acceptodds
Under review as a conference paper at ICLR 2027

Zero-Shot Generalization with CovLift: Context and Covariate Capacity

Abstract

Time-series foundation models (TSFMs) promise accurate zero-shot forecasts on unseen series. Frozen models can nevertheless fail to use predictive information already available in their inputs. We investigate how context length shapes the value of covariates and develop CovLift, a method for adapting inputs to frozen TSFMs. CovLift screens engineered covariates against observed history and matches features and context length to the forecasting model. Synthetic probes predict Chronos-2's covariate-to-numeric-only error ratio on unseen relationships. On spring electricity load with day-1 weather forecasts, eight added features harm Chronos-2 at 128 observations but reduce its geometric paired mean squared error ratio by 24% against same-length raw covariates at 1,024. This context effect replicates across sixteen unseen regions, persists under an assumed six-hour publication delay, and has the same direction on TiRex-2. CovLift also improves over raw covariates on unseen load regions. Summer evidence is inconclusive; tested additions show no confirmed Chronos-2 gain on solar or wind, and the original CovLift-Auto configuration slightly worsens aggregate benchmark performance. These findings identify context length as a determinant of covariate value and demonstrate how input design can improve zero-shot generalization on unseen load regions without updating model weights.

Primary area:
learning on time series and dynamical systems
Keywords:
time-series foundation models, zero-shot forecasting, exogenous covariates, context length, in-context learning, covariate engineering, under-generalization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_8PotOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

From the abstract, the authors supposedly propose CovLift which "screens engineered covariates against observed history and matches features and context length to the forecasting model" - but I do not understand what that means, because the paper does not explain that well.

As for the results, I understand nothing.

Critical strengths

Only if I could undersatnd the paper can I even judge it. I see no critical strengths as the entire paper was completely incomprehensible.

Critical weaknesses

The paper is incomprehensible as written. I was unable to understand anything, let alone evaluate it

The examples below are taken from the first few pages; the same pattern continues through Section 3, after which I could no longer follow the paper. I list them so the authors and the ACs can see the problem concretely.

  • Abstract: "On spring electricity load with day-1 weather forecasts, eight added features harm Chronos-2 at 128 observations but reduce its geometric paired mean squared error ratio by 24% against same-length raw covariates at 1,024."
    • What does this mean? There is no context given before these results; it jumps into low-level results directly.
  • Introduction: "A day-ahead temperature forecast cuts error tenfold for a temperature target (§4), yet can harm other targets. We study this undergeneralization across synthetic probes, energy, retail and benchmark tasks (§3–4; Fig. 1)."
    • What does this mean?
  • What is "Rounds-2/3"? Only explained in §6.
  • What is "spring load"? Is this a dataset? It is never introduced or cited in the main text.
  • What is a "preregistered contrast"?
  • What is a "reader"? Never defined.
  • What's a "channel" (Fig. 1)? Covariate? Feature?
  • "328 spring load windows" (Fig. 1 caption) - what is that? Is this a dataset? Which windows?
  • "Intervals and fixed-identity controls" (Fig. 1 caption) - what are you talking about? What does any of this mean?
  • Figure 1 is incomprehensible.
  • "vintages"? What is that? Never defined.
  • "Orthogonal probabilist–Hermite probes" - what is this? This is used for the first time on page 4 and onwards but never defined; no citation/reference.
Questions for authors

See weaknesses. I have nothing more to ask/clarify as I did not understand anything.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 8Pot: clarity, results and AI use

We thank the reviewer for identifying unclear passages. After using these terms throughout months of research, we became too accustomed to them and did not introduce them clearly enough for readers. We take responsibility for this presentation issue. We apologise for the delay while verifying additional experiments. We will introduce the task, data and terms before the results.

What problem does the paper address? A pretrained forecasting model may fail to use helpful information, such as weather forecasts. We study when this happens and whether changing the inputs can improve predictions while keeping the model's weights fixed.

What does CovLift do? CovLift constructs features from existing covariates, such as squared temperature, and screens them using observed history. It does not use future target values. A synthetic capacity profile measures which relationships each model can use. Our load gains come from expansion and screening; an additional selection benefit from the profile remains unestablished.

What does the abstract result mean? On spring electricity demand, eight added features hurt Chronos-2 with shorter histories and help with longer histories. Keeping feature identities fixed also supports the role of history length. This shows why features and history should be assessed together. These results use the original weather feed, whose timing limitation is explained below.

The terms and Figure 1.

  • A reader is the forecasting model. A covariate is an extra predictor; a channel is one input column.
  • Spring load means US Energy Information Administration hourly electricity demand from March-June 2026. Rounds 2/3 are separate evaluation batches on previously unseen grid regions.
  • A window pairs past observations with a period to predict. The 328 windows cover particular regions and forecast start times, with sufficient history to compare history lengths.
  • The geometric paired MSE ratio compares forecast errors with added features and raw covariates at the same history length. We average log ratios across windows and exponentiate; below 1 means improvement.
  • A preregistered contrast is a comparison fixed before evaluation; here, whether relative feature benefit grows with history. Bootstrap intervals estimate uncertainty by resampling data.
  • Fixed-identity controls keep the same engineered features while varying history length.
  • Vintages are versions of a weather forecast identified by reference time, which differs from publication time. The original feed includes forecasts with reference times after the demand forecast begins.
  • Undergeneralization means failing to use predictive input on a new series. The temperature example uses Open-Meteo forecasts to predict ERA5 temperature. “Tenfold” means roughly one tenth of the error without covariates; this benefit need not transfer to other targets.
  • Probabilist-Hermite probes use polynomials such as z and z^2 - 1. They are uncorrelated under standard-normal input, allowing separate tests of different degrees. We will define them and cite NIST DLMF, Table 18.3.1.
  • Figure 1: (a) constructs and screens features; (b) applies the frozen model; (c) compares relative forecast error as history length grows for Chronos-2 and TabICL. It uses the original weather feed and reselects features at each length. Fixed identities are a separate control. We will clarify the caption.

Scientific significance. The additional experiment shows that screened expansion can improve forecasts in unseen regions without updating model weights. Ridge also benefits from longer histories, while Chronos-2 shows greater sensitivity to irrelevant added features in our controls. Practically, this supports choosing features and history jointly for the model and task. The weather test remains conditional on an assumed publication delay because actual release times are unavailable.

Did we verify the AI-assisted work? Yes. We verified the code, results, citations and text, and take responsibility for the submission. We disclosed AI assistance in the submission form and AI Use Statement. The questions, hypotheses, designs and interpretations are ours. The supplement provides code and per-window evidence for submitted results, including failed tests.

We view this work as a small step towards addressing under-generalisation in time-series foundation models through covariate engineering. If this approach generalises, it could help frozen models use task-specific information, such as weather for energy demand or promotions for retail, without retraining for each task. We would appreciate it if the reviewer could assess the paper's scientific contribution using these explanations and the evidence in our responses to gZos and S1Fx, and reconsider the score and the flag. We thank the reviewer for helping us clarify the work.

Reviewer_S1FxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how the basic model of the frozen time series utilizes covariates, and how the values of artificially constructed features depend on the context length. The paper proposes an input transformation process called CovLift, which combines a fixed feature dictionary, a context-based filtering method, and diagnostic features specific to the reader, without the need to update the model weights. Through synthetic experiments, the sensitivity to changes in link functions, polynomial orders, lag terms, noise, and the relationship between covariates and the target variable was analyzed.

Critical strengths

The paper found in the spring power load experiment that engineered covariates may harm predictions in a shorter context but can turn beneficial in a longer context. Comparisons with fixed feature identities further supported the role of context length. This provides useful guidance for jointly choosing the context length and input features. The paper combined synthetic experiments, real data, multiple models, and fixed feature comparisons to validate the main findings and clearly reported the results that did not achieve clear benefits in cross-season and cross-task transfer. These analyses help understand the effective scenarios and applicable boundaries of the method.

Critical weaknesses

The authors have explicitly acknowledged that CovLift and a matched generic screening baseline produced the same model input in the audit experiment. The existing evidence provides support for the diagnostic value of profiling, but evidence for the incremental benefit of profile-guided feature selection is still limited. The authors have provided a weather data correction analysis, but the actual publication times have not been verified, and the main results of the sixteen unseen regions in Round 4 still use the legacy weather feed. Therefore, direct evidence is lacking as to whether these cross-regional benefits can be maintained when the weather forecasts are restricted to those actually available at the forecast origin.

Questions for authors

The paper shows that context length and the number of new features affect forecasting performance. Given the default settings already provided, how should users determine the initial context length and feature budget for new time series tasks based only on historical data? Could clearer and more actionable selection guidance be provided? When using the same data, features, and context lengths, would simple models such as ridge regression also exhibit the phenomenon of feature expansion being harmful in a short context and turning beneficial in a long context? Could this comparison be provided to help understand the relationship between the observed effect and general sample size effects?

1 comment
AuthorsOpenReview ↗

Response to Reviewer S1Fx: corrected feed and ridge

We thank the reviewer for the comments. We apologise for the delay while we ran and verified these experiments. We report 95% intervals; ratios below 1 mean improvement.

W1: profiling and feature selection. The evidence supports profiling as a diagnostic; its added selection benefit remains unestablished. The matched control reproduces the audited load gains, supporting a simpler screening pipeline. CovLift and the generic control select identical inputs at all 543 audited load origins. We will distinguish gains from expansion and screening from the profile's diagnostic role: frozen predictions rank Chronos-2's held-out covariate-to-numeric error ratios at rho = 0.952.

W2: forecast availability and cross-regional gains. The cross-regional gains persist under our six-hour availability assumption. We reran the same 16 unseen regions and 387 windows using weather forecasts whose reference times were at least six hours before each demand forecast began. Screened expansion still improves on raw covariates: at 1,024 hours, the error ratio is 0.826 [0.782, 0.873]. The six-hour gap is an assumed publication allowance; actual publication times remain unverified.

Table 1 compares each arm with raw covariates on the same windows, using moving-block bootstrap intervals.

history (hours) / armcorrected feedlegacy feed
128 / top-81.078 [1.020, 1.140]1.188 [1.114, 1.269]
128 / screened1.014 [0.966, 1.061]1.067 [1.012, 1.124]
512 / top-80.925 [0.879, 0.972]0.967 [0.914, 1.021]
512 / screened0.894 [0.851, 0.936]0.898 [0.853, 0.943]
1,024 / top-80.873 [0.828, 0.921]0.837 [0.779, 0.892]
1,024 / screened0.826 [0.782, 0.873]0.779 [0.723, 0.836]

Screened expansion beats raw in 13 of 16 regions at 1,024 hours.

Q1: how should users choose history and feature budget using historical data? We suggest joint rolling validation as a starting point; suitable settings may vary by task.

  1. Start with raw covariates and screened expansion at 512 observations. Where history permits, compare lengths 128, 256 and 1,024 and budgets of 2, 8 and 27 added features. These are observations; they correspond to hours in our load data.
  2. Use rolling historical forecast origins at the deployment horizon. Score only windows whose complete target horizons are observed before the current origin. Refit preprocessing and feature screening within each historical context, and use covariates available at that historical origin. Compare candidates on common validation windows.
  3. Choose the history and budget jointly using a fixed forecast-loss metric. Retain raw covariates when expansion gives no stable validation gain. Freeze the choice before the next forecast and reassess using newly completed historical windows.

Our experiments support choosing history and features jointly. On spring load, screened expansion helps Chronos-2 at 512 and 1,024 hours; eight added features can harm at 128. The proposed rolling procedure remains to be evaluated.

Q2: does ridge also change from harmful to helpful? For the tested eight-feature arm, ridge already benefits at 128 hours and improves further with longer history. This suggests validating feature budgets with the intended model. We fit ridge per window on identical features, histories and masks, using lambda = 1. Table 2 uses 328 spring windows and the legacy feed; each ratio uses that reader's own raw-covariate forecast.

history (hours)ridgeChronos-2
1280.852 [0.792, 0.916]1.120 [1.026, 1.218]
2560.732 [0.686, 0.779]1.010 [0.924, 1.096]
5120.671 [0.617, 0.722]0.920 [0.844, 0.994]
1,0240.568 [0.518, 0.617]0.760 [0.688, 0.832]

These ratios measure feature gains within each model. Across the matched grid, paired geometric Chronos-2/ridge MSE ratios span 0.34-0.66. Shared growth does not imply equal accuracy.

The log contrasts are 0.406 [0.344, 0.473] for ridge and 0.388 [0.303, 0.478] for Chronos-2. The Chronos-2 minus ridge difference is -0.018 [-0.115, 0.081]; on the corrected feed it is -0.025 [-0.125, 0.076]. These results are consistent with a shared sample-size effect. We therefore measure the growth against ridge, the matched empirical reference, in place of the 1 + p/L line.

Shuffling added features disrupts their alignment with the target. At 128 and 1,024 hours, error rises over the raw-input baseline by 13.6% and 15.2% for Chronos-2, versus 2.2% and 0.8% for ridge. This post-hoc test does not establish why the models differ.

We view this work as a small step towards addressing under-generalisation in time-series foundation models through covariate engineering. During the rebuttal, we ran additional corrected-feed and matched ridge tests. In light of these results, we would appreciate it if the reviewer could reconsider the score. We thank the reviewer for helping strengthen the work.

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

The paper asks when a frozen, covariate-capable time-series foundation model, mainly Chronos-2, can exploit future-known covariates, and whether engineering the covariate inputs alone can help. Synthetic probes build a per-reader capacity profile. On Chronos-2 a profile frozen before testing ranks held-out link ratios well (rho = 0.952), but it does not transfer to TabICL. CovLift-Auto is a training-free, per-window pipeline of truncation, lag realignment, shrinkage and screened basis expansion, gated by the profile. A preregistered prospective study on hourly EIA load in spring 2026 reports a 17 to 18% gain over raw covariates. The main finding is that added channels hurt at L = 128 and help at L = 1024. Outside spring load, the gains vanish or reverse, as the paper reports.

Critical strengths

S1. The empirical reporting is unusually transparent, controlled and reproducible, and this is a genuine strength of the paper. The evaluation is preregistered and prospective, and each round is labeled Confirmed, Replicated or Withheld (Table 1). Shuffle controls and a decontaminated checkpoint rule out leakage on the probes, and the paper reports its own negative results. The supplement ships full code, configs and per-window evidence, so every claim can be checked. This level of care makes the reported findings easy to trust.

S2. The paper isolates a clean interaction between context length and channel budget for Chronos-2, and the probes give a falsifiable predictor of how Chronos-2 reads covariate links. On spring load, the top-8 expanded channels go from harmful at L = 128 to helpful at L = 1024 (Table 12: 1.120 to 0.760). The effect holds with fixed feature identities and on 16 unseen regions. A profile frozen before testing ranks Chronos-2's held-out link ratios at rho = 0.952 (Table 4).

Critical weaknesses

W1. On the current evidence, CovLift-Auto is hard to distinguish from a generic basis-expansion-plus-screen pipeline. In the confirmed cells, expansion is the only transform that runs. A generic control with the same dictionary, screen, threshold and channel cap, but no reader profile, gives identical reader inputs at all 543 audited load origins (App. D). On retail, the selector is statistically indistinguishable from the generic control (Table 20: 1.002). The profile does not transfer to TabICL (rho = 0.19). On the 23 fev-bench covariate tasks, CovLift-Auto is 2.1% worse than raw Chronos-2. The gain over raw covariates on load is real, but the current evidence points to standard expansion and screening, rather than the reader-profiled design, as its source.

W2. The headline context-by-channel regularity is supported in a narrow setting. The clean result is for Chronos-2 on spring load (Table 13: 0.962 / 0.837 / 0.727 at L = 128 / 512 / 1024). Moirai-2.0 is harmed at every context length. At L = 1024, the winter and summer ratios are 0.979 [0.923, 1.036] and 0.991 [0.958, 1.022]. The 16-region Round 4 replication used only the legacy weather feed, where about three quarters of H = 96 horizon values have reference times after the origin (App. E). All rounds share one spring 2026 window. As it stands, the evidence supports a Chronos-2, spring-load observation rather than a general property of covariate-capable TSFMs.

W3. The comparison with existing methods sits outside the main results, is not matched, and covers the standard covariate benchmark only in part. The main tables compare CovLift only with raw covariates; external methods appear in one prose paragraph (l.446-455) and Appendix Table 21. There, TATO and TFMAdapter are not matched to CovLift: at H = 96 the baselines use L = 960 while CovLift uses L = 512, and at H = 24 the estimators differ (1.184 vs 0.998 on the same arrays). The closest covariate adapters, ChronosX, UniCA and CoRA, are only cited, and TimesFM-2.5 XReg is absent. Feature-engineering pipelines such as OpenFE or CAAFE are cited but not run, and fev-bench is used only as a subset reported in prose. Matched reruns appear feasible. As it stands, the comparison does not yet show that CovLift improves on established ways of using covariates with a frozen reader.

W4. The central regularity is empirical, context length is not a new axis, and the method combines standard parts, so the conceptual contribution is limited. The context-by-channel regularity is measured, not derived. The formal results (Props. A1 to A3) cover only the normalization study, and the observed contrast exceeds the stylized 1 + p/L reference by 7.5x without an explanation (Table 11). Its direction matches the familiar degrees-of-freedom trade-off, and look-back length is already studied for forecasters (Shi et al., NeurIPS 2024), a literature the related work does not address. On the method side, W1 shows that only the standard expansion-plus-screen part is active in the confirmed cells. The paper reads as a careful measurement study; the methodological contribution it claims is not yet supported.

Minor points, listed separately from the weaknesses above and not weighed in the assessment:

  • Several generality results, such as fev-bench, solar and wind, appear only in prose and should be tabulated with intervals.
  • Preregistered, replicated and post-hoc results are interleaved, and one table separating these tiers would help.
  • Method versions change across rounds (v2 to v6), which makes numbers hard to trace.
  • The LTSF normalization study (Table 2) is only loosely tied to the covariate thesis.
  • The legacy-feed reference-time caveat belongs in the main text, not the appendix.
  • The main text should state that the registered slope band in the basis-expansion study failed (2.06 vs [0.7, 1.3]).

Further requests, not decisive for the assessment:

  • Please rerun TATO, with re-ranking enabled, and TFMAdapter at CovLift's context length and estimator.
  • Where feasible, please add ChronosX, UniCA or CoRA, TimesFM-2.5 XReg, and a feature-engineering pipeline such as OpenFE.
  • A per-window ridge baseline on the same channels across the K x L grid would show whether the excess over 1 + p/L is specific to the frozen reader.
  • Please consider the ChronosX synthetic covariate suite as an external benchmark.
Questions for authors

Q1 (W1). Is there a held-out setting where the reader profile changes the actual inputs and improves on the generic expansion-plus-screen control with the same dictionary, screen and cap? Please report that matched comparison with its interval. An interval excluding 1 would resolve W1; if no such case exists, the contribution would rest on expansion plus screening.

Q2 (W2). Can Round 4 be rerun on the corrected weather feed? A screened/raw gain at L = 1024 with an interval below 1 would resolve the post-origin concern, while an interval including 1 would leave it open.

Q3 (W2). Is there evidence for the harmful-to-helpful crossing outside the spring 2026 window, or on a second reader such as TimesFM-2.5 with XReg? The same crossing with an interval below 1 at long context would support a general property; without it, the narrower scope would stand.

2 comments
AuthorsOpenReview ↗

Response to Reviewer gZos: evidence and controls

We thank the reviewer for the detailed comments. We apologise for the delay while we ran and verified these experiments. We report 95% intervals unless stated otherwise.

Q1/W1: is there a held-out gain over matched generic screening? The profile has predictive value, but we have not established an added selection benefit. The matched control reproduces the audited load gains with simpler screening. Both methods select identical inputs at 543 load origins; the retail selector/generic ratio is 1.002 [0.996, 1.010] (98.75% interval).

The profile's frozen predictions rank Chronos-2's held-out covariate-to-numeric error ratios at rho = 0.952. We will distinguish this result from the load gains attributable to expansion and screening. Transfer remains limited: TabICL gives rho = 0.19, and CovLift-Auto is 2.1% worse than raw Chronos-2 on 23 fev-bench tasks.

Q2/W2: do Round 4 gains persist with the corrected feed? Yes, under the six-hour availability assumption. We reran the same 16 unseen regions and 387 windows. At 1,024 hours, screened/raw is 0.826 [0.782, 0.873], meeting Q2's criterion. Top-8 harms at 128 hours (1.078 [1.020, 1.140]) and helps at 1,024 (0.873 [0.828, 0.921]). Screening improves on raw in 13 of 16 regions.

We used weather forecasts whose reference times were at least six hours before each demand forecast began. The six-hour gap is an assumed publication allowance; actual publication times remain unverified.

Q3/W2: is the crossing confirmed outside spring or on a second reader? The full crossing is currently supported only for production Chronos-2 on spring load. These settings are starting points for local validation. Winter has a positive context contrast, 0.177 [0.073, 0.287], but its long-context ratio is 0.978 [0.912, 1.044]. Summer shows no clear context effect. Six of thirteen readers show growing feature benefit with context.

W3: matched comparisons and additional baselines. To improve comparability, we aligned context and estimators by rescoring Round 1 under three estimators and running CovLift at the adapters' 960-hour context, reusing their saved forecasts. This supports testing screened expansion locally. The table uses their equal-region geometric MSE estimator, horizon 96 and legacy feed. History eligibility still gives different populations.

methodoriginsratio to raw [interval]
TATO3381.325 [1.049, 1.733]
TFMAdapter3381.013 [0.990, 1.038]
CovLift, screened3280.749 [0.661, 0.841]

At horizon 96, our adapter setups need over 512 hours. We disabled TATO test-label re-ranking to preserve prospective evaluation. Development tests include TimesFM-2.5 XReg and an output-space adapter analogue. We prioritised ridge to test linear fitting; this leaves XReg's matched performance untested. We did not add matched ChronosX/UniCA/CoRA or OpenFE runs, or the ChronosX suite.

W4: is the excess over 1 + p/L specific to the frozen reader? The ridge comparison supports a shared context benefit. Chronos-2's larger shuffled-channel penalty supports validating added features with each model. The matched grid gives log contrasts of 0.406 [0.344, 0.473] for ridge and 0.388 [0.303, 0.478] for Chronos-2; the Chronos-2 minus ridge difference is -0.018 [-0.115, 0.081]. We therefore measure the growth against ridge, the matched empirical reference, in place of the 1 + p/L line.

The table uses 328 spring windows, 14 regions and the legacy feed. Ratios use each reader's raw-covariate forecast at the same length. We permute each selected channel in time for noise-8; per-window ridge uses lambda = 1.

reader / arm128 hours1,024 hours
Chronos-2 / noise-81.136 [1.079, 1.197]1.152 [1.112, 1.194]
Chronos-2 / top-81.120 [1.026, 1.218]0.760 [0.688, 0.832]
ridge / noise-81.022 [1.012, 1.032]1.008 [1.006, 1.010]
ridge / top-80.852 [0.792, 0.916]0.568 [0.518, 0.617]

Chronos-2's feature benefit grows while its noise penalty stays near 15%. This post-hoc decomposition leaves the penalty's cause open. Across the matched grid, paired geometric Chronos-2/ridge MSE ratios range from 0.34 to 0.66.

We will discuss Shi et al.'s look-back-length work. Props. A1-A3 concern normalization. We will tabulate fev-bench, solar and wind with intervals, separate registered, replicated and post-hoc results, trace versions v2-v6, and clarify normalization's supporting role. We will put the legacy-feed caveat and failed slope criterion (2.06 versus [0.7, 1.3]) in the main text.

Together, the held-out predictions, corrected-feed regional gains and matched controls support a scoped empirical contribution: how Chronos-2 uses engineered covariates and how their value changes with history. We would appreciate it if the reviewer could reconsider the score in light of this evidence. We thank the reviewer for helping us strengthen and clarify the work.

I thank the authors for the additional experiments, and the corrected feed rerun addresses my timing concern, assuming a six-hour publication delay. The paper still sits below the bar because the rebuttal confirms rather than resolves my two decisive concerns. The authors concede they have not established any added benefit from the reader profile, and their matched generic control selects identical inputs at all 543 audited load origins, so the measured gains come from standard expansion and screening rather than the profiled design. Their own ridge comparison then shows the headline interaction between context length and channel budget is statistically indistinguishable between Chronos-2 and ridge, with an interval covering zero, so they drop their reference line and the regularity reduces to a general sample size effect rather than a property of the frozen reader. The crossing is confirmed only for Chronos-2 on spring load, and the matched comparisons against covariate adapters such as ChronosX, UniCA, CoRA, and TimesFM-2.5 XReg were not run, so it still does not show an advantage over established ways of using covariates. I maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.