pith. sign in

arxiv: 2606.12990 · v1 · pith:XPLLRJAJnew · submitted 2026-06-11 · 💻 cs.LG

Exposure Bias as Epistemic Underidentification in Recursive Forecasting

Riku Green , Zahraa S. Abdallah , Telmo M Silva Filho This is my paper

Pith reviewed 2026-06-27 07:47 UTC · model grok-4.3

classification 💻 cs.LG
keywords exposure biasrecursive forecastingepistemic underidentificationinduced statesprovenance variablesdistribution shiftmulti-step prediction
0
0 comments X

The pith

Recursive forecasting under partial observability is an epistemic underidentification problem, not only a distribution shift.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that exposure bias in recursive multi-step forecasting cannot be fully explained as distribution shift between training and deployment. It proves that, even when latent dynamics are deterministic, one-step supervision on observed data identifies the correct predictor behavior only for those observed contexts. Once the model rolls out and generates its own states, those self-induced states can have local targets that are not fixed by the numeric state values alone, leaving the deployed recursive predictor underidentified. The authors formalize this using induced states and provenance variables, derive an error decomposition, and show empirically that rollout occupies a distinct regime where provenance information can aid correction.

Core claim

Under partial observability or state truncation, recursive rollout is an epistemic underidentification problem. Even with deterministic latent dynamics, one-step Bayes supervision identifies behavior only on observed contexts and need not identify the deployed recursive predictor once rollout queries self-generated induced states whose correct local targets are not determined by numeric state alone.

What carries the argument

Induced states (model-generated states visited during rollout) and provenance variables (additional history encodings), which together decompose induced-state error into teacher-forcing/rollout mismatch, representation-class approximation, and provenance information gaps.

If this is right

  • Rollout enters a distinct induced-state regime separate from the observed training contexts.
  • Fixed induced states define a distinct local corrective task from standard teacher-forcing correction.
  • Closed-loop performance gains can arise from changing which induced states are visited, not solely from local adaptation.
  • A binary provenance encoding enables conditional further gains via provenance-aware correction.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same underidentification logic could be tested in autoregressive models outside forecasting, such as sequence generation with incomplete context.
  • Explicit measurement of provenance gaps during rollout could guide new training objectives that simulate self-generated states more directly.
  • The distinction between observed and induced states suggests examining whether similar gaps appear in reinforcement learning policies trained on partial observations.

Load-bearing premise

One-step Bayes supervision on observed contexts is insufficient to identify the correct local targets for self-generated induced states during rollout.

What would settle it

An experiment in which the one-step supervised predictor achieves identical performance on rollout queries as a predictor given full provenance information would falsify the underidentification claim.

Figures

Figures reproduced from arXiv: 2606.12990 by Riku Green, Telmo M Silva Filho, Zahraa S. Abdallah.

Figure 1
Figure 1. Figure 1: Provenance can resolve a clash between observed and corrective targets. The same numeric state (0, 1) occurs both on the observed support and as a self-generated rollout state under the Bayes-optimal one-step predictor. These two occurrences require different next-step targets. A predictor that uses only the numeric state must conflate them; augmenting the input with a binary observed/generated provenance … view at source ↗
Figure 2
Figure 2. Figure 2: Rollout regime check. Linear-probe accuracy for distinguishing observed contexts Xt from teacher-forced induced states Zh = ψ gTF h (Xt) across rollout depth h (mean ± standard deviation). Accuracy generally increases with depth, most strongly on ETTh1, moderately on MG, and weakly on Weather, suggesting that rollout progressively leaves the observed-state regime. Recursive deployment then presents not the… view at source ↗
Figure 3
Figure 3. Figure 3: Frozen induced-state evaluation. For each horizon h, we freeze teacher-forced induced states Zh = ψ gTF h (Xt) and compare the original teacher-forced predictor (TF), a probe on Zh alone (Z), and a provenance-aware probe on (Zh, Ph) (Z+P). Relearning on fixed induced states is heterogeneous: Z and Z+P often match or outperform TF on MG and Weather, but underperform strongly on ETTh1. Z+P remains close to Z… view at source ↗
Figure 4
Figure 4. Figure 4: ETTh1 raw cross-state next-step MSE. For each evaluated predictor (TF, SS, SSP), we measure the median raw next-step MSE across prediction horizon on each other’s fixed induced states. Each panel holds the evaluated model fixed and varies only the induced-state source, so this is a frozen-state diagnostic rather than a deployable rollout metric. Unlike the normalized cross-state comparison, the raw MSE mak… view at source ↗
Figure 5
Figure 5. Figure 5: MG raw cross-state next-step MSE. Median raw next-step MSE across rollout horizon for TF, SS, and SSP evaluated on fixed induced states generated by TF, SS, or SSP. Each panel fixes the evaluated model and varies only the state source. temperature, hourly) and Weather (meteorological measure￾ments). For these real-world datasets, a single univariate series is extracted (channel index 0) and standardized to… view at source ↗
Figure 6
Figure 6. Figure 6: Weather raw cross-state next-step MSE. Median raw next-step MSE across rollout horizon for TF, SS, and SSP evaluated on fixed induced states generated by TF, SS, or SSP. Each panel fixes the evaluated model and varies only the state source. These results support the main-text interpretation: correc￾tion helps not only by improving local prediction on a fixed induced-state dataset, but also by changing the … view at source ↗
Figure 7
Figure 7. Figure 7: GRU analogue of [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: GRU analogue of [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: GRU analogue of [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
read the original abstract

Recursive multi-step forecasting is usually framed as distribution shift: models are trained on observed histories but deployed on their own predictions. We show this framing is incomplete by proving that, under partial observability or state truncation, recursive rollout is also an epistemic underidentification problem. Even with deterministic latent dynamics, one-step Bayes supervision identifies behavior only on observed contexts and need not identify the deployed recursive predictor once rollout queries self-generated induced states whose correct local targets are not determined by numeric state alone. We formalize this with induced states $Z$ and provenance variables $P$, and derive a decomposition of induced-state error into teacher-forcing/rollout mismatch, representation--class approximation, and provenance information gaps. Empirically, we show that rollout enters a distinct induced-state regime, that fixed induced states define a distinct local corrective task, and that closed-loop gains arise not only from local adaptation but also from changing the induced states visited during rollout. Using a simple binary provenance encoding, provenance-aware correction can further improve performance, though gains are conditional rather than uniform. These results recast exposure bias as reasoning under self-induced epistemic uncertainty.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that exposure bias in recursive multi-step forecasting is incompletely framed as distribution shift; under partial observability or state truncation it is also an epistemic underidentification problem. Even with deterministic latent dynamics, one-step Bayes supervision on observed contexts need not identify the correct local targets for self-generated induced states Z during rollout, because those targets depend on provenance variables P not recoverable from numeric state alone. The work formalizes this via induced states Z and provenance P, derives a decomposition of induced-state error into teacher-forcing/rollout mismatch, representation-class approximation, and provenance information gaps, and reports empirical results showing that rollout enters a distinct induced-state regime, that fixed induced states define a distinct local corrective task, and that a simple binary provenance encoding yields conditional closed-loop gains.

Significance. If the formal decomposition and empirical regime distinctions hold, the reframing supplies a useful conceptual distinction between distribution shift and epistemic underidentification, together with an explicit three-term error decomposition that could inform targeted corrections. The observation that closed-loop gains arise in part from altering the induced states visited during rollout, rather than solely from local adaptation, is a concrete and potentially actionable insight. The conditional improvement from provenance-aware correction illustrates one practical direction for addressing the provenance gap.

major comments (2)
  1. [§3] §3 (formalization and proof): The central claim that one-step Bayes supervision identifies behavior only on observed contexts and need not identify the recursive predictor on induced states Z rests on showing that correct local targets are not determined by numeric state alone; the derivation must explicitly demonstrate that the provenance-gap term remains nonzero even under deterministic latent dynamics, otherwise the underidentification result reduces to a restatement of partial observability.
  2. [§4] §4 (empirical results): The claims that 'rollout enters a distinct induced-state regime' and that 'provenance-aware correction can further improve performance' are load-bearing for the practical significance of the framing; without reported quantitative metrics, dataset specifications, model architectures, or statistical comparisons against standard teacher-forcing baselines, the magnitude and reliability of these effects cannot be assessed.
minor comments (2)
  1. [Abstract] Abstract: the phrase 'using a simple binary provenance encoding' is introduced without indicating the underlying forecasting task or base model, which would help readers contextualize the conditional gains.
  2. [Notation] Notation: the distinction between induced states Z and provenance variables P would benefit from a short illustrative example immediately after their introduction to improve readability for readers unfamiliar with the provenance framing.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments. Below we respond point by point to the two major comments. We agree that both points identify areas where the manuscript can be strengthened and will revise accordingly.

read point-by-point responses

  1. Referee: [§3] §3 (formalization and proof): The central claim that one-step Bayes supervision identifies behavior only on observed contexts and need not identify the recursive predictor on induced states Z rests on showing that correct local targets are not determined by numeric state alone; the derivation must explicitly demonstrate that the provenance-gap term remains nonzero even under deterministic latent dynamics, otherwise the underidentification result reduces to a restatement of partial observability.

    Authors: We agree that the derivation would benefit from an explicit step isolating the provenance-gap term under deterministic dynamics. The current argument already separates P from the numeric state Z by definition (P encodes history or context not recoverable from the observed variables alone), but we will add a short lemma or expanded paragraph in §3 that constructs the gap explicitly: even when the latent dynamics are deterministic, the correct local target for a given induced state Z depends on the value of P, which cannot be recovered from Z. This will be inserted before the error decomposition to make the nonzero gap under determinism fully transparent. revision: yes

  2. Referee: [§4] §4 (empirical results): The claims that 'rollout enters a distinct induced-state regime' and that 'provenance-aware correction can further improve performance' are load-bearing for the practical significance of the framing; without reported quantitative metrics, dataset specifications, model architectures, or statistical comparisons against standard teacher-forcing baselines, the magnitude and reliability of these effects cannot be assessed.

    Authors: The empirical section reports performance differences and regime distinctions on concrete tasks, but we accept that the presentation would be stronger with additional quantitative detail. In the revision we will expand §4 to include: (i) exact numerical metrics (error values, deltas) in tables, (ii) full dataset names, sizes, and preprocessing steps, (iii) model architecture and hyperparameter specifications, and (iv) direct statistical comparisons against teacher-forcing baselines with confidence intervals or significance tests. These additions will allow readers to assess magnitude and reliability directly. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central argument is a theoretical decomposition of rollout error into teacher-forcing mismatch, representation-class approximation, and provenance information gaps, formalized via induced states Z and provenance variables P. This decomposition follows directly from the definitions of partial observability and self-generated states without any reduction to fitted parameters, self-citations, or ansatzes that presuppose the target result. No equations equate a 'prediction' to its own inputs by construction, and the epistemic underidentification claim is presented as a logical consequence of deterministic latent dynamics failing to determine local targets on induced states. The derivation is self-contained against external benchmarks and does not rely on load-bearing self-citation chains.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 2 invented entities

Review performed on abstract only; full paper may contain additional assumptions or entities.

axioms (2)
  • domain assumption Even with deterministic latent dynamics, one-step supervision need not identify the recursive predictor on induced states
    Stated directly in abstract as the basis for the underidentification result
  • domain assumption Partial observability or state truncation creates provenance information gaps
    Core premise enabling the distinction between observed contexts and rollout-induced states
invented entities (2)
  • induced states Z no independent evidence
    purpose: Represent states visited during recursive rollout that are not determined by numeric state alone
    New formal object used to define the distinct local corrective task and error decomposition
  • provenance variables P no independent evidence
    purpose: Capture information about how induced states were generated for provenance-aware correction
    Introduced to enable the binary encoding experiment and further performance gains

pith-pipeline@v0.9.1-grok · 5731 in / 1492 out tokens · 23107 ms · 2026-06-27T07:47:54.827679+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Deployment-Side Adaptiveness in Multi-Horizon Volatility Forecasting

    cs.LG 2026-06 unverdicted novelty 6.0

    Validation-based selection of inference-time rollout rules for multi-output volatility forecasters yields low-cost improvements over default MIMO deployment and recovers much of ensemble benefit at lower cost.

Reference graph

Works this paper leans on

25 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Realities: The Cost of MSE-Optimal Forecasting Under Conditional Uncertainty , author=

    Expectations vs. Realities: The Cost of MSE-Optimal Forecasting Under Conditional Uncertainty , author=. arXiv preprint arXiv:2606.04342 , year=

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:1810.09136 , year=

    Do deep generative models know what they don't know? , author=. arXiv preprint arXiv:1810.09136 , year=

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:1511.05101 , year=

    How (not) to train your generative model: Scheduled sampling, likelihood, adversary? , author=. arXiv preprint arXiv:1511.05101 , year=

    Pith/arXiv arXiv

  4. [4]

    Advances in neural information processing systems , volume=

    Scheduled sampling for sequence prediction with recurrent neural networks , author=. Advances in neural information processing systems , volume=

  5. [5]

    Neural computation , volume=

    A learning algorithm for continually running fully recurrent neural networks , author=. Neural computation , volume=. 1989 , publisher=

    1989

  6. [6]

    Advances in neural information processing systems , volume=

    Professor forcing: A new algorithm for training recurrent networks , author=. Advances in neural information processing systems , volume=

  7. [7]

    Proceedings of the fourteenth international conference on artificial intelligence and statistics , pages=

    A reduction of imitation learning and structured prediction to no-regret online learning , author=. Proceedings of the fourteenth international conference on artificial intelligence and statistics , pages=. 2011 , organization=

    2011

  8. [8]

    arXiv preprint arXiv:1511.06732 , year=

    Sequence level training with recurrent neural networks , author=. arXiv preprint arXiv:1511.06732 , year=

    Pith/arXiv arXiv

  9. [9]

    Chaos, Solitons & Fractals , volume=

    Robustness of LSTM neural networks for multi-step forecasting of chaotic time series , author=. Chaos, Solitons & Fractals , volume=. 2020 , publisher=

    2020

  10. [10]

    Chaos, Solitons & Fractals , volume=

    Forecasting of noisy chaotic systems with deep neural networks , author=. Chaos, Solitons & Fractals , volume=. 2021 , publisher=

    2021

  11. [11]

    arXiv preprint arXiv:2210.08959 , year=

    Flipped classroom: Effective teaching for time series forecasting , author=. arXiv preprint arXiv:2210.08959 , year=

    arXiv

  12. [12]

    Physica D: Nonlinear Phenomena , volume=

    Learning on predictions: Fusing training and autoregressive inference for long-term spatiotemporal forecasts , author=. Physica D: Nonlinear Phenomena , volume=. 2024 , publisher=

    2024

  13. [13]

    International Conference on Artificial Intelligence and Statistics , pages=

    Robust probabilistic time series forecasting , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2022 , organization=

    2022

  14. [14]

    Advances in neural information processing systems , volume=

    Predictive representations of state , author=. Advances in neural information processing systems , volume=

  15. [15]

    arXiv preprint arXiv:1207.4167 , year=

    Predictive state representations: A new theory for modeling dynamical systems , author=. arXiv preprint arXiv:1207.4167 , year=

    Pith/arXiv arXiv

  16. [16]

    Journal of Machine Learning Research , volume=

    Approximate information state for approximate planning and reinforcement learning in partially observed systems , author=. Journal of Machine Learning Research , volume=

  17. [17]

    Machine learning , volume=

    Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods , author=. Machine learning , volume=. 2021 , publisher=

    2021

  18. [18]

    Advances in neural information processing systems , volume=

    What uncertainties do we need in bayesian deep learning for computer vision? , author=. Advances in neural information processing systems , volume=

  19. [19]

    Data Mining and Knowledge Discovery , volume=

    Stratify: unifying multi-step forecasting strategies , author=. Data Mining and Knowledge Discovery , volume=. 2025 , publisher=

    2025

  20. [20]

    OpenAI blog , volume=

    Language models are unsupervised multitask learners , author=. OpenAI blog , volume=

  21. [21]

    IEEE transactions on neural networks and learning systems , volume=

    A bias and variance analysis for multistep-ahead time series forecasting , author=. IEEE transactions on neural networks and learning systems , volume=. 2015 , publisher=

    2015

  22. [22]

    arXiv preprint arXiv:2511.11461 , year=

    Epistemic Error Decomposition for Multi-step Time Series Forecasting: Rethinking Bias-Variance in Recursive and Direct Strategies , author=. arXiv preprint arXiv:2511.11461 , year=

    arXiv

  23. [23]

    Proceedings of the 3rd Workshop on Neural Generation and Translation , pages=

    Generalization in generation: A closer look at exposure bias , author=. Proceedings of the 3rd Workshop on Neural Generation and Translation , pages=

  24. [24]

    arXiv preprint arXiv:2402.08373 , year=

    Time-series classification for dynamic strategies in multi-step forecasting , author=. arXiv preprint arXiv:2402.08373 , year=

    arXiv

  25. [25]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Improving multi-step prediction of learned time series models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=