricRepaired − Raw [95% CI]
95% Coverage ↑+0.056 [+0.010, +0.129]
Interval Width ↓−0.090 [−0.133, −0.050]
CRPS ↓−0.012 [−0.019, −0.005]

Repair increases 95% coverage from 0.722 to 0.778, while reducing interval width and improving CRPS. Importantly, the higher coverage is achieved with narrower uncertainty intervals. These results provide direct evidence that coherence repair can improve inferential uncertainty quality in the evaluated settings.

We acknowledge that Table 7 does not show a statistically significant improvement on nonlinear tasks (Repaired − Raw: coverage −0.006 [−0.081, +0.069], interval width +0.028 [−0.010, +0.050], and probability error +0.025 [−0.004, +0.054]). This is consistent with the distinction between functional coherence and external calibration. When exact repair holds throughout the continuation, the expected future readout equals the current prediction, so any initial bias relative to the true probability is not automatically corrected. Similarly, TabMGP empirically demonstrates uncertainty quantification for loss-defined functionals but does not establish improvements in the underlying forecaster's predictive accuracy through self-imputation. This distinction may be particularly relevant in nonlinear settings where the initial predictive readout is misspecified, since improved coherence need not translate into better ground-truth uncertainty quality.

Preserving useful reasoning behavior.

We additionally compared native and repaired LLM continuations on the same reasoning tasks, separately from Appendix G.3. Across Llama and OLMo at H=2H=2 and H=4H=4, repair reduces squared conditional mean shifts by 25–50%, while final-answer accuracy differences are not statistically significant. Moreover, repaired continuations achieve lower NLL and ECE than native continuations across both models and horizons. These results demonstrate that self-induced drift can be substantially reduced while improving probabilistic answer quality without a statistically detectable change in final-answer accuracy. Detailed results are provided in our response to Reviewer RABn [W2, Q2].

Overall, our results support uncertainty estimation through resampling with controlled self-induced drift. The real-data experiments retain comparable uncertainty coverage and covariance scale, while logistic and LLM experiments show improved uncertainty quality.

[W2, Q2] Covariance-Structured Amortization versus Direct Moment Regression

Thank you for highlighting the need to distinguish the contribution of our covariance-based structure from generic supervised amortization, and for requesting a more carefully matched comparison with direct endpoint-moment regression. We have conducted an audit of the original comparison, followed by controlled experiments across five training seeds and additional independent evaluations.

Supervision and comparison fairness.

Importantly, ACE does not require additional observed data or independent future trajectories to obtain its covariance supervision. The same offline predictive-resampling trajectories used to estimate endpoint moments also provide the intermediate covariance information used to construct Sn,HS_{n,H}. From an amortized inference or meta-learning perspective, exploiting these intermediate quantities as auxiliary supervision is part of the proposed method. At inference time, ACE predicts the covariance from current-state responses without accessing future rollouts.

The original comparison already used the same current-state response probes and training rollout trajectories, with matched trainable-parameter counts and hyperparameter-search grids. The ACE endpoint model additionally uses the ACE-predicted S^n,H\widehat S_{n,H} as an anchor for endpoint variance estimation, with supervision derived from the same rollout trajectories. To assess the robustness of this comparison, we repeated the evaluation across five training seeds.

Endpoint prediction and independent evaluation.

At the original H=128H=128 setting, ACE achieves a CRPS of 0.04391, compared with 0.04472 for direct endpoint regression. The paired difference (ACE − Direct) is −0.00081 [−0.00190, −0.00013], corresponding to approximately 2.0% lower CRPS. Our additional analyses suggested that covariance-informed prediction could provide further benefits at shorter horizons. We therefore evaluated shorter horizons on new held-out states using prespecified comparisons.

Table R.2. Paired CRPS comparisons with direct endpoint-moment regression. Negative differences favor ACE.

HorizonΔ\DeltaCRPS [95% CI] ↓pp
H=8H=8−0.00175 [−0.00247, −0.00112]0.0004
H=32H=32−0.00071 [−0.00150, −0.00012]0.037

The H=8H=8 comparison shows an 8.2% reduction in CRPS, consistent across all five training seeds and significant after correction for the prespecified comparisons. The H=32H=32 and H=128H=128 comparisons also favor ACE, with smaller improvements and paired 95% confidence intervals excluding zero. These results support the benefit of covariance-structured amortization over direct moment regression.

We also acknowledge that the endpoint Gaussian reported in the submission achieves 84.1% coverage for nominal 90% intervals and does not meet the prespecified noninferiority criterion against the rollout reference. Additional diagnostics indicate that errors in estimating the endpoint mean and variance contribute to this undercoverage: replacing both with independently estimated reference moments increases coverage to 90.3% (see our response to Reviewer jbKc [Q3]). Thus, our results demonstrate improved amortized approximation over direct regression in the evaluated settings, without establishing a fully calibrated replacement for explicit predictive resampling.

We will incorporate these additional comparisons and diagnostics into the final manuscript and clarify both the benefits and limitations of covariance-structured amortization.

Comments on the paper

2 comments

When Does Coherence Repair Improve Uncertainty Estimation?

Thank you for this concrete question. We investigated when repair improves uncertainty estimation and when native resampling provides useful corrections. We then tested prespecified hypotheses on 576 new initial states across nine tasks (64 per task, 128 paths per law, H=32H=32), holding the forecaster and selected readout fixed.

When does native resampling introduce additional bias?

Let PP and QQ denote the native and repaired continuation laws, θ0\theta_0 the initial readout, p⋆p^\star the ground-truth probability, and μP=EP[θH∣D0]\mu_P=\mathbb{E}_P[\theta_H\mid D_0] the native endpoint mean. We assess how native resampling changes the initial squared bias:

GP=(μP−p⋆)2−(θ0−p⋆)2.G_P=(\mu_P-p^\star)^2-(\theta_0-p^\star)^2.

A positive GPG_P indicates that native continuation increases the initial squared bias, while a negative value indicates that it moves the endpoint mean closer to the truth. We estimated GPG_P on half the native paths and evaluated uncertainty quality on the other half, then exchanged their roles. We classify readouts as initially accurate if ∣θ0−p⋆∣<0.15|\theta_0-p^\star|<0.15.

Table R.9. Uncertainty quality conditional on native squared-bias changes (H=32H=32). Negative CRPS differences favor repair.

ConditionCRPS Q−PQ-P [95% CI]95% Coverage P→QP\to Q
Native continuation increases squared bias−0.0100 [−0.0141, −0.0062]0.531 → 0.577
Native continuation reduces squared bias+0.0140 [+0.0098, +0.0184]0.578 → 0.528

Both directional CRPS comparisons are significant after Holm correction (p=0.0005p=0.0005), as is the coverage improvement when native resampling increases squared bias (p=0.018p=0.018). When the initial readout is accurate and native self-imputation introduces additional bias, repair improves CRPS by 0.0082 [0.0049, 0.0116] and coverage from 0.733 to 0.815. Conversely, native continuation can be preferable when it corrects an initially biased prediction. For initially inaccurate readouts, repair shows no statistically significant average CRPS difference from native continuation (+0.0004 [−0.0023, +0.0031]). This is consistent with the objective of coherence repair: it controls drift induced by self-imputation but is not designed to correct errors already present in the frozen forecaster's initial prediction. Thus, our results identify conditions where repair improves uncertainty quality without requiring improvements in the underlying predictive model.

Explaining the original nonlinear results.

A post-hoc breakdown of the original Table 7 shows that the nearly unchanged nonlinear coverage (0.456 → 0.450) masks opposing effects. Among 82 states where native resampling increases squared bias, repair improves coverage from 0.329 to 0.427. In contrast, among 78 states where native resampling reduces squared bias, coverage changes from 0.590 to 0.474. The pattern is confirmed on new states and observed in eight exploratory logistic and nonlinear task variants.

Interpretation and scope.

The competing effects can be understood through the endpoint bias–variance decomposition. For either continuation law R∈{P,Q}R\in\{P,Q\}, let μR=ER[θH∣D0]\mu_R=\mathbb{E}_R[\theta_H\mid D_0] and VR=Var⁡R(θH∣D0)V_R=\operatorname{Var}_R(\theta_H\mid D_0). Then,

ER[(θH−p⋆)2∣D0]=VR+(μR−p⋆)2.\mathbb{E}_R[(\theta_H-p^\star)^2\mid D_0]=V_R+(\mu_R-p^\star)^2.

Writing MSER\mathrm{MSE}_R for the left-hand side and defining GQG_Q analogously to GPG_P, subtracting the two expressions gives

MSEQ−MSEP=(VQ−VP)+GQ−GP.\mathrm{MSE}_Q-\mathrm{MSE}_P=(V_Q-V_P)+G_Q-G_P.

Exact mean repair throughout the continuation gives μQ=θ0\mu_Q=\theta_0 and GQ=0G_Q=0. It therefore preserves the initial squared bias, preventing harmful changes in the endpoint mean while also suppressing potentially useful corrections from native resampling. Endpoint variance still matters, and the identity alone does not determine CRPS or coverage, which we evaluate empirically. Approximate repair can leave residual drift. Our post-hoc analysis finds larger-magnitude residual endpoint mean shifts in nonlinear tasks than in logistic tasks (−0.034 vs. −0.013), which may partly explain their different uncertainty-quality outcomes.

Together, these results clarify the role of coherence repair in predictive resampling. When the initial readout is close to the truth and native self-imputation increases squared bias, repair improves uncertainty quality. When the initial readout is inaccurate, repair does not resolve the underlying predictive error, but we observe no statistically significant average CRPS difference from native continuation. This highlights the value of controlling self-induced drift while preserving the forecaster, without requiring improvements in its predictive accuracy. Our analysis uses known synthetic ground truth and may motivate future work on distinguishing useful native corrections from self-induced drift without access to that truth. We will incorporate these conditional findings into the revised manuscript.

AuthorsOpenReview ↗

Manuscript Update Incorporating Rebuttal Results

Dear Area Chairs and Reviewers,

We have uploaded an updated PDF incorporating the additional theoretical and empirical analyses presented during the rebuttal. All additions and revisions are highlighted in blue.

In particular, we have added:

  1. Theoretical extensions: Wasserstein bounds under residual drift, a characterization of the limiting distribution, and an explicit Beta–Bernoulli example, together with their proofs.
  2. TabPFN experiments: Additional analyses of uncertainty quality under coherence repair, covariance-projection distortion, direct evaluation of ACE's centered-variation target, endpoint calibration, and computational efficiency relative to explicit rollouts.
  3. LLM experiments: Comparisons of native and repaired continuations, including probabilistic answer quality, fallback behavior, adaptive candidate expansion, and residual-drift diagnostics.

We have incorporated these additional results into the existing manuscript structure. In the final version, we will further refine the organization and presentation to better integrate the new findings and highlight their contributions.

We sincerely thank the reviewers for their constructive feedback, which has helped us strengthen the paper both theoretically and empirically.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.