Pith. sign in

REVIEW 4 major objections 6 minor 18 references

This paper claims that a prior-data fitted network can do zero-shot forecasting with exogenous covariates if the synthetic training prior and the attention mechanism are made time-aware.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-02 18:03 UTC pith:T3U55ZW2

load-bearing objection Plausible PFN adaptation to exogenous time series, but the zero-shot claim exceeds the synthetic prior and the reported experiments. the 4 major comments →

arxiv 2603.15802 v2 pith:T3U55ZW2 submitted 2026-03-16 cs.LG

Time-Aware Prior Fitted Networks for Zero-Shot Forecasting with Exogenous Variables

Andres Potapczynski , Ravi Kiran Selvam , Tatiana Konstantinova , Malcolm Wolff , Kin G. Olivares , Ruijun Ma , Michael W. Mahoney , Andrew Gordon Wilson
show 2 more authors
Boris N. Oreshkin Dmitry Efimov
This is my paper
classification cs.LG
keywords zero-shot forecastingprior-data fitted networkstime series foundation modelsexogenous covariatessynthetic data generationpositional encodingselectricity price forecastingM5 competition
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

ApolloPFN is a forecasting model built on the prior-data fitted network idea: instead of training on labeled real series, it learns to imitate Bayesian inference from synthetic datasets. The paper's claim is that the reason earlier PFN-based forecasters fail is that both their training distribution and architecture treat time points as exchangeable, and that fixing this—by generating graphs with a single root and periodic root-node dynamics, and by adding rotary and absolute position encodings plus full attention—lets a small 11M-parameter model handle exogenous variables such as promotions, prices, load and weather zero-shot. On electricity price forecasting and M5 aggregations, ApolloPFN outperforms the next-best model in most settings, and it also holds its own on univariate M-series benchmarks. Why this matters: most deployed time series foundation models ignore exogenous covariates, so a model that natively uses them without fine-tuning could improve forecasting where external drivers cause sharp changes.

Core claim

The central discovery is that the i.i.d. assumption embedded in TabPFN-TS-style training is the main obstacle, not the PFN paradigm itself. By replacing the random growing network with a Single Root Node Growing Network (SRNGN) and by exciting root nodes with randomly sampled sine and cosine functions over time, the synthetic datasets acquire temporal autocorrelation and long causal paths; combining this with rotary position embeddings (RoPE), absolute positional encodings, and full sample attention produces a model that treats order and nearness as relevant. The paper reports that these changes resolve characteristic failure modes—missed ordered patterns, weak trend extrapolation, lack of r

What carries the argument

The load-bearing machinery is the synthetic data prior. The paper introduces SRNGN, a graph generator that yields a single root node and long causal paths, and redefines root-node values as α1 sin(φ1 t) + α2 cos(φ2 t) with randomly sampled amplitudes and log-uniform frequencies. Those periodic roots propagate through the graph so every feature and target is time-correlated. On the architecture side, rotary position embeddings (RoPE) are applied in the sample-attention head to make nearby observations interact more strongly, absolute sinusoidal encodings add global position, and full attention lets future exogenous values inform each forecast horizon. Together these make order and relative di

Load-bearing premise

The weakest premise is that the synthetic training distribution—graphs with a single root whose values are a sum of one sine and one cosine—is representative enough of real time series with exogenous drivers; if real dynamics are nonstationary, regime-switching, or long-range dependent, the learned zero-shot prior will not transfer.

What would settle it

Train ApolloPFN with root nodes sampled as random walks (or regime-switching processes) instead of sine/cosine mixtures, keeping everything else fixed, and compare on M5 and electricity price benchmarks; a sharp performance drop would confirm the periodic-root prior is the load-bearing component. Alternatively, run ApolloPFN zero-shot on series with known long-range dependence (e.g., ARFIMA or Markov-switching data) and look for substantial degradation relative to models trained on such data.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Zero-shot forecasting with exogenous variables is achievable without task-specific fine-tuning, reducing deployment overhead.
  • A prior-data fitted network with only 11M parameters can be competitive with—and on several benchmarks better than—models 30 to 70 times larger.
  • The specific failure modes of TabPFN-TS (order invariance, weak trends, lack of recency, over-wide intervals) are attributable to its i.i.d. training prior and can be corrected by data and architecture changes.
  • Because ApolloPFN learns from synthetic data, the same PFN training loop can be reused for new domains by changing the prior rather than by collecting real data.

Where Pith is reading between the lines

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

  • A reader might infer that the sine/cosine root process is the real bottleneck: regimes with strong nonstationarity, random walks, or long-memory dependencies are not represented in the prior, so the model may silently revert to poor extrapolation in those settings.
  • The SRNGN single-root topology couples all exogenous features to one common cause; datasets whose exogenous variables act independently on the target may be harder for the model than datasets where features and target share a common driver.
  • The time-aware interventions (RoPE + full attention) could be transferred to other transformer-based forecasters or tabular PFNs as a generic recipe, though the paper does not test that.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces ApolloPFN, a prior-data fitted network (PFN) for zero-shot time-series forecasting with exogenous covariates. The authors modify TabPFN's synthetic data generation by (i) proposing a single-root graph-generation algorithm (SRNGN) and (ii) making the root process time-dependent via sinusoidal excitation. Architecturally, they add RoPE and absolute positional encodings and remove the train/test attention mask so that future exogenous values inform all forecast points. The model is evaluated on electricity price forecasting (sCRPS), M5 aggregations (RMSSE), and classical univariate M-series benchmarks, claiming state-of-the-art zero-shot performance against TabPFN-TS, Moirai, Chronos, and Sundial, with only 11M parameters.

Significance. If the claims hold, the paper makes a useful contribution: it extends the PFN paradigm to time series with exogenous covariates while remaining zero-shot, and its ablations (SRNGN vs. RGN, RoPE, full attention) provide concrete evidence about which design choices matter. The model is comparatively small and competitive with much larger foundation models. However, the empirical support is narrower and weaker than the text asserts: two benchmark families named in the abstract and conclusion are not evaluated, all tables are point estimates without uncertainty quantification, and the synthetic prior lacks lagged or nonstationary temporal structure. The core zero-shot generalization claim is therefore not yet established at the level of confidence the paper projects.

major comments (4)
  1. [Abstract and Section 5] The abstract states that ApolloPFN is demonstrated on 'M5, electric price forecasting, UCI Air Quality, and Solar Energy datasets.' Section 5, however, evaluates only electricity price, M5, and univariate M-series benchmarks. UCI Air Quality and Solar Energy never appear in the full text. Since the paper's central claim is SOTA performance 'across a diverse set' of exogenous benchmarks, this missing evidence is load-bearing. The authors should either add the missing experiments or remove the unsubstantiated dataset names.
  2. [Tables 1, 2, 3] All results are point estimates with no standard errors, confidence intervals, number of seeds, or significance tests. Several reported differences are small (e.g., Table 3: M4(M) ApolloPFN 0.092 vs. Moirai 0.093; M1(Y) ApolloPFN 0.142 vs. Sundial 0.119, where ApolloPFN is not best). Without uncertainty quantification, the reader cannot tell whether the SOTA claims reflect stable improvements or noise. The paper should report variability across repeated evaluations (or at least bootstrap intervals) and, where appropriate, paired significance tests.
  3. [Section 4.1 and Section 6] The synthetic prior's temporal structure is limited to a sum of one sine and one cosine per root node, with all SCM edges connecting variables at the same time step. There are no lagged dependencies (x_{t-ℓ} → y_t or y_{t-ℓ} → y_t), no stochastic trends, and no regime-switching processes in the training distribution. The conclusion (Sec. 6) explicitly concedes that dependencies absent from the training data cannot be recovered at test time. Yet the paper claims SOTA 'particularly in regimes where exogenous covariates drive sharp changes, discontinuities, or heterogeneous responses.' The only exogenous benchmarks shown involve contemporaneous covariates (load and price/promotion). This does not establish transfer to lagged or nonstationary regimes. The authors should either extend the prior with lagged edges and nonstationary root processes, or temper the claim and add benchmarks (e.g., t
  4. [Section 4.2, Eq. (2)] The absolute positional encoding formula is written as Ω_{t,2d+1} = sin(2π t 2^{2d+1} / 2^{12}) and Ω_{t,2d} = cos(2π t 2^{2d} / 2^{12}). As written, the frequency increases with d, so for d ≥ 6 the period is less than 2π and aliasing occurs. This is not the standard Vaswani-style encoding referenced in the text. If this is a typo in the manuscript, it must be corrected; if it is actually implemented, it is unclear how the architecture works. Since positional encodings are a core contribution, this formula needs clarification and the correct definition should be stated.
minor comments (6)
  1. [Section 5 opening] Typo: 'strong zero-short performance' should be 'zero-shot performance.'
  2. [Appendix B.2] Typo: 'TrancatedGamma' should be 'TruncatedGamma.'
  3. [Appendix A] Typo: 'Riemman approximation' should be 'Riemann approximation.'
  4. [Section 4.1] Typo: 'generatied' should be 'generated.'
  5. [Table 1 caption] Missing space in 'ApolloPFNbeats.' Also consider stating explicitly the number of independent trials used to produce each table.
  6. [Reproducibility] The paper does not state whether code, model checkpoints, or evaluation scripts will be released. Given the dependence on a custom synthetic-data generator and architectural choices, a reproducibility statement would strengthen the paper.

Circularity Check

0 steps flagged

No significant circularity: ApolloPFN is trained on synthetic data and evaluated on held-out real benchmarks; self-citations are background and not load-bearing.

full rationale

The paper's central claim is that a PFN trained on a synthetic prior (Section 4.1) zero-shot forecasts real benchmarks with exogenous variables (Section 5). The evaluation is genuinely held-out: Tables 1-3 report electricity price, M5, and M-series results against TabPFN-TS, Moirai, Chronos, and Sundial, and there is no evidence that test-set outcomes were used to adjust the model or its training prior. The model is not fitted to the test data, and no prediction is constructed from the benchmark values by definition. Self-citations to Potapczynski et al. (2024) and Yu et al. (2025a,b) are contextual or point to external analyses; the core architecture and data-generation contributions are introduced in this paper itself, not imported from an unverified self-citation. The conclusion's limitation paragraph ('behaviors or dependencies that are not represented in the training data cannot be reliably recovered at test time') is an honest boundary condition rather than a circular step. The most defensible criticism is a correctness-risk concern about the limited temporal structure in the synthetic prior (root excitation is a sum of one sine and one cosine, and graph edges are contemporaneous), but that is a generalization-risk critique, not a circularity reduction. Accordingly, no circular steps are identified and the score is 1.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The main free parameters are the choices defining the synthetic data prior. These are not fitted to the target benchmark data, so they do not constitute test-set fitting, but they are hand-selected and may strongly influence generalization. The key axiom is that the synthetic prior matches real-world temporal dependencies.

free parameters (5)
  • Graph attachment probability ρ = sampled from TruncatedGamma(α,β), α,β unspecified
    Controls SRNGN graph structure; values are hand-chosen and not reported, so a reader cannot reproduce a key part of the synthetic data prior.
  • Node count range = uniform between 20 and 150
    Chosen prior for the number of nodes in the synthetic SCM graph.
  • Frequency distribution for root excitation = log φ ~ U(1,10)
    The sine/cosine root nodes sample frequencies from a log-uniform range; this defines the timescale diversity of the synthetic prior.
  • Amplitude distribution for root excitation = α ~ N(0,1)
    Random amplitudes for sine/cosine components of root nodes.
  • Activation function set = tanh, sine, abs, identity, log, sigmoid, smooth relu, modulo, step
    Chosen set of nonlinearities for MLP edges in the SCM; affects the complexity of generated dependencies.
axioms (4)
  • standard math The PFN training loss L(θ) = -E log qθ(ytest|xtest,Dtrain) yields a valid approximation to the posterior predictive.
    This is the theoretical grounding of prior-data fitted networks, cited from Müller et al.; the paper relies on it without re-deriving it.
  • domain assumption The synthetic data distribution (SRNGN graphs plus sine/cosine root nodes) is representative enough of real-world time series with exogenous variables for zero-shot transfer.
    This is the load-bearing premise of the method: if the synthetic prior does not capture the structure of, e.g., M5 promotion dynamics or electricity price regimes, the model cannot generalize. The paper provides only indirect evidence via benchmark results.
  • standard math Z-scoring the target and features before inference and inverting the z-scoring on outputs preserves the predictive information.
    Standard normalization used in TabPFN and carried over here; this is a benign preprocessing assumption.
  • domain assumption RoPE and absolute positional encodings provide a beneficial inductive bias for time series order.
    The paper hypothesizes that relative and absolute temporal position information will help the attention mechanism; the ablation supports this, but it is still an assumption about the model family.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Time-Aware Prior Fitted Networks for Zero-Shot Forecasting with Exogenous Variables." pith.science (2026). https://pith.science/paper/T3U55ZW2

@misc{pith2026260315802,
  author       = {Pith},
  title        = {Pith review of: Time-Aware Prior Fitted Networks for Zero-Shot Forecasting with Exogenous Variables},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T3U55ZW2}},
  note         = {Machine review of arXiv:2603.15802}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In many time series forecasting settings, the target time series is accompanied by exogenous covariates, such as promotions and prices in retail demand; temperature in energy load; calendar and holiday indicators for traffic or sales; and grid load or fuel costs in electricity pricing. Ignoring these exogenous signals can substantially degrade forecasting accuracy, particularly when they drive spikes, discontinuities, or regime and phase changes in the target series. Most current time series foundation models (e.g., Chronos, Sundial, TimesFM, TimeMoE, TimeLLM, and LagLlama) ignore exogenous covariates and make forecasts solely from the numerical time series history, thereby limiting their performance. In this paper, we develop ApolloPFN, a prior-data fitted network (PFN) that is time-aware (unlike prior PFNs) and that natively incorporates exogenous covariates (unlike prior univariate forecasters). Our design introduces two major advances: (i) a synthetic data generation framework that injects realistic temporal patterns, structural changes, and exogenous dependencies into the PFN prior; and (ii) time-aware architectural modifications that embed inductive biases needed to exploit temporal context. We demonstrate that ApolloPFN outperforms existing baselines across several forecasting benchmarks with exogenous information, including M5, electric price forecasting, UCI Air Quality, and Solar Energy datasets.

Figures

Figures reproduced from arXiv: 2603.15802 by Andres Potapczynski, Andrew Gordon Wilson, Boris N. Oreshkin, Dmitry Efimov, Kin G. Olivares, Malcolm Wolff, Michael W. Mahoney, Ravi Kiran Selvam, Ruijun Ma, Tatiana Konstantinova.

Figure 1
Figure 1. Figure 1: Despite the clear value of exogenous information, most existing time series foundation models [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: (a) Not using exogenous information can lead to catastrophic forecasting errors. We compare the predictions of ApolloPFN with and without exogenous information for the weekly sales of a real product from the M5 benchmark. Ignoring the price information leads the forecaster to predict a decreased demand via context parroting (brown), whereas the tracking of price dynamics helps the model focus on most up-to… view at source ↗
Figure 2
Figure 2. Figure 2: Failure modes of TabPFN-TS for time series data that ApolloPFN addresses. We provide illustrative examples of each failure case with different real time series: we use a time series in Tourism Monthly for (a), in Tourism Yearly for (b), in M5 Weekly for (c), and in M1 Monthly for (d). In the plots, the training data is to the left of the black line, and the forecasts are to the right. (a) When TabPFN-TS is… view at source ↗
Figure 3
Figure 3. Figure 3: SRNGN graph generation algorithm used by ApolloPFN accelerates learning. We compare the test benchmark performance of our ApolloPFN model trained with the random growing network (RGN) algorithm and our Single Node Growing Network (SRNGN) algorithm at different training steps. With SRNGN, we achieve better performance at 20K iterations than at 80K with RGN. with RGN and one with SRNGN, leaving the rest of t… view at source ↗
Figure 4
Figure 4. Figure 4: shows a clear and consistent trend across test benchmarks: ApolloPFN achieves its strongest perfor￾mance only after all proposed modifications are introduced. In particular, the most substantial improvement arises from incorporating positional encodings. RoPE is the primary driver of this effect, as it biases the model toward prioritizing temporally proximal observations when forming predictions. At the sa… view at source ↗
Figure 5
Figure 5. Figure 5: How TabPFN combines attention across features and samples. Taken from Hollmann et al. (2025), the figure illustrates the main components of the TabPFN architecture, discussed in Equation 1, plus the translation of the embedding into a Riemann distribution approximation of the PPD p(ytest|xtest, Dtrain). The transformation of Z ∈ R N×F ×D into the Riemman approximation of the PPD is done with another linear… view at source ↗
Figure 6
Figure 6. Figure 6: Example graphs from distinct graph generation algorithms. that generating data using SRNGN accelerates training (see [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

18 extracted references · 9 linked inside Pith

  1. [1]

    GIFT-Eval: A Benchmark For General Time Series Forecasting Model Evaluation.arXiv 2410.10393,

    Taha Aksu, Gerald Woo, Juncheng Liu, Xu Liu, Chenghao Liu, Silvio Savarese, Caiming Xiong, and Doyen Sahoo. GIFT-Eval: A Benchmark For General Time Series Forecasting Model Evaluation.arXiv 2410.10393,

    Pith/arXiv arXiv

  2. [2]

    As illustrated in Figure 6 (Top), a characteristic of Algorithm 2 is that it generates graphs with many root nodes (as each added root node in might never get an incoming edge) and, if the redirection probabilityρis high, then several of the root nodes might point to the first node. When selecting which features to use from a graph, the root nodes are alw...

    2025

  3. [6]

    From Tables to Time: How TabPFN-v2 Outperforms Specialized Time Series Forecasting Models .arXiv:2501.02945,

    Shi Bin Hoo, Samuel Müller, David Salinas, and Frank Hutter. From Tables to Time: How TabPFN-v2 Outperforms Specialized Time Series Forecasting Models .arXiv:2501.02945,

    arXiv

  4. [7]

    P. L. Krapivsky and S. Redner. The Magic of Networks Grown by Redirection.arXiv 2305.10628,

    Pith/arXiv arXiv

  5. [8]

    Sundial: A Family of Highly Capable Time Series Foundation Models.arXiv 2502.00816,

    Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. Sundial: A Family of Highly Capable Time Series Foundation Models.arXiv 2502.00816,

    Pith/arXiv arXiv

  6. [11]

    RoFormer: Enhanced Transformer with Rotary Position Embedding.arXiv 2104.09864,

    Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. RoFormer: Enhanced Transformer with Rotary Position Embedding.arXiv 2104.09864,

    Pith/arXiv arXiv

  7. [12]

    Maddix, Boran Han, Xiyuan Zhang, Abdul Fatir Ansari, Oleksandr Shchur, Christos Faloutsos, Andrew Gordon Wilson, Michael W

    Annan Yu, Danielle C. Maddix, Boran Han, Xiyuan Zhang, Abdul Fatir Ansari, Oleksandr Shchur, Christos Faloutsos, Andrew Gordon Wilson, Michael W. Mahoney, and Yuyang Wang. Understanding the implicit biases of design choices for time series foundation models.arXiv preprint arXiv:2510.19236, 2025a. Annan Yu, Danielle C. Maddix, Boran Han, Xiyuan Zhang, Abdu...

    arXiv

  8. [13]

    Then, the other columns of ˜Yare filled with ˜Yi,2 = 0if the entry is observed (i= 1,...,N train) and ˜Yi,2 =−2if not (i=N train + 1,...,N)

    In terms of the target, we first create a tensor ˜Y∈R N×2 by first z-scoring all the train targets,˜Yi,1 = (yi−µ train)/σtrain, whereµ train = 1 Ntrain ∑Ntrain i=1 yi andσ 2 train = 1 Ntrain−1 ∑Ntrain i=1 (yi−µ train)2, for the positions ofi= 1,...,N train, and then by setting the rest of theNtest positionsi=N train + 1,...,Nas ˜Yi,1 =µ train. Then, the o...

    2025

  9. [14]

    In other words, we doXi←X i +Ωfor alli= 1,...,N

    After the embeddingXis constructed, we then add a fixed random positional encodingΩ∈RF−1×D to each feature shared across all Nsamples. In other words, we doXi←X i +Ωfor alli= 1,...,N. Finally, we setZ= [X,Y]∈R N×F×D , which is the embedding passed to the architecture seen in Figure 5 and discussed Equation 1 in Section 2.1. Figure5:HowTabPFNcombinesattent...

    2025

  10. [17]

    Our empirical results imply 15 0 1 2 3 4 5 6 (a) Growing random network with redirection and pref- erential attachment (Krapivsky & Redner, 2023)

    are characterized by multiple root nodes and short path lengths, as illustrated in Figure 6 (Top). Our empirical results imply 15 0 1 2 3 4 5 6 (a) Growing random network with redirection and pref- erential attachment (Krapivsky & Redner, 2023). 0 1 2 3 4 5 6 (b) Single-root growing random network inducing longer causal paths. Figure 6:Example graphs from...

    2023

  11. [18]

    Table 4 summarizes the dataset and citations for reference

    huggingface repository. Table 4 summarizes the dataset and citations for reference. In terms of Electric Prices (Lago et al., 2021), we have the following. •The Nord pool (NP) market which is one of the largest European power markets containing hourly measurements from 2023-01-01 to 2018-12-24. The NP dataset comes with exogenous variables measuring the g...

    2021

  12. [2009]

    Olivares, Malcolm Wolff, Andrew Gordon Wilson, Dmitry Efimov, and Vincent Quenneville-Belair

    Andres Potapczynski, Kin G. Olivares, Malcolm Wolff, Andrew Gordon Wilson, Dmitry Efimov, and Vincent Quenneville-Belair. Effectively Leveraging Exogenous Information across Neural Forecasters.NeurIPS TSALM 2024,

    2024

  13. [2016]

    On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258,

    Rishi Bommasani et al. On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258,

    Pith/arXiv arXiv

  14. [2019]

    Lag-Llama: Towards Foundation Models for Probabilistic Time Series Fore- casting.arXiv:2310.08278,

    Kashif Rasul, Arjun Ashok, Andrew Robert Williams, Hena Ghonia, Rishika Bhagwatkar, Arian Kho- rasani, Mohammad Javad Darvishi Bayazi, George Adamopoulos, Roland Riachi, Nadhir Hassen, Marin Biloš, Sahil Garg, Anderson Schneider, Nicolas Chapados, Alexandre Drouin, Valentina Zantedeschi, Yuriy Nevmyvaka, and Irina Rish. Lag-Llama: Towards Foundation Model...

    Pith/arXiv arXiv

  15. [2021]

    A decoder-only foundation model for time- series forecasting.arXiv 2310.10688,

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time- series forecasting.arXiv 2310.10688,

    Pith/arXiv arXiv

  16. [2023]

    more realistic

    was through a MLP, where each node is connected to all other nodes in the next layer, and the depth of the MLP is the depth of the graph which culminates with 1 node at the end which would be the target. To illustrate, if we have a 3-layered MLP with a width of10, then we would have a graph with21 = 10 + 10 + 1nodes and110 = 10×10 + 10×1edges (assuming th...

    2023

  17. [2024]

    Maddix, Hao Wang, Michael W

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Maddix, Hao Wang, Michael W. Mahoney, Kari Torkkola, Andrew Gordon Wilson, Michael Bohlke-Schneider, and Yuyang Wang. Chronos: Learning the Language ...

    Pith/arXiv arXiv

  18. [2025]

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer Normalization.arXiv 1607.06450,

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.