Pith. sign in

REVIEW 3 major objections 6 minor 49 references

Text-conditioned forecasting works better when the LLM plans over frozen time-series model trajectories instead of writing the numbers.

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 · grok-4.5

2026-07-31 12:29 UTC pith:6JNZU2XT

load-bearing objection Clean training-free split (TSFM simulator + LLM Ranker/Judge + MCTS) with broad multi-backbone evidence; gains are real but sometimes small, and text-specific credit is thinner than the abstract implies. the 3 major comments →

arxiv 2607.24892 v1 pith:6JNZU2XT submitted 2026-07-27 cs.LG cs.AI

LLM as Forecasting Planner: Training-Free Text Conditioning for Time-Series Foundation Models

Huu Hiep Nguyen , Dung Nguyen , Minh Hoang Nguyen , Dai Do , Hung Le This is my paper
classification cs.LG cs.AI
keywords text-conditioned forecastingtime-series foundation modelslarge language modelsMonte Carlo tree searchtraining-free multimodal fusionRanker-Judge planningContext-is-KeyTime-MMD
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.

Numerical history alone cannot capture shutdowns, policy shifts, or other events that live in text. Time-series foundation models already produce plausible futures, and language models can read context, but letting the language model invent or edit forecast values often wrecks the temporal structure the specialist model learned. This paper reframes the job as planning: the frozen time-series model proposes candidate continuations, and the language model only ranks partial options and scores finished paths against the text. Instantiated as LAFP, the method runs Monte Carlo tree search over the horizon with a Ranker as policy and a Judge as value function, without training either model. On Context-is-Key and Time-MMD, across two backbones and four language models, the approach matches or beats the matching context-blind backbone in every setting and usually beats flat best-of-N under the same evaluation budget. The practical claim is that sequential search is a stable, training-free way to let text steer forecasts while leaving the numbers under the specialist model’s control.

Core claim

Formulating text-conditioned forecasting as Monte Carlo tree search over trajectories generated by a frozen time-series foundation model, with an LLM Ranker supplying the policy prior and an LLM Judge supplying the value signal, yields consistent gains over the matching context-blind backbone in all sixteen LLM–backbone configurations on Context-is-Key and Time-MMD, and generally outperforms matched-budget flat Best-of-N reranking.

What carries the argument

LAFP (LLM As Forecasting Planner): MCTS over forecast windows in which the frozen TSFM is the sole numerical simulator, the Ranker LLM decides relevance and ranks candidate continuations into PUCT priors, and the Judge LLM returns an expected 0–9 score used as backed-up value; the final forecast is read from visit statistics and path values, never from LLM-generated numbers.

Load-bearing premise

The frozen time-series model must already be able to generate at least some futures that fit the text; if nothing in its candidate pool matches the event or constraint, search cannot invent the right forecast.

What would settle it

On CiK and Time-MMD, under matched Judge calls and rollouts, if LAFP failed to match or beat the context-blind TSFM in most of the sixteen LLM–backbone cells, or if flat Best-of-N systematically beat LAFP once candidate diversity is held fixed, the claim that sequential search is the effective training-free bridge would not hold.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Text can condition forecasts without any multimodal retraining if the language model is restricted to ranking and scoring TSFM samples.
  • Direct LLM forecasting remains brittle to model choice and corrupted context; planner-style use is more stable across the same LLMs.
  • Gains scale with how diverse and joint the backbone’s samples are: richer candidate pools give sequential search more room to improve.
  • Matched-budget Best-of-N is a fair control: when it loses to MCTS, the advantage is how evaluations are reused, not how many Judge calls are spent.
  • Heterogeneous foundation models can be composed by assigning each only the decisions its native training supports.

Where Pith is reading between the lines

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

  • The same planner split—specialist simulator plus LLM policy/value—could transfer to other structured prediction settings where language describes constraints but should not emit the structured object.
  • If future TSFMs expose richer conditional sampling given soft textual side information, LAFP’s ceiling would rise without changing the search design.
  • Judge score–error correlation is stronger when candidates are diverse; weaker Judges may still filter bad branches even when they cannot rank the best ones precisely.
  • Cost-sensitive deployments can cut the simulation budget roughly in half with only small error increase, suggesting a practical accuracy–latency knob the paper already plots.

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

3 major / 6 minor

Summary. The paper formulates text-conditioned time-series forecasting as a planning problem: a frozen TSFM proposes candidate continuations, and a frozen LLM acts as Ranker (policy prior) and Judge (value function) inside an MCTS loop over forecast windows (LAFP). All numerical values remain TSFM-generated; the LLM only selects and evaluates. The evaluation covers CiK (weighted RCRPS) and Time-MMD (geometric-mean MSE), two backbones (Chronos-T5-Large, TimesFM-1), four open LLMs, a matched-Judge/matched-budget TSFM-BoN control, component ablations, budget curves, cross-task text robustness, Judge–error analysis, per-domain tables, and per-seed standard deviations. Headline results: LAFP matches or beats the context-blind anchor in all 16 LLM×backbone configurations, and generally edges out matched-budget BoN. The method is clean, the evaluation is unusually thorough, and the negative results (TimesFM-1 Judge carries no signal) are reported candidly.

Significance. If the results hold, the paper offers a practical, training-free composition pattern for heterogeneous foundation models with a clear division of labor, and it is backed by a stronger-than-typical empirical package: a genuinely matched Best-of-N control (same Judge, same rollout budget, same output rule), ablations that isolate Ranker and MCTS, budget curves over all eight LLM×backbone pairs, honest cost accounting (Table 3), and disclosed failure analysis (App. I). The matched-budget design and the candor about where the method does not help (TimesFM-1) are real strengths. However, the significance of the central *framing* claim — that gains derive from textual conditioning steering the search — is currently not isolated from context-free effects: the read-off machinery (visit-proportional sampling, trust-region geometric-median pick) and anchor stabilization account for a share of the headline improvement that the manuscript does not quantify. The claims are also supported with falsifiable controls that would be cheap to run with the existing pipeline.

major comments (3)
  1. [§5, §6.2, Fig. 3a/b, Eq. (8)] Attribution of the gain to text conditioning is not isolated. The context-free components can in principle produce improvement on their own: (i) on CiK the read-off is visit-proportional sampling from the tree; (ii) on Time-MMD the trust-region pick (Eq. 8, k=B limit) is explicitly a context-free robust aggregator; (iii) Fig. 3b shows the Judge has no error signal for TimesFM-1, and Table 2 shows w/o Ranker *improves* TimesFM-1 on Time-MMD (0.751 vs 0.772), yet LAFP-TimesFM still beats its anchor — so that improvement must come from the aggregation machinery, not text. Fig. 3a does not settle this: the noise condition *augments* the true context with cross-task sentences (§6.2), so relevant information remains and the ~2% degradation cannot measure text's contribution. Please add: (a) a fully shuffled/replaced-context condition (no relevant text present), and (b) a context-blind LAFP con
  2. [§5, Table 1, Table 10, Table 11, App. I] The Time-MMD headline ('up to 60%') is inflated by degenerate anchor draws, and the LAFP-vs-BoN margin is within seed noise. The Chronos anchor's geo-mean is dominated by Agriculture zero-shot MSE 34.999 (Table 11, 'documented deterministic wild single draws') and Security 169.4 (test split +4σ, App. I); since the anchor uses one sample per window while LAFP uses 64 rollouts plus a trust-region pick, much of the 1.883→0.768 improvement is variance reduction/stabilization rather than conditioning. Consistent with this, BoN (which shares the trust-region pick) recovers nearly all of it (0.754–0.782 vs. 0.768), and the LAFP-over-BoN margins (≤0.014 geo-mean MSE) are smaller than the reported per-cell stds (0.029–0.067, Table 10). Please: (a) add a stabilization-matched context-blind anchor (e.g., median/geometric-median of 64 samples, or multi-sample anchor) so the anchor comparison isolate
  3. [§5 ('Effect of the TSFM backbone'), App. B] Backbone-generality claim rests on a surrogate sampler for TimesFM-1. App. B constructs TimesFM 'candidates' by independent per-step inverse-CDF draws from marginal deciles, which by construction destroys temporal correlation, and assigns a Gaussian-proxy pseudo-likelihood. The paper is candid that the abstention-recovers-pθ property holds only for Chronos, but the main text still frames the two-backbone results as evidence that 'LAFP benefits both backbone families' (§5). Given that the TimesFM candidate pool is not a set of backbone trajectory samples, and that both the Judge (Fig. 3b) and Ranker (Table 2) show no useful signal for this backbone, the TimesFM column currently demonstrates robustness of the read-off machinery more than backbone generality of the planner. The main text should state the quantile-reconstruction caveat prominently and soften the generality language according
minor comments (6)
  1. [§3.2, Eq. (4)] The continuous Judge (Eq. 4) requires token-level logprobs over digit tokens, restricting the method to open/self-hosted models; this limitation and its interaction with the vLLM serving setup deserve one sentence in §3.2 or §4.3.
  2. [§3.2, Eq. (1)] PUCT calibration: Judge values lie in [0,9] while cpuct=1.25, and unvisited actions get Q=0 against all-nonnegative values, so the exploration term is small relative to value differences after a few visits. A brief sensitivity note on cpuct (or normalizing values to [0,1]) would help.
  3. [§3.2 Ranking] Report the fraction of decision vs. chance nodes per benchmark (how often the Ranker abstains). This is central to the claimed gating mechanism and is currently unquantified.
  4. [Table 3] Table 3 counts 64 'TSFM traj. calls' but expansion also generates K=8 candidate blocks per expanded node (~2/3 × B expansions). Please clarify total TSFM generation cost relative to BoN's 64 full trajectories, since per-step compute is not obviously matched.
  5. [Fig. 2c, Fig. 3c] Fig. 2c (5× fewer Judge queries) and Fig. 3c are single illustrative instances; the captions should say so explicitly, and ideally be complemented by aggregate statistics (e.g., mean queries-to-best-score over instances).
  6. [References] Several bibliography entries carry 2026 dates and CoRR identifiers (Das et al. 2026, Li et al. 2026, Liu et al. 2026a/b); please verify these are stable citations rather than forthcoming preprints that may change.

Circularity Check

0 steps flagged

No significant circularity: empirical claims rest on external held-out metrics, not on self-defined scores or fitted identities.

full rationale

LAFP is a training-free systems/method paper. Its load-bearing claims are comparative forecast accuracy on external benchmarks (CiK weighted RCRPS; Time-MMD geo-mean MSE/MAE in StandardScaler space) against frozen TSFM anchors, LLM-as-forecaster baselines, and matched-budget TSFM-BoN. The Judge score s(τ)∈[0,9] is an internal MCTS backup signal (Eq. 4–5), not the reported target; final read-off still evaluates numerical trajectories produced by the frozen TSFM against held-out ground truth. There is no parameter fit that is then relabeled as a prediction, no uniqueness theorem imported from overlapping authors to force the method, and no definitional identity equating the claimed improvement to the inputs. Self-citations (e.g., Nguyen et al. on multimodal fusion) are peripheral related work, not premises of the derivation. Skeptical concerns about attributing gains to robust aggregation versus text conditioning are about causal interpretation of ablations, not circular reduction of the result to its inputs. Score 0 is appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

Load-bearing content is methodological and empirical, not axiomatic physics. The claim rests on standard MCTS/PUCT machinery, the assumption that TSFM samples span useful futures, LLM ranking/judging quality, fixed search hyperparameters, and benchmark protocols (CiK weighted RCRPS, Time-MMD geo-mean MSE). No new physical entities; Ranker/Judge/LAFP are engineering roles.

free parameters (5)
  • MCTS budget B and branching K = B=64, K=8
    Default B=64 simulations and K=8 candidates per expansion set the search/compute tradeoff and are chosen as the reference configuration, not derived.
  • Search depth D / window width W = D=8 (default)
    Horizon is split into D=⌈H/W⌉ blocks with default max depth 8; this discretizes the planning problem by hand.
  • PUCT constant cpuct and rank prior temperature η = cpuct=1.25, η=1.0
    Control exploration vs Ranker prior; set to fixed defaults (cpuct=1.25, η=1.0) without theoretical necessity.
  • Output mix α and trust-region shortlist k = α=0.5, k=5
    value(τ)= (s(τ)+α pathQ)/(1+α) with α=0.5; Time-MMD picks among top-5 nearest geometric median. Authors note α and k fixed on offline development trees before final runs.
  • Chronos sampling temperature T = T=1.5
    T=1.5 used for candidate generation to increase diversity; affects candidate support and abstention behavior.
axioms (5)
  • standard math PUCT-guided MCTS with policy prior and backed-up values is a valid way to search a large trajectory tree under a finite simulation budget.
    Imported from AlphaGo/AlphaZero-style search; used in §3.2 Selection without re-derivation.
  • domain assumption Frozen TSFM samples (or quantile-reconstructed paths) are sufficiently diverse and temporally plausible that selection among them can improve context-conditioned error.
    Stated limit in §3.2 and backbone analysis §5; TimesFM product-of-marginals reconstruction weakens joint diversity (App. B).
  • domain assumption An instruction-tuned LLM can act as a useful policy (relevance+ranking) and value function (0–9 expected score) over serialized trajectories without fine-tuning.
    Core operational premise of Ranker/Judge; quality varies by LLM and is weakest for Gemma-3-4B on Time-MMD Chronos.
  • domain assumption CiK weighted RCRPS and Time-MMD geometric-mean MSE (StandardScaler, TaTS-style windows) are appropriate primary metrics for the claim of consistent improvement.
    Evaluation protocol §4–5 and App. F; aggregation choices affect headline numbers.
  • ad hoc to paper When Ranker marks context irrelevant, uniform sampling among TSFM candidates is the correct local transition (chance node).
    Design choice in §3.2 to avoid irrelevant text steering; for Chronos claimed to recover backbone sampling, not for TimesFM.
invented entities (2)
  • LAFP (LLM As Forecasting Planner) framework no independent evidence
    purpose: Name the training-free MCTS composition of frozen TSFM simulator + Ranker policy + Judge value.
    Methodological system, not a latent physical object; evaluated only via forecast metrics in this paper.
  • Decision vs chance nodes conditioned on Ranker relevance no independent evidence
    purpose: Gate where textual priors enter the search tree versus uniform TSFM sampling.
    Paper-specific search ontology; falsifiable only insofar as ablations show Ranker on/off effects.

pith-pipeline@v1.2.0-grok45-kimik3 · 31212 in / 3668 out tokens · 69353 ms · 2026-07-31T12:29:02.564414+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of LLM as Forecasting Planner: Training-Free Text Conditioning for Time-Series Foundation Models." pith.science (2026). https://pith.science/paper/6JNZU2XT

@misc{pith2026260724892,
  author       = {Pith},
  title        = {Pith review of: LLM as Forecasting Planner: Training-Free Text Conditioning for Time-Series Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6JNZU2XT}},
  note         = {Machine review of arXiv:2607.24892}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Text-conditioned time-series forecasting predicts a series from both its numerical history and natural-language context, allowing forecasts to account for events and constraints that the past alone cannot reveal. This requires both reliable numerical forecasting and the ability to interpret contextual information. Time-series foundation models (TSFMs) provide strong numerical forecasts, while large language models (LLMs) can reason over text, but combining their strengths remains challenging because asking an LLM to generate or revise forecast values directly can distort the temporal structure captured by the TSFM. We instead formulate forecasting as a planning problem over TSFM-generated trajectories. The frozen TSFM acts as a simulator that proposes numerical continuations, while the LLM acts as a policy and value function that guides candidate selection and evaluates completed trajectories against the context. We instantiate this as \rc{} (\textbf{L}LM \textbf{A}s \textbf{F}orecasting \textbf{P}lanner), a training-free framework that bridges the modality gap without retraining either model, using Monte Carlo tree search (MCTS) over the forecast horizon with a \emph{Ranker} LLM as policy and a \emph{Judge} LLM as value function. Experiments on Context-is-Key and Time-MMD across two TSFM backbones (Chronos and TimesFM) and four LLMs show that \rc{} delivers consistent improvements across model choices, supporting sequential search as an effective training-free approach to text-conditioned forecasting.

Figures

Figures reproduced from arXiv: 2607.24892 by Dai Do, Dung Nguyen, Hung Le, Huu Hiep Nguyen, Minh Hoang Nguyen.

Figure 1
Figure 1. Figure 1: LAFP overview. The frozen TSFM serves as the simulator, generating candi￾date numerical continuations and completing forecast trajectories. The LLM provides both the policy and value signals: the Ranker assesses the relevance of the context and ranks candidate continuations, while the Judge evaluates completed trajectories. Each of the B simulations proceeds through Selection, Expansion, Ranking, Simulatio… view at source ↗
Figure 2
Figure 2. Figure 2: Budget curve, Chronos-T5-Large TSFM with Qwen2.5-7B Ranker/Judge. (a) CiK (weighted RCRPS) and (b) Time-MMD (geo-mean MSE): LAFP vs. TSFM-BoN at matched Judge-call budget B. (c) Best-so-far Judge score vs. Judge queries. from numerical generation, LAFP maintains substantially more stable performance across the same LLM choices. Comparison with Best-of-N baseline. TSFM-BoN uses the same TSFM backbone, Judge… view at source ↗
Figure 3
Figure 3. Figure 3: (a) CiK weighted RCRPS, true vs. cross-instance shuffled context; dashed line: context-blind anchor. (b) Mean within-window error percentile vs. Judge score. (c) Fore￾cast comparison of LAFP and baselines. Failure modes of direct numerical generation. The per-domain results in App. G reveal several extreme failures in Security, Health, and Environment, primarily for weaker LLM-as-forecaster configurations.… view at source ↗
Figure 4
Figure 4. Figure 4: CiK budget curves per Ranker/Judge LLM × TSFM pair: weighted RCRPS (↓) vs. Judge-call budget B, LAFP vs. TSFM-BoN at matched budget, averaged over three seeds. K.4 LLM-CoT LLM-CoT shares DirectPrompt’s template and differs only in the output instruction, which asks for explicit reasoning before the forecast; the full prompt is: SYSTEM: You are a useful forecasting assistant. USER: I have a time series fore… view at source ↗
Figure 5
Figure 5. Figure 5: Time-MMD budget curves per Ranker/Judge LLM × TSFM pair: geometric￾mean MSE (↓) over the nine domains vs. Judge-call budget B, LAFP vs. TSFM-BoN at matched budget [PITH_FULL_IMAGE:figures/full_fig_p026_5.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

49 extracted references · 1 canonical work pages

  1. [1]

    Gonzalez and Ion Stoica , booktitle=

    Lianmin Zheng and Wei-Lin Chiang and Ying Sheng and Siyuan Zhuang and Zhanghao Wu and Yonghao Zhuang and Zi Lin and Zhuohan Li and Dacheng Li and Eric Xing and Hao Zhang and Joseph E. Gonzalez and Ion Stoica , booktitle=. Judging. 2023 , url=

    2023

  2. [2]

    CoRR , volume =

    Karl Cobbe and Vineet Kosaraju and Mohammad Bavarian and Mark Chen and Heewoo Jun and Lukasz Kaiser and Matthias Plappert and Jerry Tworek and Jacob Hilton and Reiichiro Nakano and Christopher Hesse and John Schulman , title =. CoRR , volume =

  3. [3]

    The Twelfth International Conference on Learning Representations , year=

    Let's Verify Step by Step , author=. The Twelfth International Conference on Learning Representations , year=

  4. [4]

    Transactions on Machine Learning Research , issn=

    Chronos: Learning the Language of Time Series , author=. Transactions on Machine Learning Research , issn=. 2024 , url=

    2024

  5. [5]

    Proceedings of the 41st International Conference on Machine Learning , pages =

    Unified Training of Universal Time Series Forecasting Transformers , author =. Proceedings of the 41st International Conference on Machine Learning , pages =. 2024 , editor =

    2024

  6. [6]

    Proceedings of the 41st International Conference on Machine Learning , pages =

    A decoder-only foundation model for time-series forecasting , author =. Proceedings of the 41st International Conference on Machine Learning , pages =. 2024 , editor =

    2024

  7. [7]

    arXiv preprint arXiv:2310.08278 , year =

    Lag-Llama: Towards Foundation Models for Probabilistic Time Series Forecasting , author =. arXiv preprint arXiv:2310.08278 , year =

    Pith/arXiv arXiv

  8. [8]

    Garza, Azul and Challu, Cristian and Mergenthaler-Canseco, Max , journal =

  9. [9]

    Attention is All you Need , url =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and Kaiser, ukasz and Polosukhin, Illia , booktitle =. Attention is All you Need , url =

  10. [10]

    Forty-second International Conference on Machine Learning , year=

    Context is Key: A Benchmark for Forecasting with Essential Textual Information , author=. Forty-second International Conference on Machine Learning , year=

  11. [11]

    Sasanur and Megha Sharma and Jiaming Cui and Qingsong Wen and Chao Zhang and B

    Haoxin Liu and Shangqing Xu and Zhiyuan Zhao and Lingkai Kong and Harshavardhan Kamarthi and Aditya B. Sasanur and Megha Sharma and Jiaming Cui and Qingsong Wen and Chao Zhang and B. Aditya Prakash , booktitle=. Time-. 2024 , url=

    2024

  12. [12]

    Zhang and Xiaoming Shi and Pin-Yu Chen and Yuxuan Liang and Yuan-Fang Li and Shirui Pan and Qingsong Wen , booktitle=

    Ming Jin and Shiyu Wang and Lintao Ma and Zhixuan Chu and James Y. Zhang and Xiaoming Shi and Pin-Yu Chen and Yuxuan Liang and Yuan-Fang Li and Shirui Pan and Qingsong Wen , booktitle=. Time-. 2024 , url=

    2024

  13. [13]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Large Language Models Are Zero-Shot Time Series Forecasters , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  14. [14]

    Beyond Na

    Arjun Ashok and Andrew Robert Williams and Vincent Zhihao Zheng and Irina Rish and Nicolas Chapados and. Beyond Na. Recent Advances in Time Series Foundation Models Have We Reached the 'BERT Moment'? , year=

  15. [15]

    arXiv preprint arXiv:2412.15115 , year =

    Pith/arXiv arXiv

  16. [16]

    Lillicrap and Fan Hui and Laurent Sifre and George van den Driessche and Thore Graepel and Demis Hassabis , title =

    David Silver and Julian Schrittwieser and Karen Simonyan and Ioannis Antonoglou and Aja Huang and Arthur Guez and Thomas Hubert and Lucas Baker and Matthew Lai and Adrian Bolton and Yutian Chen and Timothy P. Lillicrap and Fan Hui and Laurent Sifre and George van den Driessche and Thore Graepel and Demis Hassabis , title =. Nat. , volume =

  17. [17]

    European Conference on Machine Learning , year=

    Bandit Based Monte-Carlo Planning , author=. European Conference on Machine Learning , year=

  18. [18]

    Annals of Mathematics and Artificial Intelligence , volume =

    Multi-Armed Bandits with Episode Context , author =. Annals of Mathematics and Artificial Intelligence , volume =

  19. [19]

    Reasoning with Language Model is Planning with World Model

    Hao, Shibo and Gu, Yi and Ma, Haodi and Hong, Joshua and Wang, Zhen and Wang, Daisy and Hu, Zhiting. Reasoning with Language Model is Planning with World Model. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.507

    doi:10.18653/v1/2023.emnlp-main.507 2023

  20. [20]

    2023 , isbn =

    Kwon, Woosuk and Li, Zhuohan and Zhuang, Siyuan and Sheng, Ying and Zheng, Lianmin and Yu, Cody Hao and Gonzalez, Joseph and Zhang, Hao and Stoica, Ion , title =. 2023 , isbn =. doi:10.1145/3600006.3613165 , booktitle =

    arXiv 2023

  21. [21]

    Journal of the American Statistical Association , volume =

    Strictly Proper Scoring Rules, Prediction, and Estimation , author =. Journal of the American Statistical Association , volume =

  22. [22]

    International Journal of Forecasting , volume =

    Another Look at Measures of Forecast Accuracy , author =. International Journal of Forecasting , volume =

  23. [23]

    Matheson and Robert L

    James E. Matheson and Robert L. Winkler , journal =. Scoring Rules for Continuous Probability Distributions , urldate =

  24. [24]

    The Eleventh International Conference on Learning Representations , year=

    A Time Series is Worth 64 Words: Long-term Forecasting with Transformers , author=. The Eleventh International Conference on Learning Representations , year=

  25. [25]

    Liu, Xu and Liu, Juncheng and Woo, Gerald and Aksu, Taha and Liang, Yuxuan and Zimmermann, Roger and Liu, Chenghao and Li, Junnan and Savarese, Silvio and Xiong, Caiming and Sahoo, Doyen , booktitle =. Moirai-. 2025 , editor =

    2025

  26. [26]

    Nexus : An Agentic Framework for Time Series Forecasting , journal =

    Sarkar Snigdha Sarathi Das and Palash Goyal and Mihir Parmar and Nanyun Peng and Vishy Tirumalashetty and Chun. Nexus : An Agentic Framework for Time Series Forecasting , journal =

  27. [27]

    CoRR , volume =

    Yuhua Liao and Zetian Wang and Qiangqiang Nie and Zhenhua Zhang , title =. CoRR , volume =

  28. [28]

    2025 , howpublished =

    2025

  29. [29]

    One Fits All: Power General Time Series Analysis by Pretrained

    Tian Zhou and Peisong Niu and Xue Wang and Liang Sun and Rong Jin , booktitle=. One Fits All: Power General Time Series Analysis by Pretrained. 2023 , url=

    2023

  30. [30]

    , title =

    Chow, C. , title =. 2006 , issue_date =. doi:10.1109/TIT.1970.1054406 , month = sep, pages =

    arXiv 2006

  31. [31]

    International Conference on Algorithmic Learning Theory , year=

    Learning with Rejection , author=. International Conference on Algorithmic Learning Theory , year=

  32. [32]

    Selective Classification for Deep Neural Networks , url =

    Geifman, Yonatan and El-Yaniv, Ran , booktitle =. Selective Classification for Deep Neural Networks , url =

  33. [33]

    Predict Responsibly: Improving Fairness and Accuracy by Learning to Defer , url =

    Madras, David and Pitassi, Toni and Zemel, Richard , booktitle =. Predict Responsibly: Improving Fairness and Accuracy by Learning to Defer , url =

  34. [34]

    Proceedings of the 37th International Conference on Machine Learning , pages =

    Consistent Estimators for Learning to Defer to an Expert , author =. Proceedings of the 37th International Conference on Machine Learning , pages =. 2020 , editor =

    2020

  35. [35]

    Language Models Still Struggle to Zero-shot Reason about Time Series

    Merrill, Mike A and Tan, Mingtian and Gupta, Vinayak and Hartvigsen, Thomas and Althoff, Tim. Language Models Still Struggle to Zero-shot Reason about Time Series. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.201

    work page doi:10.18653/v1/2024.findings-emnlp.201 2024

  36. [36]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Are Language Models Actually Useful for Time Series Forecasting? , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  37. [37]

    The Fourteenth International Conference on Learning Representations , year=

    Language in the Flow of Time: Time-Series-Paired Texts Weaved into a Unified Temporal Narrative , author=. The Fourteenth International Conference on Learning Representations , year=

  38. [38]

    The 29th International Conference on Artificial Intelligence and Statistics , year=

    Spectral Text Fusion: A Frequency-Aware Approach to Multimodal Time-Series Forecasting , author=. The 29th International Conference on Artificial Intelligence and Statistics , year=

  39. [39]

    Lillicrap and Karen Simonyan and Demis Hassabis , title =

    David Silver and Thomas Hubert and Julian Schrittwieser and Ioannis Antonoglou and Matthew Lai and Arthur Guez and Marc Lanctot and Laurent Sifre and Dharshan Kumaran and Thore Graepel and Timothy P. Lillicrap and Karen Simonyan and Demis Hassabis , title =. CoRR , volume =

  40. [40]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  41. [41]

    Guan, Xinyu and Zhang, Li Lyna and Liu, Yifei and Shang, Ning and Sun, Youran and Zhu, Yi and Yang, Fan and Yang, Mao , booktitle =. r. 2025 , editor =

    2025

  42. [42]

    Learning to summarize with human feedback , url =

    Stiennon, Nisan and Ouyang, Long and Wu, Jeffrey and Ziegler, Daniel and Lowe, Ryan and Voss, Chelsea and Radford, Alec and Amodei, Dario and Christiano, Paul F , booktitle =. Learning to summarize with human feedback , url =

  43. [43]

    CoRR , volume =

    Reiichiro Nakano and Jacob Hilton and Suchir Balaji and Jeff Wu and Long Ouyang and Christina Kim and Christopher Hesse and Shantanu Jain and Vineet Kosaraju and William Saunders and Xu Jiang and Karl Cobbe and Tyna Eloundou and Gretchen Krueger and Kevin Button and Matthew Knight and Benjamin Chess and John Schulman , title =. CoRR , volume =

  44. [44]

    2026 , eprint=

    Does Text Actually Help? Uncovering and Resolving Text Collapse in Multimodal Time Series Forecasting , author=. 2026 , eprint=

    2026

  45. [45]

    Estimation of the Continuous Ranked Probability Score with Limited Information and Applications to Ensemble Weather Forecasts , volume =

    Zamo, Michaël and Naveau, Philippe , year =. Estimation of the Continuous Ranked Probability Score with Limited Information and Applications to Ensemble Weather Forecasts , volume =. Mathematical Geosciences , doi =

  46. [46]

    Evaluating Forecasts for High-Impact Events Using Transformed Kernel Scores , volume =

    Allen, Sam and Ginsbourger, David and Ziegel, Johanna , year =. Evaluating Forecasts for High-Impact Events Using Transformed Kernel Scores , volume =. SIAM/ASA Journal on Uncertainty Quantification , doi =

  47. [47]

    Aditya Prakash and Abhimanyu Das , title =

    Haoxin Liu and Yichen Zhou and Rajat Sen and B. Aditya Prakash and Abhimanyu Das , title =. CoRR , volume =

  48. [48]

    Proceedings of the 43rd International Conference on Machine Learning , year =

    Rethinking Multimodal Time-Series Forecasting Evaluation , author =. Proceedings of the 43rd International Conference on Machine Learning , year =

  49. [49]

    Advances in Neural Information Processing Systems , editor=

    Chain of Thought Prompting Elicits Reasoning in Large Language Models , author=. Advances in Neural Information Processing Systems , editor=. 2022 , url=

    2022