acceptodds
Under review as a conference paper at ICLR 2027

Learning Composable Forecasting Operators via Block Infilling

Abstract

Long-horizon forecasting is usually formulated either as a fixed mapping to a target window or as a recursive process that advances through its own predictions. We ask whether forecasting can instead be organized around a reusable local prediction problem. We introduce Augure, which learns a shared masked block-infilling operator and composes it across future chunks. The same operator is used at every chunk, predicts each chunk jointly, and can be applied additional times at inference without changing its parameters or output head. Across standard probabilistic forecasting benchmarks, this inductive bias is competitive with direct, recurrent, autoregressive, and diffusion-based alternatives. Controlled ablations show that the local training problem matters: masked infilling, exposure to varying prefix lengths, and the absence of explicit rollout-depth specialization each affect the quality of the resulting operator. A checkpoint trained at can be reused up to , with little native-relative cost on some datasets and larger degradation on others. These results suggest that masked block infilling provides a useful compositional inductive bias while also exposing a central limitation of repeated reuse:performance can degrade as model-generated prefixes accumulate.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Forecasting, Masked Modeling, Compositional Generalization, Inductive Bias

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_iHKrOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Augure, a probabilistic long-horizon forecaster built around a single reusable, chunk-level prediction operator, as an alternative to direct multi-step or step-wise recursive forecasting. The future horizon is split into chunks of B patches of P timesteps each (default B=6, P=8). A causal transformer encoder encodes the lookback window and any previously committed chunks, maintained as a KV cache. A decoder predicts the current chunk with bidirectional attention inside the chunk, conditioned on that prefix. A residual MLP head outputs 9 quantiles (0.1–0.9) per timestep. The model is channel-independent and uses non-affine RevIN, RoPE, and roughly 2.5M parameters.

Training is framed as masked block infilling. At each optimization step, one target chunk index is sampled uniformly and shared across the minibatch. All earlier chunks are given as ground-truth prefix. Each patch of the target chunk is masked (zeroed) independently with probability m=0.5, and the pinball loss is computed only on masked positions. No free-running rollout is used during training.

Inference applies the same operator repeatedly. Each new chunk is fully masked and predicted in one decoder pass. Its median is then encoded and appended to the prefix cache, and the process repeats until the requested horizon is covered. Forecast length is therefore set by the number of operator applications rather than by a horizon-specific output head.

Experiments cover seven standard benchmarks (ETTh1/h2/m1/m2, Electricity, Exchange Rate, Weather) with L=336 and H ∈ {96, 192, 336, 720}. MAE, a quantile-based CRPS approximation, and WQL are reported in original data units. Baselines are DLinear, PatchTST, DeepAR, SegRNN, and NsDiff. The paper claims:

  1. Competitive native-horizon accuracy. Augure is best on all metrics and horizons for Weather and Electricity. SegRNN leads most ETT settings, while Augure obtains the lowest MAE on ETTh1/ETTh2 at H=720.
  2. Reuse beyond the training horizon. A checkpoint trained at H=96 (two chunk positions) can run unchanged up to H=720 (fifteen applications). Relative to native training, the cost is negligible on ETTh2 (−1.1% MAE) and small on ETTm2 (+4.8%), but substantial on ETTm1 (+21.8%) and Weather (+17.6%).
  3. Mechanistic ablations. Three controls are compared with the full model:
  • removing infilling (Direct block);
  • restricting training to a single chunk index (Fixed chunk);
  • adding explicit rollout-depth conditioning (Scalar depth).

The full model is the strongest when each variant is trained natively for the evaluation horizon, while some ablations degrade more slowly under deep reuse. The authors interpret this as a fidelity–stability trade-off: the infilling objective learns a stronger local rule that may be more sensitive to errors accumulating in a model-generated prefix.

The authors acknowledge the following limitations:

  • the train/inference mismatch: training uses ground-truth prefixes and partial masks, while inference uses generated prefixes and full masks;
  • forecast uncertainty is not propagated, since only the median is committed to the prefix;
  • cross-variate dependencies are not modeled;
  • reuse is evaluated only up to 7.5× the training horizon.
Critical strengths
  1. Clear, well-motivated formulation. The paper presents masked block infilling as the forecasting operator itself, not as an auxiliary pretraining objective, and composes it causally across chunks. Its core argument is simple and plausible: since no output position is permanently tied to a fixed forecast offset, the operator should become reusable across chunk positions. The method is described precisely, with explicit equations, algorithms for training and inference, and a complexity analysis of the cached rollout.

  2. Horizon-flexible inference from a single checkpoint. Forecast length is set by how many times the operator is applied, not by a horizon-specific output head. A model trained at H=96 can therefore run unchanged up to H=720. On some datasets this reuse costs very little relative to native training (ETTh2 −1.1% MAE, ETTm2 +4.8%), which isnpractically useful when the target horizon is unknown at training time.

  3. Informative Direct-block control. The comparison against a variant without infilling, with chunk geometry and model scale held fixed, is the right control for isolating the objective. Under native training, the full model beats Direct block at every horizon on all four ablated datasets. This is the most convincing evidence that the infilling objective contributes something beyond blockwise generation.

  4. Strong results on some benchmarks. Augure obtains the best MAE, CRPS, and WQL at every horizon on Weather and Electricity, and the lowest MAE at H=720 on ETTh1 and ETTh2. It does so with a small model (~2.5M parameters).

  5. Honest framing and explicit limitations. The authors claim competitiveness rather than dominance and explicitly reject a claim of horizon invariance. They report datasets where reuse degrades substantially (ETTm1 +21.8%, Weather +17.6%). They also openly discuss the train/inference mismatch (ground-truth prefixes and partial masks in training versus generated prefixes and full masks at inference), the lack of uncertainty propagation, and channel independence.

  6. Rigorous model selection. The chunk size and masking probability are selected on validation, and the test-set sweep is reported only as a post-selection check. The metrics and normalization schemes are documented precisely (Table 6, Appendix A.5), and the authors state that code, configurations, and processed data are provided.

  7. Useful conceptual distinction. The paper separates applicability (an operator can be executed at greater composition depth) from compositional stability (whether the evolving prefix remains useful). This is a worthwhileframing for horizon-flexible forecasters, even though the evidence for the specific trade-off is not yet convincing.

Critical weaknesses
  1. The reuse tables cannot be reproduced from the paper's own native results, and these tables carry the core mechanistic claims. The percentages in Tables 10 and 11 should equal (reused − native) / native, using the native values in Tables 2 and 9. For the ablation variants they do not:
    • Weather, Scalar depth, H=720: reported +132.7%, but 28.934 vs. native 34.540 gives −16.2%.
    • ETTm1, Scalar depth, H=720: reported +35.8%, recomputed −1.4%. At H=336: reported +14.7%, recomputed −22.2%.
    • ETTm1, Fixed chunk, H=192: reported −6.5%, recomputed −21.6%.
    • ETTm1, Direct block, H=720: the reported +12.4% implies a native MAE of 2.082, which is Augure's native value; Direct block's own native value is 2.134 (Table 2).

The H=96 columns are also inconsistent. Evaluating an H=96 checkpoint at H=96 should reproduce the native H=96 result, yet the numbers differ: - Scalar depth: 2.424 vs. 1.484 on ETTm1, and 18.145 vs. 10.138 on Weather. - Fixed chunk: 1.669 vs. 1.479 on ETTm1, and 12.250 vs. 10.334 on Weather.

Augure's own numbers are consistent across tables (Table 8 matches Table 1), so the problem is specific to the ablations. The "fidelity–stability trade-off" claimed in Sec. 5.3, Sec. 5.4, and the conclusion, including the claim that Scalar depth "becomes markedly unstable on Weather," rests on these numbers. As reported, the claims are unsupported. The authors should explain which checkpoints produced each table and recompute all percentages.

  1. No variance estimates. All results come from single runs, with no seeds, standard deviations, or significance tests. The H=96 discrepancies above suggest that results for an identical configuration can differ by up to ~80% between training runs. Several claims rest on margins that are likely within this noise:

    • ETTh2 reuse being "essentially cost-free" (−1.1% MAE / −0.3% CRPS);
    • variant rankings separated by 1–3%;
    • many of the comparisons against SegRNN on ETT.
  2. Weak and partly broken baselines.

    • DeepAR is effectively non-functional. Its WQL is ~0.76–0.79 on every dataset and horizon. The paper attributes this to MeanScaler but does not fix it, so DeepAR provides no real point of comparison.
    • PatchTST behaves erratically across horizons, which suggests training failures. On ETTm2 its MAE is 5.978 at H=96, 2.501 at H=192, and 6.275 at H=336. On ETTh1 it is 2.686 at H=96 and 2.075 at H=192.
    • Some baseline CRPS values exceed MAE (DLinear on Weather; NsDiff on Electricity at every horizon). This means their predicted quantile spread makes them worse than the median alone, i.e., they are poorly calibrated.
    • The tuning budget is asymmetric. No hyperparameter search is reported for the baselines, while Augure's B and m are validation-tuned.
    • Normalization differs across models (Table 6). Augure and SegRNN use per-window RevIN, NsDiff fits train-split statistics, and DeepAR applies no centering. This confounds the comparison.
  3. Key competitors and controls are missing. The paper's central selling point is horizon-flexible composition, but it compares against none of the cited models that already offer this: TimesFM, Timer, TiRex, and Toto, which are patch-based models that generate to arbitrary horizons. It also omits standard long-horizon baselines (iTransformer, N-HiTS) and probabilistic ones (CSDI, TimeGrad, Chronos-Bolt). The most informative same-backbone controls are also absent:

    • K=1 (B·P ≥ H), i.e., the same backbone used as a single-pass masked forecaster. This control exists only implicitly, for Weather at H=96 (B=12 in Tables 7 and 12). Extending it to longer horizons and other datasets would test the claim that B·P < H is necessary for learning a compositional operator. It also changes chunk size and removes chunk-position sampling, so it measures composition together with chunk granularity rather than composition alone.
    • B=1: next-patch autoregression with the same backbone.
  4. Limited novelty, and the closest prior work is not discussed. The architecture combines known components: patching, RevIN, channel independence, a causal encoder with KV cache, block-bidirectional/inter-block-causal attention, RoPE, and a quantile head. The training and inference scheme is very close to semi-autoregressive masked generation and block diffusion (e.g., Arriola et al., 2025):

    • masked training within a block, conditioned on a clean prefix;
    • bidirectional attention within the block and causal attention across blocks;
    • block-by-block generation with a KV cache.

Augure is effectively a single-step, continuous-valued instance of this scheme. Feeding zero placeholders for the future into an encoder-decoder also echoes Informer's generative-style decoder (Zhou et al., 2021). Neither line of work is cited. The novelty should be positioned against them.

  1. The metric protocol prevents comparison with prior work.

    • Raw units, pooled across variates. MAE and CRPS are computed in original units and pooled across variates, so high-magnitude channels dominate (e.g., the 321 Electricity clients; the heterogeneous Weather variables). The datasets where Augure leads most clearly (Weather, Electricity) are also the ones most affected by this pooling.
    • No normalized MSE/MAE. The standard protocol from PatchTST onward is not reported, so results cannot be checked against published numbers.
    • The WQL definition has a flaw. Its denominator is max(mean|y|, 1), so on Exchange Rate WQL is identical to CRPS and adds no information.
    • Limited probabilistic evaluation. CRPS is approximated from only nine quantiles (0.1–0.9), which ignores the tails. No coverage or calibration metrics are reported.
  2. The reuse analysis does not test the proposed mechanism.

    • Reused Augure is compared only against native Augure. It is never compared with baselines, or with other models that can generate to arbitrary horizons, at the extended horizons.
    • There is no per-chunk error curve and no oracle-prefix experiment (ground-truth prefixes at inference). These are needed to attribute degradation to error accumulation in the generated prefix rather than to horizon difficulty.
    • Reuse is tested only from H₀=96.
  3. The objective is not explained, and ablations are missing.

    • Why infilling helps is unclear. Inference always uses the fully masked case, which is only 1/64 of training examples, and training additionally exposes ground-truth target patches that never exist at test time. The paper does not explain why this improves forecasting, e.g., regularization, an auxiliary-task effect, or a change in effective loss weighting.
    • Several design choices are not ablated:
      • a learned mask token instead of zeros;
      • committing a sampled path instead of the median;
      • a decoder-only model with a block-causal mask instead of encoder-decoder;
      • a training mix weighted toward the fully masked case.
    • The Scalar depth control is not a fair test of depth conditioning. Under reuse, the H=96 checkpoint is trained only on r ∈ {0.5, 1}, yet it is evaluated with r up to 7.5 because r = i/K_train is left unclamped. The degradation therefore mainly reflects inputs far outside the training range. A clamped or learned depth embedding would test whether depth information itself is harmful. The variant also degrades sharply at native horizons, where no extrapolation occurs (ETTm1 H=96: 2.424 vs. 1.514 MAE; Weather: 18.145 vs. 9.816), even though the same configuration scores 1.484 and 10.138 in Table 10.
    • Several configuration details are missing:
      • the value of i_fixed in the Fixed chunk control is not given;
      • the B×m sweep is run only on Weather, but its result is applied to all datasets;
      • ablations exclude ETTh1, Electricity, and Exchange Rate.
  4. Minor issues.

    • Exchange Rate is listed as "hourly" in Table 5; it is daily.
    • Traffic and ILI, part of the standard long-horizon benchmark suite, are omitted, and so are broader benchmarks such as GIFT-Eval.
    • No runtime, latency, or baseline parameter counts are reported.
    • There are small rounding mismatches between tables (e.g., ETTm1 H=192 reuse is +1.9% in Table 10 vs. +2.0% in Table 8).
Questions for authors
  1. Which checkpoints were used for Tables 2/9 and for Tables 10/11? For the ablation variants, the H=96 values differ between these tables (e.g., Scalar depth on Weather: 18.145 vs. 10.138 MAE). Many reported reuse percentages also cannot be reproduced from the native values in Tables 2/9. For example, Weather / Scalar depth / H=720 is reported as +132.7%, but 28.934 vs. 34.540 gives −16.2%; the ETTm1 / Direct block / H=720 figure of +12.4% appears to use Augure's native MAE (2.082) as its reference. Could you provide corrected tables and state which native reference each percentage uses?

  2. If the discrepancies above come from separate training runs of the same configuration, how large is the variance across seeds? Could you report mean ± std over at least 3 seeds for Table 1, the mechanistic ablations, and the reuse study? Do the ETTh2 "cost-free reuse" result and the fidelity–stability trade-off survive this?

  3. Why does Scalar depth degrade so strongly at H=96 (ETTm1: 2.424 vs. 1.514 MAE; Weather: 18.145 vs. 9.816), where K=2 and the depth feature reduces to a learned per-chunk offset? Was its training stable? For reuse, how does a clamped (r = min(i/K_train, 1)) or learned depth embedding behave compared with the unclamped linear extrapolation?

  4. Inference always uses a fully masked chunk, which occurs in only ~1/64 of training examples, and training exposes ground-truth target patches that are never available at test time. What explains the gain over Direct block: regularization, an auxiliary-task effect, a different effective loss weighting, or something else? Have you tried a mixture that oversamples the fully masked case, or a curriculum toward full masking?

  5. Section 3.4 states that B·P < H "is what forces the model to learn compositional KV-cache passing." Could you compare against K=1 (one chunk spanning the full horizon, with matched supervision) at H ∈ {192, 336, 720} and on datasets beyond Weather? Could you also include a B=1 next-patch autoregressive variant on the same backbone?

  6. The central claim is that one operator can be reused across horizons. How does Augure compare with cited patch-based models that also do this (TimesFM, Timer, TiRex, Toto), either fine-tuned in-domain or zero-shot, especially at the reused horizons?

  7. Baselines:

    • DeepAR's WQL is ~0.77 on every dataset, which is near-trivial. Why was its scaling not fixed, e.g., by using StdScaler or RevIN as for the other models?
    • PatchTST's MAE is non-monotonic across horizons (ETTm2: 5.978 / 2.501 / 6.275 / 6.401). Were these runs checked for training failures?
    • What hyperparameter tuning, if any, was done for the baselines?
    • How do the results change if all models use the same normalization?
  8. Could you report the standard normalized MSE/MAE protocol so results can be compared with published numbers? How do the rankings change if raw-unit metrics are computed per variate and then averaged, rather than pooled, so that high-magnitude channels in Electricity and Weather do not dominate?

  9. Can you show per-chunk error curves for native vs. reused checkpoints, and a ground-truth-prefix oracle run at inference? These would separate generated-prefix distribution shift from the intrinsic difficulty of longer horizons, and would directly test the explanation given in Sec. 5.4.

  10. How does Augure differ from block-diffusion language models (e.g., Arriola et al., 2025), which also train with masked blocks conditioned on clean prefixes, use bidirectional attention within blocks and causal attention across them, and generate block by block with a KV cache? Augure looks like a one-step, continuous-valued instance of that scheme. Relatedly, how does it differ from Informer's generative-style decoder applied chunk-wise?

  11. Since only the median is fed back into the prefix, how well calibrated are the quantiles at later chunks? Could you report coverage or PIT diagnostics as a function of chunk index, and compare committing the median with committing sampled paths?

2 comments
AuthorsOpenReview ↗

Response to Reviewer iHKr

We are deeply grateful to Reviewer iHKr for an extraordinarily thorough technical review. Your audit identified critical reporting errors and missing controls that we have now resolved in the revision. Below, we address every weakness and question systematically.

A. Table Discrepancies & 3-Seed Variance (Weaknesses 1-2; Questions 1-2)

  • Root Cause: In our original submission, an evaluation script bug pulled single-seed preliminary checkpoints for Tables 10/11 instead of the final checkpoints reported in Tables 2/9.
  • Corrected Numbers & 3-Seed Variance: We audited all scripts and re-ran 3-seed evaluations (mean ± std) using unified checkpoints. Below are the corrected reuse comparisons for the exact rows flagged by the reviewer:
DatasetVariantH=96H=96 (Native)H=192H=192 (Reused/Δ\Delta)H=336H=336 (Reused/Δ\Delta)H=720H=720 (Reused/Δ\Delta)
ETTh2Augure2.437 ± 0.0442.663 ± 0.036 / +3.4%2.791 ± 0.048 / +1.5%2.965 ± 0.126 / -2.6%
ETTm1Direct block1.590 ± 0.0371.741 ± 0.015 / +0.1%1.902 ± 0.045 / +1.7%2.356 ± 0.154 / +15.7%
ETTm1Fixed chunk1.515 ± 0.1331.666 ± 0.021 / -9.7%1.918 ± 0.028 / -2.7%2.530 ± 0.155 / +0.0%
ETTm1Scalar depth2.724 ± 0.2701.684 ± 0.050 / -40.1%2.011 ± 0.125 / -27.2%2.849 ± 0.293 / +0.1%
WeatherScalar depth17.629 ± 0.46411.047 ± 0.448 / -52.0%13.174 ± 2.252 / -40.0%20.195 ± 7.571 / -36.8%

- Impact on Claims: For ETTh2 at H=720H=720, paired per-seed generated-prefix-minus-native MAE differences are -1.1%, +2.4%, and -9.0%; their mean is -2.6% with sample SD 5.9 percentage points. Corresponding CRPS changes are -0.3%, +1.4%, and -11.8%, mean -3.6% with SD 7.2 percentage points. The variation spans improvement and penalty, supporting near-zero average reuse cost, not a consistent per-seed gain.

B. Why Infilling Helps & Objective Ablations (Weakness 9; Question 4)

  • Why m=0.5m=0.5 beats Direct Prediction despite m=1.0m=1.0 inference: Crucially, our Direct Block ablation is mathematically equivalent to training exclusively at the fully masked endpoint (m=1.0m=1.0), where the decoder sees zero target patches and is supervised on all positions. The consistent improvement of Augure (m=0.5) over Direct Block (m=1.0) both natively and under our new Oracle-Prefix test demonstrates that exposing visible intra-chunk patches during training acts as a structural regularizer: it forces the bidirectional decoder to capture fine-grained local temporal dependencies rather than memorizing a rigid prefix-to-block mapping.
  • As shown in our masking-ratio sweep (Tables 7/12), m=0.5m=0.5 outperforms both lighter masking (m=0.3m=0.3, which over-relies on intra-chunk shortcuts) and full masking (Direct Block, m=1.0m=1.0).

C. External Baselines & Foundation Models (Weakness 4; Question 6)

  • Foundation Models: Augure is a 2.5M-parameter in-domain model trained from scratch on a single dataset. We therefore focus on controlled in-domain comparisons and do not present Augure as a foundation model.
  • Diffusion & Long-Horizon Baselines: We include NsDiff (ICML 2025), a more recent score-based diffusion forecaster than CSDI/TimeGrad, alongside PatchTST and SegRNN. As noted to Reviewer Y6f8, iTransformer suffered NaN gradient explosions under quantile-head training.

D. Fixing Baseline Parity: Quantile Crossing, Normalization, and DeepAR (Weakness 3; Question 7)

  • CRPS-MAE consistency: Reviewer iHKr noted that CRPS exceeded MAE for DLinear on Weather and NsDiff on Electricity. After quantile sorting, DLinear's Weather CRPS is below MAE at all horizons (three-seed means): 15.250 vs. 15.263 at H=96H=96, 15.443 vs. 16.067 at H=192H=192, 15.772 vs. 16.569 at H=336H=336, and 16.496 vs. 16.676 at H=720H=720. For NsDiff too, its CRPS on Electricity no longer exceeds the MAE.
  • DeepAR (StdScaler Re-evaluation): The two added DeepAR seeds use StdScaler; the original paper seed used MeanScaler, so this is not a three-seed same-scaler comparison.
  • Tuning Disclosure: We explicitly disclose in Section 4.2 that BB and mm were validation-tuned for Augure while baselines used parameter-matched reference configurations.

E. WQL Fix

We removed the max⁡(⋅,1)\max(\cdot, 1) floor in the WQL denominator (now strictly CRPS/mean(∣y∣)\text{CRPS} / \text{mean}(|y|)) so Exchange Rate WQL accurately reflects the error rate.

AuthorsOpenReview ↗

Response to Reviewer iHKr (continued)

F. Reuse Mechanism: Oracle-Prefix & Per-Chunk Error Curves (Weakness 8; Question 9)

As requested, we ran an oracle-prefix vs. generated-prefix evaluation across chunk indices k=1,…,15k=1,\dots,15 (H=96→720H=96 \to 720) on ETTh2 and Weather (Figure 2):

  • Per-Chunk Errors (Chunk 1→5→10→151 \to 5 \to 10 \to 15): ETTh2 Augure oracle-prefix MAE is 2.307 at chunk 1, 2.262 at chunk 5, 2.271 at chunk 10, and 2.282 at chunk 15; generated-prefix MAE is 2.307 at chunk 1 and 3.183 at chunk 15. Weather Augure oracle-prefix MAE is 8.722 at chunk 1 and 10.370 at chunk 15; generated-prefix MAE is 8.722 at chunk 1 and 16.831 at chunk 15. Values are three-seed means.
  • These curves separate horizon difficulty under true prefixes from the additional change under generated prefixes; they do not imply that generated chunks add information about the future.

G. Positioning vs. Block Diffusion LMs (Arriola et al., 2025) & Informer (Zhou et al., 2021)

We thank the reviewer for highlighting these connections, which we now cite and contrast in Section 2:

  • vs. Block Diffusion LMs (Arriola et al., 2025): While BD3-LMs share block-causal attention and KV caching across blocks, they operate on discrete tokens using multi-step iterative denoising within each block. Augure operates directly on continuous patches and predicts all quantiles of a chunk in a single forward pass (O(1)\mathcal{O}(1) per chunk) without noise schedules or iterative refinement, followed by value-space median re-encoding.
  • vs. Informer (Zhou et al., 2021): Informer feeds zero-padded placeholders into a non-autoregressive decoder for the entire fixed horizon in one shot, which cannot extrapolate beyond training length HH. Augure decodes sequentially chunk-by-chunk and extends a causal KV cache, enabling variable-horizon composition.

H. Minor Corrections

  • Corrected Exchange Rate frequency to daily (Table 5).
  • Added runtime/parameter counts for all baselines (Table 5), and verified all cross-table rounding.
Reviewer_Y6f8OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Augure, which treats masked block filling directly as the prediction operator itself, and assembles arbitrarily long time domains block by block using a single shared decoder.

Critical strengths
  1. This paper proposes that mask filling can be directly used as a composable predictive operator in inference, rather than merely as a pre-training target. This sparked my interest in the paper.
  2. This paper learns a general block-filling operator and repeatedly applies it during inference to generate the future block by block, with the total prediction length determined by the number of operator invocations. This leads me to believe that this method may have certain value in the future.
Critical weaknesses
  1. There is an internal data inconsistency between the original ablation tables and the reused ablation tables (comparing Tables 2 and 9 with Tables 10 and 11). This undermines my confidence in the experimental results.
  2. TiRex, Toto 2.0, and SegRNN were repeatedly discussed in related work, but none of them were compared in the main table. Even the iTransformer mentioned in the main text is not in the baseline. This makes me doubt the conclusions of the paper.
Questions for authors
  1. Could you explain the motivation behind your method? Why perform variable-length prediction?
  2. During training, Augure samples only one target chunk per optimization step and masks half of its patches, while running for a fixed maximum of 50 epochs regardless of H. Can you explain this?
  3. On line 35 of the paper, you stated "the same local block-inflling problem is solved throughout the future." Could you explain why this is the case?
1 comment
AuthorsOpenReview ↗

Response to Reviewer Y6f8

We thank Reviewer Y6f8 for the thoughtful review and for noting the promise of using masked infilling as a composable inference operator.

1. Internal Table Consistency

As detailed in our General Response, the inconsistency between Tables 2/9 and Tables 10/11 was caused by a checkpoint-aggregation mismatch in our table-generation script. We have re-verified every checkpoint, recomputed all tables over 3 seeds, and confirmed 100% alignment between H=96H=96 native and reuse columns.

2. Choice of Baselines (SegRNN, TiRex, Toto 2.0, iTransformer)

- SegRNN: SegRNN was included in Table 1 across all 7 datasets and horizons (where it performs very strongly on ETT, while Augure leads on Weather and Electricity and achieves lower MAE on ETTh1/ETTh2 at H=720H=720).

- TiRex and Toto 2.0: These are large-scale foundation models pretrained on massive cross-domain corpora. Because Augure is trained strictly from scratch on a single dataset (2,221,200 parameters) to isolate an inductive bias rather than pretraining scale, we focus our primary controlled benchmark on in-domain models trained under identical splits.

- iTransformer: We attempted to train iTransformer with a 9-quantile pinball head under our unified probabilistic pipeline, but repeatedly encountered non-finite (NaN) gradient explosions despite gradient clipping and reducing the learning rate to 3×10−53\times10^{-5}. No valid checkpoint was obtained, so we do not report iTransformer metrics.

3. Motivation for Variable-Length Prediction

In operational settings (e.g., cloud telemetry or energy grids), required forecast horizons often vary at deployment time. Fixed-horizon direct forecasters (PatchTST, DLinear) bake HH into their projection heads and require training separate models for every HH. Composing a single shared local operator allows one trained checkpoint to generate forecasts for arbitrary horizons without architectural changes or retraining.

4. Training Mechanics (Single-Chunk Sampling, m=0.5m=0.5, Max 50 Epochs)

- Single-chunk sampling: Sampling one target chunk i∼U1,…,Ki \sim \mathcal{U}\\{1,\dots,K\\} per step trains the shared operator across varying prefix lengths while keeping memory and step compute bounded to one chunk rather than rolling out the full horizon during backpropagation.

- Masking ratio m=0.5m=0.5: Selected via validation sweep (Table 7) as the best balance between local structural conditioning and extrapolation difficulty.

- 50-epoch cap: 50 is the maximum epoch count. Models typically converge in the range of 15 to 35 epochs.

5. Clarification of Line 35 ("Same local block-infilling problem is solved throughout the future")

At every inference step kk, the exact same decoder parameters and chunk geometry (B×P×QB \times P \times Q) solve the prediction task given a fully masked chunk and the current causal prefix; only the conditioning prefix changes. We have reworded Line 35 in Section 1 to clarify that the local problem formulation is invariant across steps, not the difficulty or error rate.

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

The paper introduces Augure, a forecasting model trained through masked block infilling and applied recursively across future chunks. It evaluates both native-horizon forecasting and the reuse of short-horizon checkpoints for longer predictions.

Critical strengths
  • Interesting research question. Investigating whether a locally trained prediction rule can be reused effectively beyond its training horizon is meaningful. Distinguishing native forecasting accuracy from stability under repeated composition provides a useful perspective on variable-horizon forecasting.
Critical weaknesses
  • Unclear methodological and practical advantage. The framework builds on established blockwise autoregressive prediction and masked modeling. Although infilling improves some native-horizon results, the existing Direct Block control remains competitive and sometimes performs better under deep reuse. The experiments therefore do not clearly establish the distinctive benefit of the proposed training objective or a favorable accuracy–cost trade-off.
  • Experimental consistency and uncertainty require clarification. Some same-horizon ablation results differ substantially between tables: for example, the ETTm1 Scalar Depth variant has an H=96H=96 MAE of 2.424 in Table 2 but 1.484 in Table 10. It is unclear whether different checkpoints or settings were used. Together with the lack of clearly reported seed-level variability, this complicates interpretation of the reuse comparisons.
Questions for authors

Could the authors provide more solid experimental results to support the central claims?

1 comment
AuthorsOpenReview ↗

Response to Reviewer NgPN

We thank Reviewer NgPN for recognizing the value of distinguishing native forecasting accuracy from stability under repeated composition.

1. Methodological Advantage Over Direct Block & Accuracy-Cost Trade-off

- Oracle-Prefix Isolation: To decouple the local forecasting rule from recursive error accumulation, we added oracle-prefix and free-running per-chunk evaluations comparing Augure against the matched-backbone Direct Block baseline on ETTh2 and Weather (Figure 2 / Table 3). With ground-truth prefixes, Augure reduces global MAE by 10.0% on ETTh2 (2.284 vs. 2.539) and 14.2% on Weather (9.467 vs. 11.030), using three-seed means. These are raw-unit MAE values over the full H=720H=720 forecast, not individual chunk values.

- Free-Running Reuse & Trade-off: Under free-running rollout, the gap between oracle and generated prefixes quantifies how error compounds through the conditioning cache. While Direct Block occasionally degrades more slowly at H=720H=720 on ETTm1/Weather, Augure maintains superior native accuracy across horizons and superior reuse on ETTh2/ETTm2. We now also report inference latency and parameter counts for all baselines in the appendix; the ETTh1 CPU and GPU measurements are summarized in Table 5.

2. Experimental Consistency and Seed-Level Variability

- Table Discrepancy Fix: The discrepancy on ETTm1 Scalar Depth at H=96H=96 (2.424 in Table 2 vs. 1.484 in Table 10) resulted from an aggregation script error that accidentally mixed preliminary screening checkpoints with final checkpoints in Tables 10/11. All tables now evaluate the exact same unified checkpoints.

- 3-Seed Uncertainty: We report mean ± std where three valid seeds are available; incomplete seed coverage is shown with a dash (ETTh1 H=96H=96 Augure MAE: 1.799 ± 0.031; ETTm1 H=96H=96 Augure MAE: 1.484 ± 0.033).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.