acceptodds
Under review as a conference paper at ICLR 2027

LTTS: A Predictor-Corrector for Long-Horizon Counterfactual Outcome Prediction

Abstract

Comparing treatment plans requires predicting how outcomes evolve beyond a common observed history. Recursive models construct these trajectories by using generated states as inputs to later predictions. Without new measurements, errors in these inputs can affect the remaining continuation. To address this problem, we propose Long-term Trajectory Stabilization (LTTS), a predictor–corrector for long-horizon outcome prediction under sustained treatment and control from the same observed prefix. We use the predictor (TSA) to advance each treatment-specific trajectory, while the adaptation module (TAM) uses historical reconstruction to adjust selected generated states before they are reused. Using a frozen monitor, we trigger limited private adaptation on the current generated block, retaining the shared reconstruction modules without using new outcome labels. The updated window supports the next prediction; prescribed treatments and reported outcomes are preserved. Across two synthetic benchmarks and a MIMIC-IV-derived Cox model-recovery benchmark, we obtain 4.2–32.6% lower mean NMAE than the best evaluated external baseline per setting in 150-step evaluations over 30 seeds. With the predictor held fixed, we observe lower mean outcome error with feedback. Subsequent predictions also improve after a single correction when further corrections are disabled. Together, these results support combining trajectory prediction with selective state feedback to improve long-horizon outcome prediction under the evaluated treatment continuations.

Primary area:
causal reasoning
Keywords:
Counterfactual outcome prediction, Long-horizon prediction, Autoregressive modeling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_x2BfOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies longitudinal counterfactual outcome prediction under sequential treatemt interventions, an important problem in causal inference. In particular, it attempts to stabilize the prediction over long horizons by monitoring and selectively reconstructing generated intermediate states. The paper uses TSA for the standard g-computation-style sequential prediction and then uses an additional module, called TMA, to audit and correct the intermediate state. Finally, empirical validations suggest that the architecture could be effective.

Critical strengths

The paper identifies an important issue in longitudinal counterfactual outcome prediction: instability over long horizons. To address this challenge, it proposes a new architecture, which both predicts and corrects the intermediate states. The empirical validations are comprehensive and convincing. The proposed architecture could indeed be helpful in long-horizon settings.

Critical weaknesses

Weakness:

  1. The central contribution is a combination of existing techniques, with limited methodological novelty. In addition, one should not confuse "reconstruction ability" with "counterfactual correction ability". The reconstruction error which TAM optimizes, is misaligned with the counterfactual outcome prediction.
  2. No theoretical justification on why the proposed correction mechanism should really work.
  3. The paper does not use standard terminology, randomly replace words with other words, introduces new words without appropriate explaining.
Questions for authors

How can you justify, that by minimizing the reconstruction error, the TMA can be effective in correction of the counterfactual state?

5 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

Thank you for your constructive comments, which have helped us clarify the contribution and improve the presentation. Below, we explain how prediction and feedback work together, outline the supporting intuition and evidence, and clarify the workflow and terminology.

Critical Weaknesses and Questions For Authors

W1a

W1a. The central contribution is a combination of existing techniques, with limited methodological novelty.

Response to W1a. The methodological contribution lies in how we couple trajectory generation with feedback under the information constraints of recursive counterfactual prediction. Our design intuition is to address local prediction error together with the propagation of error through generated inputs. With future observations unavailable, TSA advances the two prescribed treatment branches, while TAM uses historical reconstruction to selectively adjust generated inputs before they are reused. We preserve treatments, the observed prefix, and reported outcomes (§4, pp. 4–5).

Our experiments support the predictive value of this joint design. We observe lower mean NMAE with feedback across prediction depths using the same TSA checkpoints (Figure 2, p. 8). Full also yields lower mean NMAE than validation-selected clipping and prototype-L2L_2 shrinkage in the equal-weight three-benchmark macro-average under both continuations (Table 8, p. 26). The complete system achieves lower mean NMAE than the listed external baselines over 150 steps (Table 2, p. 7).

W1b/W2/Q1

W1b. In addition, one should not confuse "reconstruction ability" with "counterfactual correction ability". The reconstruction error which TAM optimizes, is misaligned with the counterfactual outcome prediction.

W2. No theoretical justification on why the proposed correction mechanism should really work.

Q1. How can you justify, that by minimizing the reconstruction error, the TMA can be effective in correction of the counterfactual state?

Response to W1b/W2/Q1. 1. Why reconstruction may help. Our intuition is that relationships learned from factual histories can help correct generated inputs when they remain informative under the prescribed continuation. We freeze the shared reconstruction modules and fit the private adapter to the current generated block for five updates before TSA reuses the corrected inputs.

2. Experimental support. We observe reductions in reference-window error and subsequent outcome error against references withheld from adaptation. Full reduces reference-window error at 67.3–72.7% of sampled positions across four synthetic settings (Table 9(b), p. 28). From a shared TSA-only state, we apply one correction and then disable TAM for 20 steps. Nonlinear all-zero NMAE is 0.803±0.116\mathbf{0.803}\pm0.116 without correction, 0.688±0.101\mathbf{0.688}\pm0.101 with calibrated Frozen, and 0.641±0.089\mathbf{0.641}\pm0.089 with Full (Tables 9(b) and 3(b), pp. 28 and 7; protocol: I.2, p. 27). Across 30 paired seeds, the paired NMAE difference (Full minus calibrated Frozen) is −0.047-0.047 (95% CI: [−0.073,−0.021][-0.073,-0.021]).

3. Auxiliary error analysis. We motivate LTTS through a design intuition and test its predictive value in controlled experiments. B.1 describes error propagation; it does not prove that minimizing reconstruction loss improves outcome prediction. For the synthetic window interface, under local Lipschitz and reference-window error conditions, it gives (p. 17, lines 877–900)

Eh+1≤(1+L)Eh+ϵh+1−Δh.E_{h+1}\le(1+L)E_h+\epsilon_{h+1}-\Delta_h.

Here EhE_h is window error, ϵh+1\epsilon_{h+1} is reference-window prediction error, and Δh\Delta_h is the signed reduction in window error from writeback. The bound identifies where feedback acts in the recursion.

AuthorsOpenReview ↗

Rebuttal (2/2)

W3

W3. The paper does not use standard terminology, randomly replace words with other words, introduces new words without appropriate explaining.

Response to W3. We use TSA and TAM as module names for trajectory prediction and reconstruction-based state feedback, respectively. At each step, TSA predicts the next state and outcome and records the outcome. TAM uses a frozen monitor to check complete blocks of generated windows. If the gate triggers feedback, TAM updates the private adapter and writes selected reconstructed coordinates into the input for subsequent predictions (Figure 1, p. 2; §4, pp. 4–5).

The terms mean:

  • Prototypes: memory vectors learned from observed training histories and frozen during prediction. TAM weights them by similarity to the encoded current block and uses their weighted combination for reconstruction.
  • Gate: a threshold test performed by the frozen monitor to decide whether to adapt and write back.
  • Session: one patient–continuation rollout. We retain private weights across triggers within that rollout and initialize them from offline weights for each new rollout.
  • Continuation grouping: we assign training windows to branches according to treatment at the target time, retaining the recorded treatment history within each window (§4.2–4.4, pp. 4–5).

If our responses address your concerns, we would be grateful if you would consider raising your score in light of the clarifications and additional evidence.

Thank you for your response. However, I think my concerns, e.g. W1, W2, are not resolved. And I agree with reviewer 1qat that the presenting of the paper needs substantial revision. I will therefore keep my score.

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

The paper proposes Long-term Trajectory Stabilization (LTTS), combining a temporal prediction architecture with a temporal adaptation module (TAM). TAM reconstructs generated history windows and selectively adjusts permitted state variables when a discrepancy-based trigger is activated. Experiments compare LTTS with seven baselines on two synthetic benchmarks and a MIMIC-IV-derived Cox recovery benchmark, with additional feedback controls, ablations, and computational analysis.

Critical strengths
  1. The predictor-corrector design is clearly specified, and the experiments include matched comparisons, multiple feedback controls, and 30-seed evaluations.
  2. The paper examines both predictive performance and computational overhead, including comparisons with alternative predictors and fixed-rule correction methods.
Critical weaknesses
  1. The evaluation uses synthetic outcomes and a MIMIC-IV-derived Cox recovery benchmark rather than observed clinical counterfactual outcomes. Consequently, the effectiveness of the proposed correction in real-world clinical settings remains unverified.
  2. TAM adapts using generated windows without new outcome labels. The experiments show improvements on the selected benchmarks, but it remains unclear how reliably this mechanism corrects trajectories under distribution shifts rather than moving them toward learned reconstruction patterns.
Questions for authors

See the above weakness. Besides, what evidence supports the claim that TAM improves counterfactual trajectories under distribution shift, given that its online adaptation uses generated windows without external outcome feedback?

4 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

Thank you for your helpful suggestions on clinical evaluation and the feedback mechanism. They have helped us strengthen the manuscript. Below, we explain how we use real clinical observations and present the design intuition and evidence for feedback under the tested distribution shifts.

Critical Weaknesses and Questions For Authors

W1

W1. The evaluation uses synthetic outcomes and a MIMIC-IV-derived Cox recovery benchmark rather than observed clinical counterfactual outcomes. Consequently, the effectiveness of the proposed correction in real-world clinical settings remains unverified.

Response to W1. Using real clinical records from MIMIC-IV, we obtain lower mean NMAE with Full LTTS than with TSA-only on observed mean arterial pressure (MAP) trajectories. With the same TSA checkpoint, we forecast MAP over 24 hours from a 48-hour observed prefix. In this retrospective evaluation, both methods receive the same recorded binary IMV sequence; future covariates and outcomes are withheld from inference and adaptation. We evaluate factual MAP trajectories under the recorded treatment sequence, scoring observed hourly values and leaving missing targets unscored. We include 2,864 adults selected from the 9,003 patients in our source dataset, using each patient's first eligible ICU stay of at least 72 hours in MIMIC-IV. We split patients into disjoint training/validation/test sets (70/15/15; seed 0). We use within-hour medians and normalize MAE by the population SD of observed MAP in training prefixes. Table 1 reports mean NMAE ±\pm SD over 30 paired seeds.

Table 1: Observed-outcome prediction on MIMIC-IV

ComparisonHours 1–24 NMAEHours 13–24 NMAE
TSA-only0.585±0.0700.585 \pm 0.0700.676±0.0840.676 \pm 0.084
Full LTTS0.439±0.0450.439 \pm 0.0450.491±0.0640.491 \pm 0.064

In the original paired benchmark, each patient contributes one observed history and an event or censoring time; the records do not contain complete risk curves under both prescribed continuations. We fit Cox to these survival records using training patients' observed covariates. We then freeze it to generate paired reference risk curves from the same prefix for model-recovery evaluation (§5.1.2, p. 6; D.2, pp. 19–20).

AuthorsOpenReview ↗

Rebuttal (2/2)

W2/Q1

W2. TAM adapts using generated windows without new outcome labels. The experiments show improvements on the selected benchmarks, but it remains unclear how reliably this mechanism corrects trajectories under distribution shifts rather than moving them toward learned reconstruction patterns.

Q1. See the above weakness. Besides, what evidence supports the claim that TAM improves counterfactual trajectories under distribution shift, given that its online adaptation uses generated windows without external outcome feedback?

Response to W2/Q1. 1. Design intuition. Our intuition is that historical relationships can help correct generated inputs when they remain informative under the prescribed continuation. We learn relationships among variables and time steps from observed histories. We freeze the shared reconstruction modules and fit the private adapter to the current block for five updates before TSA reuses the corrected inputs.

2. Experimental support. We test this intuition against reference trajectories withheld from adaptation and obtain lower subsequent outcome error with Full. On nonlinear all-zero, Full reduces reference-window error at 69.3% of sampled positions (Table 9(b), p. 28). From a shared state, one correction followed by 20 steps with TAM disabled yields mean NMAE of 0.803 without correction, 0.688 with calibrated Frozen, and 0.641 with Full (Tables 9(b) and 3(b), pp. 28 and 7; protocol: I.2, p. 27). Across 30 paired seeds, the paired NMAE difference (Full minus calibrated Frozen) is −0.047-0.047 (95% CI: [−0.073,−0.021][-0.073,-0.021]).

3. Tested shifts and error analysis. We obtain lower mean NMAE with Full after both treatment switches and at the largest tested input perturbation. In the nonlinear benchmark, we switch the prescribed action at t=101t=101 while retaining generated history. Table 2 reproduces mean NMAE ±\pm SD over the first 20 post-switch steps from submission Table 9(c) (p. 28).

Table 2: Action-switch comparison: submission Table 9(c)

Action switchTSA-onlyFull(5)
0→10\mathbin{\to}11.082±0.1741.082\pm0.1740.417±0.0640.417\pm0.064
1→01\mathbin{\to}01.011±0.1581.011\pm0.1580.435±0.0690.435\pm0.069

In the additional nonlinear all-zero test, a fixed-direction b=2b=2 perturbation to generated X4/X5X_4/X_5, scaled in training-SD units, gives next-20-step mean NMAE of 0.904 with Full, 1.073 with calibrated Frozen, and 1.166 without correction, after one operation with further TAM disabled.

For the synthetic window interface, B.1 provides auxiliary analysis of how writeback changes window-error propagation (p. 17, lines 877–900). It does not prove that reconstruction-based adaptation improves outcome prediction. We test the predictive value of our design intuition through the controlled comparisons above.

If our responses address your concerns, we would be grateful if you would consider raising your score in light of the clarifications and additional evidence.

Dear Reviewer eQJd,

We greatly appreciate your time and effort in reviewing our work. We hope that our responses have clarified our work and addressed your concerns. If so, we would sincerely appreciate your consideration of updating your rating accordingly. If you have any further concerns, we would be happy to provide further clarification.

We highly value every exchange during the review process and sincerely hope to receive further feedback from you at your earliest convenience.

Sincerely,

The Authors

AuthorsOpenReview ↗

Follow-up on our rebuttal

Dear Reviewer,

As the discussion deadline approaches, we would appreciate your feedback on our rebuttal and additional results. We would be happy to clarify any remaining concerns. If we have addressed your concerns, please consider updating your rating accordingly.

Thank you for your time and consideration.

Reviewer_Aq1TOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This authors have proposed a novel framework entitled LTTS aimed at two-arm counterfactual outcome prediction in longitudinal settings with time-varying treatments. The method combines a trajectory advancement component (TSA) with a selective state feedback module (TAM) that reconstructs and adjusts subsequent input windows based on generated states. The key claim of the paper is that this feedback loop reduces mean normalized absolute error (NMAE) across prediction horizons and is argued to outperform frozen reconstruction and fixed-rule baselines, and transfers across different backbone architectures. Evaluation was conducted on synthetic outcome mechanisms and a fitted-Cox recovery dataset, with ablations intended to isolate the feedback contribution.

Critical strengths

Overall this was an interesting approach from the authors. While some details might need more clarification (more on this below) there were a few key strengths as:

  • The selective state feedback mechanism is an interesting approach and, to this reviewer's best knowledge, an understudied topic in longitudinal counterfactual forecasting. The apporach to dynamically adjust input windows without requiring future observations, can be seen as a sort of "self-correction" mechanism and could align well with real-world monitoring and clinical follow-up scenarios.

  • The ablation design using matched predictors and shared starting states provides a structured way to isolate the feedback contribution, and the consistent error reductions across early, middle, and late prediction depths lends some validity to the mechanism's ability to capture beyond short-term effects

  • The authors made an effort to demostrate the architectural flexibility of the framework, maintaining performance when the TSA backbone is replaced with TimeKAN or TimeMixer, indicating that the feedback component is somewhat decoupled from the specific predictor choice.

Critical weaknesses

Coming to the possibly addressable weakness

  • First, the methods description could be improved upon. The presentation can feel highly convoluted, with entangled terminology around TSA, TAM, gated sessions, memory prototypes, and continuation-grouped inputs, severely hindering readability and making it difficult to trace how data flows through the feedback loop.
  • Coming to the feedback loop itself, this step can arguably be considered as the key contribution. It would have been interesting to observe whether such mechanisms could be added to existing SOTA methods like G-Transformer that are arguably much simpler in their construction. Going on to the baselines itself, its not evident if the gains are due to the architectural complexity or the feedback mechanisms - for example, it would have been interesting to see comparisons to Time series foundation models and/or newer variants of TimeMixer. Lastly, the theoretical claims regarding how feedback changes the input window and affects error propagation was asserted empirically without formal justification, stability analysis, or bounds on error accumulation, leaving the long-horizon behavior of the feedback loop theoretically ungrounded.
  • The authors made a commendable effort to document the computational profiling. However, this substantial computational overhead may need to be balanced against a clear utility analysis. For example, a demonstrated threshold where accuracy gains outweigh latency costs to justify the trade-off for time-sensitive clinical or equipment monitoring applications
  • On the experimental results itself, the design relies heavily on synthetic outcome mechanisms and a single fitted-Cox recovery setting. Thes may not adequately capture the confounding structures, non-stationarity, irregular sampling, or missingness patterns typical of real longitudinal data, limiting external validity. Several numerical improvements reported in the paper are presented without confidence intervals and/or standardized effect sizes
Questions for authors
  • How does the feedback mechanism avoid data leakage or violate causal assumptions, particularly regarding post-treatment confounding, collider bias, or compounding reconstruction errors, and what formal safeguards or sensitivity analyses are in place?

  • If the TAM feedback component were attached to a simpler architecture like G-Transformer, would it still yield meaningful NMAE reductions, and what does this reveal about whether the gains come from the feedback loop or the underlying backbone complexity?

  • How does the method theoretically bound or empirically mitigate error accumulation over long horizons, and what happens to the feedback loop when initial TSA predictions are substantially misaligned with the true trajectory?

  • How would the framework perform under realistic data conditions such as irregular observation intervals, missing covariates, or time-varying confounders, and can you extend the evaluation beyond synthetic and fitted-Cox settings to at least one observational longitudinal dataset?

7 comments
AuthorsOpenReview ↗

Rebuttal (1/5)

Thank you for your careful review and constructive suggestions, which have helped us improve the clarity and evaluation of our work. Below, we clarify the prediction setting and feedback procedure, and address your questions about experimental support, computational cost, and statistical uncertainty.

Critical Weaknesses and Questions For Authors

Q1

Q1. How does the feedback mechanism avoid data leakage or violate causal assumptions, particularly regarding post-treatment confounding, collider bias, or compounding reconstruction errors, and what formal safeguards or sensitivity analyses are in place?

Response to Q1. 1. Data leakage. We use post-t0t_0 reference outcomes only for scoring, never for weight initialization or online updates (p. 29, lines 1535–1539). We initialize the private adapter from offline-trained weights for each patient–arm call (§4.4, p. 5). Online updates use the current generated block as the reconstruction target (p. 5, lines 234–235).

2. Prediction target and causal concerns. We predict individual outcome trajectories under prescribed treatment continuations, rather than estimate treatment effects. Section 3 states consistency, positivity, and sequential exchangeability as conditions for identifying intervention-outcome distributions from observational histories, including for treatment-effect estimation (pp. 3–4, lines 156–163). Appendix B.1 analyzes prediction-error propagation using separate conditions on the predictor and reference trajectory.

Regarding post-treatment confounding, we advance each branch under its prescribed treatment and protect treatment and observed records. Corrected states serve as recursive inputs, not adjustment variables for a treatment-effect comparison (§4.3–4.4, p. 5). For collider bias through selection, the gate controls writeback; we score all test patients over the full 150-step suffix without selecting successful corrections (p. 20, lines 1052–1058).

We observe lower mean NMAE with Full than TSA-only at all three confounding levels, including the setting with no hidden assignment confounding (Table 1). In the nonlinear synthetic benchmark, we scaled both hidden treatment-assignment terms by γ∈0,0.5,1\gamma\in\\{0,0.5,1\\}. At γ=0\gamma=0, assignment uses only the full observed history and independent noise, giving an unconfounded assignment mechanism conditional on that history, as required by sequential exchangeability. At γ=1\gamma=1, we recover the original confounded rule. We retained the outcome equations and probability clipping. Table 1 reports mean NMAE over the full 150-step suffix; each TSA-only/Full pair uses the same predictor checkpoint at its respective γ\gamma.

Table 1

γ\gammaAll-zero: TSA-onlyAll-zero: FullAll-one: TSA-onlyAll-one: Full
00.6300.2040.5930.201
0.50.6550.2180.6370.208
10.6900.2240.6600.218

3. Error accumulation. One correction lowers nonlinear all-zero mean NMAE from 0.688 with calibrated Frozen to 0.641 with Full over the next 20 steps, with further TAM disabled (Table 3(b), p. 7). Over the full 150-step rollout, we obtain lower mean NMAE with feedback in every early, middle, and late bin across all six benchmark–arm settings, using the same TSA checkpoints (Figure 2, p. 8). For error propagation, B.1 uses local Lipschitz continuity and bounded reference-window error (p. 17), separately from Section 3's causal identification conditions. The conditional recursion in §4.4 (p. 5) shows how writeback changes the window error passed to subsequent predictions.

AuthorsOpenReview ↗

Rebuttal (2/5)

W1

W1. First, the methods description could be improved upon. The presentation can feel highly convoluted, with entangled terminology around TSA, TAM, gated sessions, memory prototypes, and continuation-grouped inputs, severely hindering readability and making it difficult to trace how data flows through the feedback loop.

Response to W1. We use TSA to predict and record outcomes, then TAM to selectively update inputs for subsequent predictions. The data flow is current window →\rightarrow prediction and recording →\rightarrow block accumulation →\rightarrow frozen monitoring →\rightarrow private update if triggered →\rightarrow masked writeback →\rightarrow next prediction (Figure 1, p. 2; §4, pp. 4–5).

Gate and session. We gate adaptation with the frozen monitor, retain private weights within each patient–continuation call, and reset them to offline weights for a new call. We check each complete generated block against the fixed, validation-calibrated threshold. A discrepancy above this threshold triggers five private updates and masked writeback (§4.3–4.4, p. 5).

Prototypes and training groups. We learn prototypes as memory vectors for reconstructing blocks of observed training histories, then freeze them for prediction. TAM gives more weight to prototypes similar to the current generated block's encoded representation and uses their weighted combination to help reconstruct that block (§4.3, p. 5; implementation, pp. 21–22). For training, we group windows by the treatment at the target time, retaining the factual treatment history within each window (§4.2, p. 4, lines 207–208; Algorithm 1, p. 22).

W2a/W2b/Q2

W2a. Coming to the feedback loop itself, this step can arguably be considered as the key contribution. It would have been interesting to observe whether such mechanisms could be added to existing SOTA methods like G-Transformer that are arguably much simpler in their construction.

W2b. Going on to the baselines itself, its not evident if the gains are due to the architectural complexity or the feedback mechanisms - for example, it would have been interesting to see comparisons to Time series foundation models and/or newer variants of TimeMixer.

Q2. If the TAM feedback component were attached to a simpler architecture like G-Transformer, would it still yield meaningful NMAE reductions, and what does this reveal about whether the gains come from the feedback loop or the underlying backbone complexity?

Response to W2a/W2b/Q2. 1. Feedback contribution. TAM lowers mean NMAE with TSA, TimeKAN, and TimeMixer when each predictor checkpoint is held fixed. With TSA, we obtain lower mean NMAE in every early, middle, and late bin across all six benchmark–arm settings (Figure 2, p. 8; results, p. 7, lines 354–358).

The first three rows of Table 2 are reproduced from submission Tables 8 and 10 (pp. 26 and 29). The TimeMixer++ row reports our additional comparison. Table 3 reproduces the feedback controls from submission Table 8. Both report mean 150-step NMAE over 30 paired seeds, equally macro-averaged across three benchmarks. Each TAM-on/off pair shares its predictor checkpoint; we train TAM separately for each backbone. For the reproduced results, SDs and validation-selection rules are provided in the submission tables and Appendix I.

Table 2: Predictor comparison

PredictorAll-zero: TAM-off →\to TAM-onAll-one: TAM-off →\to TAM-on
TSA0.655→0.2950.655\to0.2950.609→0.2550.609\to0.255
TimeKAN (2025)0.774→0.3220.774\to0.3220.721→0.4180.721\to0.418
TimeMixer (2024)0.617→0.3830.617\to0.3830.681→0.2640.681\to0.264
TimeMixer++ (2025)0.602→0.3600.602\to0.3600.653→0.2890.653\to0.289

With TSA retained, TAM also yields lower mean NMAE than validation-selected clipping and prototype-L2L_2 shrinkage. Among the completed comparisons in these two tables, TSA+TAM has the lowest mean NMAE under both continuations.

Table 3: Feedback comparison with TSA retained: submission Table 8

Prediction procedureAll-zero NMAEAll-one NMAE
TSA + clipping0.3810.394
TSA + prototype-L2L_20.4170.436
TSA + TAM (Full)0.2950.255
AuthorsOpenReview ↗

Rebuttal (3/5)

W2a/W2b/Q2 (continued)

2. G-Transformer with TAM. We also obtain lower mean NMAE with TAM in all six G-Transformer settings. Table 4 compares TAM-on/off at the same G-Transformer checkpoint under matched rollout settings and reports mean NMAE over the full 150-step suffix. The TAM-off values match the G-Transformer baseline in submission Table 2 (p. 7). This comparison extends the observed feedback benefit to G-Transformer.

Table 4: G-Transformer feedback comparison

BenchmarkAll-zero: TAM-offAll-zero: TAM-onAll-one: TAM-offAll-one: TAM-on
Synthetic linear0.4140.3880.3860.341
Synthetic nonlinear0.3180.2960.3320.291
MIMIC-derived Cox recovery0.3520.3480.3740.355

3. TimeMixer coverage. We include TimeMixer++ (2025) as the requested newer TimeMixer variant. Table 2 also includes TimeKAN (2025) and TimeMixer (2024).

W2c/Q3

W2c. Lastly, the theoretical claims regarding how feedback changes the input window and affects error propagation was asserted empirically without formal justification, stability analysis, or bounds on error accumulation, leaving the long-horizon behavior of the feedback loop theoretically ungrounded.

Q3. How does the method theoretically bound or empirically mitigate error accumulation over long horizons, and what happens to the feedback loop when initial TSA predictions are substantially misaligned with the true trajectory?

Response to W2c/Q3. 1. Role of the theory. Our theory characterizes error propagation; we test the benefit of the learned correction empirically. For the synthetic window-based interface, Appendix B.1 (p. 17, lines 877–900) gives Eh+1≤(1+L)Eh+ϵh+1−ΔhE_{h+1}\le(1+L)E_h+\epsilon_{h+1}-\Delta_h, under local Lipschitz continuity and bounded prediction error on reference windows. Here EhE_h is window error, ϵh+1\epsilon_{h+1} is reference-window prediction error, and Δh\Delta_h is the signed reduction from writeback. This analysis identifies where feedback acts; it does not establish that the learned update makes Δh\Delta_h positive.

2. Insight and experimental evidence. Our controlled experiments show that this feedback improves subsequent outcome prediction in the tested settings. Our insight is to use relationships learned from observed histories to adjust generated states before TSA reuses them. Relationships among variables and time steps may help correct these inputs when they remain informative under the prescribed continuation. We keep the shared reconstruction modules fixed and use five private updates to tailor the response to the current generated block.

On nonlinear all-zero, Full reduces reference-window error at 69.3% of sampled positions (Table 9(b), p. 28). One correction followed by 20 steps with TAM disabled yields NMAE of 0.803 without correction, 0.688 with calibrated Frozen, and 0.641 with Full (Tables 9(b) and 3(b), pp. 28 and 7). Over 150 steps, the same-checkpoint comparison gives lower mean NMAE with TAM in every early, middle, and late bin across all six settings (Figure 2, p. 8).

3. Substantial initial misalignment. Full has lower mean NMAE than both controls at every tested perturbation level, including b=2b=2. At b=2b=2, NMAE is 0.904 with Full, versus 1.166 without correction and 1.073 with calibrated Frozen (Table 5).

For nonlinear all-zero, we add a fixed-direction perturbation to X4/X5X_4/X_5 in the last generated row; bb scales it in training-standard-deviation units. All branches share the perturbed state and block. Frozen retains its validation-selected coefficient; Full starts from offline private weights and takes five updates. We then disable TAM in every branch and report mean NMAE over the next 20 steps. The b=0b=0 row uses the manuscript's single-correction results with no added perturbation (I.2, p. 27, lines 1437–1444).

Table 5: Starting-state perturbation, nonlinear all-zero

Perturbation bbNo correctionCalibrated FrozenFull
0, manuscript0.8030.6880.641
0.50.8330.7170.668
10.9150.8160.793
21.1661.0730.904

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Contributions and Rebuttal Evidence

Dear Program Chairs, Senior Area Chair, Area Chair, and Reviewers,

Thank you for your time and feedback. We summarize our contribution and the evidence provided during rebuttal.

  1. Selective feedback improves recursive prediction. With LTTS, we use relationships learned from observed histories to adjust generated inputs before reuse, without future measurements or labels. We preserve prescribed treatments and reported outcomes. Our added G-Transformer comparison holds predictor checkpoints fixed and lowers NMAE in all six benchmark–treatment settings (Aq1T, Table 4).

  2. Correction improves subsequent outcome prediction. We evaluate against reference trajectories withheld from adaptation. After one correction with further feedback disabled, nonlinear all-zero NMAE over the next 20 steps is 0.803 without correction, 0.688 with calibrated Frozen, and 0.641 with Full, averaged over 30 paired seeds (submission, Table 3(b); x2Bf, W1b/W2/Q1). Our conditional error analysis explains how feedback enters error propagation (Appendix B.1). We also retain gains under treatment switches and starting-state perturbations (eQJd, Table 2; Aq1T, Table 5).

  3. Added validation on observed clinical outcomes. We added factual mean arterial pressure (MAP) forecasting under recorded treatment sequences using 2,864 MIMIC-IV adults, with patient-disjoint training, validation, and test sets. From a 48-hour observed history, we forecast 24-hour MAP trajectories while withholding future covariates and outcomes. With the same predictor checkpoint, Full LTTS lowers NMAE relative to TSA-only over both the full forecast and its later half (Aq1T, Table 6; eQJd, Table 1).

We appreciate your consideration of these findings alongside our detailed responses.

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.