acceptodds
Under review as a conference paper at ICLR 2027

SpecSpin: Spectral Truncation and Rotation for 4-bit Zero-Shot Time-Series Foundation Models

Abstract

Time-series foundation models (TSFMs) forecast zero-shot across domains, but they are expensive to serve, and 4-bit weight quantization, the standard remedy borrowed from large language models, is not free for them: the autoregressive loop compounds rounding error over long horizons. We show that TSFM attention has spectral structure that a quantizer can exploit. Along the key/time axis, pre-softmax score rows are smooth and recency-biased, so most of their energy sits in the lowest DCT frequencies; along the input-feature axis, the q/k/v projections concentrate their energy in a small set of DCT directions. We introduce (Spectral Spin quantization), a training-free, calibration-free 4-bit recipe built on both observations. truncates each q/k/v projection to its highest-energy DCT directions, applies a Haar rotation inside the kept subspace, quantizes all linear layers to 4-bit NormalFloat (NF4), and low-pass filters every attention score row before the softmax. On three architecturally distinct TSFMs (Chronos, TimeMoE, Moirai), six benchmarks and four horizons, is the best 4-bit method on overall average for all three backbones and lowers average MSE below the BF16 model on each (, , , versus , , for uniform 4-bit quantization). Its advantage grows with the horizon, reaching (TimeMoE) and (Moirai) at .

Primary area:
learning on time series and dynamical systems
Keywords:
TSFM

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RfKxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes SpecSpin, a training-free and calibration-free 4-bit quantization method for zero-shot time-series foundation models. The key motivation is that standard LLM-style quantization does not fully fit TSFMs because forecasting errors can compound over long autoregressive horizons, and TSFM attention has exploitable spectral structure. SpecSpin uses three main steps: it applies a DCT transform to q/k/v projections and keeps the highest-energy feature directions, applies a Haar rotation within the retained subspace, quantizes linear layers with NF4, and low-pass filters attention score rows before the softmax. The method is evaluated on Chronos, TimeMoE, and Moirai across six long-term forecasting benchmarks and four horizons. The paper reports that SpecSpin achieves the best average 4-bit performance for all three backbones, with gains increasing at longer forecast horizons.

Critical strengths
  1. The paper clearly argues that TSFMs differ from language models because their attention patterns are smooth, recency-biased, and spectrally concentrated, and because autoregressive forecasting can compound quantization error across horizons.

  2. The evaluation includes three architecturally different TSFMs: Chronos, TimeMoE, and Moirai, covering encoder-only, autoregressive mixture-of-experts, and decoder-only models. This makes the empirical claim more convincing than testing on a single architecture.

Critical weaknesses
  1. The paper compares against BF16, AWQ-4bit, and FQ-4bit, but does not directly evaluate stronger modern post-training quantization methods such as GPTQ, AWQ with calibration, SmoothQuant-style activation handling, QuIP/QuaRot/SpinQuant, or other competitive low-bit methods adapted to TSFMs. This makes it harder to assess whether SpecSpin is superior to strong quantization baselines rather than mainly to simple group-wise quantization.

  2. The experiments report each result cell as a single deterministic run, so there is no uncertainty estimate over seeds, quantization randomness, or sampling variability. This is especially important for a quantization paper because the method is intended for deployment, where reliability and predictable accuracy preservation matter as much as average performance. Since some improvements over BF16 or other 4-bit baselines are modest, repeated runs, confidence intervals, or robustness checks across quantization seeds and deployment settings would make the empirical claims more convincing.

  3. Some components appear less essential than the method name suggests. The ablation indicates that removing the spin changes average MSE by at most about 0.1% on the tested Chronos settings, so the “Spin” component seems accuracy-neutral rather than a major contributor.

  4. The method relies on retention hyperparameters, sometimes tuned per backbone and dataset, which weakens the calibration-free framing. The conclusion notes that retention settings are tuned per backbone and in some cases per dataset, so practical deployment may still require nontrivial hyperparameter selection.

  5. The deployment motivation is not fully convincing. Although the paper emphasizes memory-constrained edge deployment, many time-series forecasting workloads can be served in cloud or server settings, and the evaluated TSFMs are not extremely large compared with modern foundation models. The authors should better justify why aggressive 4-bit quantization is necessary for these TSFMs and identify the concrete deployment scenarios where BF16 or 8-bit inference is insufficient.

  6. The paper provides useful empirical evidence that TSFM attention has spectral structure and that SpecSpin can exploit it, but the claim is not rigorously established. The spectral sparsity argument rests mainly on observed DCT energy concentration and downstream forecasting results, while the theoretical perturbation bound appears only as a loose sanity check. Therefore, the paper convincingly motivates the method empirically, but it does not provide a strong theoretical proof that SpecSpin generally solves quantization error accumulation in TSFMs.

Questions for authors
  1. Can the authors compare SpecSpin with stronger modern post-training quantization baselines, such as GPTQ, calibrated AWQ, SmoothQuant-style methods, QuIP/QuaRot/SpinQuant, or TSFM-adapted variants?

  2. Can the authors provide uncertainty estimates or robustness checks? Since each result cell is reported from a single deterministic run, it is unclear how stable the gains are across quantization seeds, model-loading randomness, sampling choices, or deployment settings.

  3. What is the practical contribution of the “spin” component? The ablation suggests that removing the spin changes average MSE by at most about 0.1% on the tested Chronos settings. Is the rotation essential for other backbones, lower bit-widths, or harder datasets?

  4. How much hyperparameter tuning is required in practice? The paper states that retention settings are tuned per backbone and sometimes per dataset, which seems to weaken the calibration-free framing. Can the authors provide a default setting or sensitivity analysis showing that SpecSpin works robustly without dataset-specific tuning?

  5. Can the authors better justify the need for aggressive 4-bit quantization in TSFMs? Many time-series forecasting workloads can be served in cloud or server environments, and the evaluated TSFMs are not extremely large compared with modern foundation models. What concrete deployment scenarios make BF16 or 8-bit inference insufficient?

  6. Can the authors strengthen the theoretical justification for why SpecSpin mitigates quantization error accumulation?

3 comments

Thank you for your careful review, and for recognizing the argument for why TSFMs differ from language models and the evaluation across three architecturally different backbones. Below we quote each of your comments and respond to it. Changes in the revised PDF are marked in blue.

Summary of the new results (new App. E). Following your suggestions, we added calibrated AWQ, GPTQ, QuaRot-style rotation and NF4 alone, and ran SpecSpin with one fixed setting per backbone:

  • Best 4-bit method on Chronos-2 and Moirai 2.0, without calibration data. SpecSpin is −1.6% and −5.3% vs. BF16 (Table 10). On Moirai 2.0 it is below BF16 on all six datasets, and its gain grows to −7.8% at H = 720.
  • Robust. Across five Haar seeds, the standard deviation of the six-dataset average is at most 0.0023 on these backbones. Under a paired block bootstrap over test windows, SpecSpin is significantly ahead of calibrated AWQ and rotation+GPTQ on both (App. E.2).
  • Real savings. With real NF4 kernels, weights shrink 3.6–7.6× and peak memory 1.5–5.0× (App. E.5).

Weakness 1 and Question 1.

The paper compares against BF16, AWQ-4bit, and FQ-4bit, but does not directly evaluate stronger modern post-training quantization methods such as GPTQ, AWQ with calibration, SmoothQuant-style activation handling, QuIP/QuaRot/SpinQuant, or other competitive low-bit methods adapted to TSFMs.

Response. Thank you for your comment. We added calibrated AWQ, GPTQ, QuaRot-style rotation (+RTN and +GPTQ) and NF4 alone (new App. E.1, Table 10; per horizon in Tables 11–13). The Table 1 "AWQ-4bit" baseline is round-to-nearest and is referred to as RTN; on TimeMoE it now also quantizes the expert gates, routers and output heads.

  • Setup. All methods quantize exactly the layers SpecSpin quantizes, at group size 128, from one shared implementation. AWQ and GPTQ are calibrated on 32 windows of each dataset's training split; SpecSpin uses no data and one fixed setting per backbone (Weakness 4).
  • Results. Average MSE change vs. BF16 over 6 datasets × 4 horizons:
RTNAWQGPTQRot.+RTNRot.+GPTQNF4SpecSpin
Chronos-2−0.1%−0.1%−0.4%−0.1%+0.3%−0.1%−1.6%
Moirai 2.0−3.6%−4.0%−1.1%+0.1%−0.2%−5.1%−5.3%
TimeMoE−3.8%−4.0%−2.0%−2.0%−1.8%+0.6%+1.1%
  • SpecSpin is the best 4-bit method on Chronos-2 and Moirai 2.0, ahead of every calibrated baseline. On Chronos-2, every baseline stays within 0.4% of BF16.
  • On Moirai 2.0, the NF4 grid accounts for much of the gain (NF4 alone: −5.1%), and spectral truncation adds a further 1.5 points on top of it (App. E.3).
  • On TimeMoE, SpecSpin is the best method on ETTm2 and Exchange; with one setting shared by all datasets, its average is on par with NF4 alone, and calibrated AWQ benefits most from its training data.
  • Other method families.
    • SmoothQuant targets activation quantization (W8A8). In our weight-only setting, its scale migration takes the same form as AWQ's per-channel scaling, which we include with a searched exponent.
    • SpinQuant learns its rotations by optimization, so the training-free QuaRot-style rows stand in for the rotation family.

Weakness 2 and Question 2.

The experiments report each result cell as a single deterministic run, so there is no uncertainty estimate over seeds, quantization randomness, or sampling variability. [...] repeated runs, confidence intervals, or robustness checks across quantization seeds and deployment settings would make the empirical claims more convincing.

Response. Thank you for your comment. We added uncertainty estimates for the two sources of variability in our setting (new App. E.2, Tables 14 and 15, which also include TimeMoE). Inference is deterministic given the weights, and all methods use the mean forecast, so model loading and sampling add no randomness.

  • Rotation seed. The Haar rotation is the only random element of SpecSpin. We reran SpecSpin with five seeds at H = 96 and H = 720 (six-dataset average MSE):
HSpecSpin mean ± stdRTNGPTQNF4
Chronos-2960.2296 ± 0.00040.23310.23270.2335
7200.5100 ± 0.00230.52010.51770.5205
Moirai 2.0960.2409 ± 0.00070.24120.24130.2377
7200.5090 ± 0.00200.52120.54380.5124

On Chronos-2 and Moirai 2.0, the seed spread is small compared with SpecSpin's margin over RTN and GPTQ at H = 720 (and at H = 96 on Chronos-2).

  • Test windows. Neighbouring test windows overlap, so we use a paired moving-block bootstrap: both methods' per-window errors are resampled with the same blocks of L + H consecutive windows (2,000 replicates). 95% intervals of SpecSpin's relative MSE difference from each baseline, over 6 datasets × 4 horizons (negative = SpecSpin better):
SpecSpin vs.RTNAWQGPTQRot.+GPTQNF4
Chronos-2−1.5 [−2.9, +0.04]−1.5 [−2.9, −0.3]−1.1 [−2.4, +0.07]−1.8 [−3.6, −0.3]−1.5 [−2.8, −0.4]
Moirai 2.0—−1.4 [−2.3, −0.5]−4.2 [−4.9, −3.2]−5.1 [−5.8, −4.2]−0.1 [−1.0, +0.3]

(Bold: SpecSpin significantly better. Moirai RTN: per-window errors were not stored.)

  • SpecSpin is significantly more accurate than calibrated AWQ and rotation+GPTQ on both Chronos-2 and Moirai 2.0, than GPTQ on Moirai 2.0, and than NF4 on Chronos-2.
  • Its margins over RTN and GPTQ on Chronos-2 are of similar size (−1.5% and −1.1%), with intervals reaching +0.04% and +0.07%. On Moirai 2.0, it matches NF4 alone, with which it shares the quantization grid.

Weakness 3 and Question 3.

Some components appear less essential than the method name suggests. The ablation indicates that removing the spin changes average MSE by at most about 0.1% on the tested Chronos settings, so the "Spin" component seems accuracy-neutral rather than a major contributor.

Is the rotation essential for other backbones, lower bit-widths, or harder datasets?

Response. Thank you for your comment. Following your question, we tested the spin at lower bit-widths and on all backbones. Its benefit is clearest at 3 bits and on the autoregressive backbones.

  • Lower bit-widths. We repeated the spin on/off comparison with a 3-bit NormalFloat grid (NF3), built the same way as NF4 (new App. E.4, Table 18). Average MSE change vs. BF16:
HNF3NF3 +SNF3 +T+FNF3 +T+S+F
Chronos-296+3.0%+0.6%−1.7%−1.3%
720−2.5%−3.5%−5.3%−4.6%
TimeMoE96+24.4%+15.0%+20.6%+21.6%
Moirai 2.096−1.2%−1.2%−0.8%−1.1%
720−10.9%−11.3%−11.7%−11.9%
  • On plain 3-bit weights, the spin lowers the error by 2.4 and 1.0 points on Chronos-2 and by 9.4 points on TimeMoE, compared with at most 1.2 points at 4 bits.
  • With truncation and the filter, 3-bit Chronos-2 stays 1.3–5.3% below BF16, and on top of these stages the spin adds little.
  • Other backbones. We ran all eight on/off combinations of truncation (T), spin (S) and filter (F) on NF4 weights, on all three backbones (new App. E.3, Table 16). Average MSE change vs. BF16:
NF4+S+T+T+S+T+F+T+S+F (SpecSpin)
Chronos-2−0.1%−0.4%−1.1%−0.9%−1.8%−1.6%
Moirai 2.0−5.1%−5.0%−6.6%−5.9%−6.1%−5.3%
TimeMoE+0.6%−0.3%+3.4%+1.4%+3.2%+1.1%
  • On TimeMoE, the spin is the stage that helps most, and on TimesFM (App. D.1) removing it raises the error on 5 of 6 datasets, by 1.5–5.7%.
  • On Chronos-2 and Moirai 2.0, truncation with NF4 already captures the gain.

We therefore view truncation with NF4 as the core of SpecSpin, and the spin as a stage whose benefit is largest on autoregressive backbones and at low bit-widths. Table 10 reports full SpecSpin because that configuration was fixed before these experiments.


Weakness 4 and Question 4.

The method relies on retention hyperparameters, sometimes tuned per backbone and dataset, which weakens the calibration-free framing.

Can the authors provide a default setting or sensitivity analysis showing that SpecSpin works robustly without dataset-specific tuning?

Response. Thank you for your comment. All results in App. E use one setting per backbone for every dataset and horizon:

  • Chronos-2: (ρ_W, ρ) = (0.90, 0.87).

  • TimeMoE: (0.95, 0.99).

  • Moirai 2.0: (0.90, 0.90).

  • Robustness without dataset-specific tuning. With these fixed settings, SpecSpin remains the best 4-bit method on Chronos-2 and Moirai 2.0 (Weakness 1). On TimeMoE, the fixed setting keeps SpecSpin on par with NF4 alone, and a per-dataset filter setting gives larger gains on Exchange.

  • Sensitivity. The parameter study (App. D.2) varies ρ_W, ρ and the rotated fraction one at a time. Averaged over its three datasets, every truncation level from ρ_W = 0.7 to 0.9 is at least as accurate as no truncation on Chronos-2 and Moirai 2.0. For the filter, the defaults are within 0.3% of no filter (ρ = 0.87 on Chronos-2, ρ = 0.90 on Moirai 2.0) and ρ = 0.7 is within 1.5%, whereas strong filtering hurts (ρ ≤ 0.5 on Chronos-2, ρ = 0.3 on Moirai 2.0, both about +17%).

  • Recommended setting. For a new dataset, we recommend the backbone default. The Chronos-2 and Moirai 2.0 defaults were chosen with the parameter study on ETTh1, ETTh2 and Exchange, and the TimeMoE default on ETTh1. SpecSpin needs no calibration data; its two retention ratios are set once per backbone.


Weakness 5 and Question 5.

The deployment motivation is not fully convincing. [...] The authors should better justify why aggressive 4-bit quantization is necessary for these TSFMs and identify the concrete deployment scenarios where BF16 or 8-bit inference is insufficient.

Response. Thank you for your comment. Our target is devices that must hold the model locally, where storage and memory, not throughput, are the constraint; for cloud or server serving, BF16 or 8-bit inference is often adequate. Examples are gateways and embedded monitors in grids, vehicles or medical devices, where data often cannot leave the device and several forecasters may share a few hundred MB of memory.

  • Measurements. We now measure real 4-bit inference with bitsandbytes NF4 kernels (new App. E.5, Table 19). At batch 1, L = 512 (FP = released checkpoint precision):
H = 96Weights (MB): FP / NF4 / SpecSpinPeak memory (MB)Latency (ms)
Chronos-2 (FP32 ckpt.)455.8 / 79.5 / 79.6476 / 95 / 9969 / 85 / 106
TimeMoE (BF16 ckpt.)888.4 / 246.9 / 248.61291 / 610 / 103785 / 187 / 359
Moirai 2.0 (FP32 ckpt.)43.6 / 5.7 / 6.297 / 50 / 53101 / 80 / 132

4-bit kernels shrink the weights on the GPU by 3.6–7.6× and the peak memory of a forecast by 1.5–5.0×. For example, TimeMoE-200M drops from 888 to 249 MB of weights, and Chronos-2 from 456 to 80 MB. SpecSpin's truncation and rotation add at most 2 MB.

  • Latency. At batch size 1, off-the-shelf 4-bit kernels are dominated by dequantization. SpecSpin's weight stages need no custom kernel; its latency is set by the score filter on the unfused attention path (8–92% over NF4 alone), which a fused filter kernel would remove. SpecSpin is therefore best suited to deployments bound by storage or memory.

Weakness 6 and Question 6.

The spectral sparsity argument rests mainly on observed DCT energy concentration and downstream forecasting results, while the theoretical perturbation bound appears only as a loose sanity check. [...] it does not provide a strong theoretical proof that SpecSpin generally solves quantization error accumulation in TSFMs.

Response. Thank you for your comment. We agree with this characterization: the paper provides exact properties of each stage and a stability guarantee, while the link to forecast accuracy is established empirically. Specifically:

  • Truncation removes precisely the energy of the discarded DCT directions (Eq. 4).
  • The spin is exact before rounding (Eq. 6).
  • Proposition 1 bounds the change in the attention output caused by the score filter, guaranteeing that the filter does not amplify error.

The new results support the link to forecast error most clearly on Moirai 2.0, where truncation lowers the NF4 error but leaves the unquantized model unchanged, and SpecSpin's gain grows from −0.8% at H = 96 to −7.8% at H = 720. On Chronos-2, truncation also improves the unquantized model, acting as a spectral regularizer. We will present compounding error as the motivation of the method, supported by these empirical results.


Thank you again for your review. We hope the added baselines, the uncertainty estimates and the component study address your concerns, and we would be glad to answer any further questions.

Reviewer_cQ4fOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses training-/calibration-free 4-bit post-training quantization of time-series foundation models (TSFMs). It observes two forms of spectral sparsity in TSFMs: (i) temporal — pre-softmax attention score rows are smooth/recency-biased, so energy concentrates in few low DCT frequencies; (ii) along the input-feature axis — q/k/v projection column energy concentrates in few DCT directions. Based on this, SpecSpin truncates each q/k/v projection to its highest-energy DCT directions, applies a Haar-random rotation ("spin") inside the retained subspace, quantizes all linear layers to 4-bit NormalFloat, and low-pass filters every score row before softmax. Validation: three backbones (Chronos-2, TimeMoE-200M, Moirai-2.0-small), six LTSF datasets, four horizons; claims the best 4-bit method on every backbone and average MSE below BF16 (−0.9%, −5.5%, −5.2%), with larger gains at longer horizons.

Critical strengths
  1. The empirical observation of temporal spectral sparsity in TSFM attention (and the recency-bias/smoothness explanation) is concrete, grounded, and measurable — a real property distinguishing TSFMs from NLP transformers that genuinely motivates a spectral quantization recipe rather than reusing LLM methods.
  2. On Moirai-2.0 the results are robust and non-artifactual: SpecSpin beats AWQ-4bit on 5–6/6 datasets and beats BF16 on all 6 (ETTm2 −14.3%), with consistent long-horizon gains — the one backbone where the method's contribution survives scrutiny.
Critical weaknesses
  1. The TimeMoE headline ("−5.5% vs BF16") is entirely an artifact of one dataset: Exchange alone contributes −37.8%, while on the other 5 datasets SpecSpin is +5.8% worse than BF16 and +4.2% worse than AWQ-4bit
  2. Baselines are too weak to validate the method: "AWQ-4bit" is actually group-wise RTN (no activation-aware scaling —AWQ's defining feature), inflating SpecSpin's relative advantage; and NF4-alone is absent, so the Moirai wins cannot be attributed to the spectral components rather than to merely swapping uniform grids for NormalFloat. Real AWQ/GPTQ/QuaRot are also missing.
Questions for authors

1.Please report TimeMoE results with and without Exchange and explain the −37.8% gain on Exchange from 4-bit compression. Does the TimeMoE claim survive excluding this dataset?

1 comment

Thank you for your careful review, and for recognizing that the temporal spectral sparsity of TSFM attention is a concrete, measurable property that motivates a spectral recipe, and that the Moirai results are robust. Below we quote each of your comments and respond to it. Changes in the revised PDF are marked in blue.

Summary of the new results (new App. E). Against calibrated AWQ, GPTQ, QuaRot-style rotation and NF4 alone, all on the same layers, and with one fixed SpecSpin setting per backbone, SpecSpin is the most accurate 4-bit method on Chronos-2 (−1.6% vs. BF16) and Moirai 2.0 (−5.3%; −7.8% at H = 720), without calibration data. On both backbones its margins over calibrated AWQ and rotation+GPTQ are statistically significant over test windows.


Weakness 1 and Question 1.

The TimeMoE headline ("−5.5% vs BF16") is entirely an artifact of one dataset: Exchange alone contributes −37.8%, while on the other 5 datasets SpecSpin is +5.8% worse than BF16 and +4.2% worse than AWQ-4bit

Please report TimeMoE results with and without Exchange and explain the −37.8% gain on Exchange from 4-bit compression. Does the TimeMoE claim survive excluding this dataset?

Response. Thank you for this careful analysis; we have verified your numbers. As you observe, the TimeMoE average in Table 1 is concentrated on Exchange, so we now report both views.

  • With and without Exchange (new App. E.1, Table 10 reports both). Average MSE change vs. BF16 on TimeMoE:
all 6 datasetswithout Exchange
Table 1 SpecSpin (per-dataset settings)−5.5%+5.8%
SpecSpin, one fixed setting+1.1%+4.4%
RTN, same layers−3.8%−2.5%
AWQ, calibrated on training windows−4.0%−3.2%

With the fixed setting, SpecSpin achieves the best 4-bit results on TimeMoE ETTm2 and Exchange, and its average is on par with NF4 alone (+0.6%); on this backbone, calibrated AWQ benefits most from its training data. We will state the TimeMoE result in Sec. 3.2 in this form.

  • Where the −37.8% came from. The Table 1 TimeMoE Exchange cells used a per-dataset setting with an aggressive score filter (ρ = 0.40, versus ρ = 0.99 in the fixed TimeMoE setting). At H = 720 on Exchange:

    • NF4 alone: 1.235.
    • SpecSpin at the fixed setting: 1.232.
    • SpecSpin with ρ = 0.40: 0.832.

    The gain therefore comes from the score filter, which alone substantially improves this series: BF16 TimeMoE has a high Exchange error (MSE 0.602, vs. about 0.39 for Chronos-2), and strong filtering smooths its near-random-walk forecast. With the fixed setting, the Exchange change is −8.2%, still the best of all 4-bit methods (RTN −7.3%).

  • Baseline layers. In App. E, all TimeMoE baselines quantize the same 400 layers as SpecSpin, including the expert gates, routers and output heads, which the Table 1 AWQ-4bit baseline kept in BF16.


Weakness 2.

Baselines are too weak to validate the method: "AWQ-4bit" is actually group-wise RTN (no activation-aware scaling —AWQ's defining feature), inflating SpecSpin's relative advantage; and NF4-alone is absent, so the Moirai wins cannot be attributed to the spectral components rather than to merely swapping uniform grids for NormalFloat. Real AWQ/GPTQ/QuaRot are also missing.

Response. Thank you for your comment. We added NF4 alone, calibrated AWQ, GPTQ and QuaRot-style rotation (+RTN, +GPTQ), and refer to the Table 1 baseline as RTN (new App. E.1). All methods quantize the same layers at group size 128 from one implementation; AWQ and GPTQ are calibrated on 32 training-split windows per dataset.

  • Results. Average MSE change vs. BF16 over 6 datasets × 4 horizons:
RTNAWQGPTQRot.+RTNRot.+GPTQNF4SpecSpin
Chronos-2−0.1%−0.1%−0.4%−0.1%+0.3%−0.1%−1.6%
Moirai 2.0−3.6%−4.0%−1.1%+0.1%−0.2%−5.1%−5.3%
TimeMoE−3.8%−4.0%−2.0%−2.0%−1.8%+0.6%+1.1%
  • SpecSpin stays ahead of calibrated methods. Without calibration data, it is the best 4-bit method on Chronos-2 and Moirai 2.0. Over test windows, it is significantly ahead of AWQ and rotation+GPTQ on both backbones, and of GPTQ on Moirai 2.0 (−4.2%, 95% interval [−4.9%, −3.2%]; new Table 15).
  • Chronos-2: the gain comes from the spectral stages. NF4 alone stays at the BF16 level (−0.1%), while SpecSpin reaches −1.6% and is significantly ahead of NF4.
  • Moirai 2.0. You anticipated this correctly: NF4 alone already reaches −5.1%, so much of the Moirai gain comes from the NormalFloat grid. The component study (new Table 16) shows that spectral truncation lowers the NF4 error by a further 1.5 points (−6.6%), while leaving the unquantized model unchanged.

Thank you again for your review. We hope the added baselines and the TimeMoE analysis address your concerns, and we would be glad to answer any further questions.

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

The paper proposes SpecSpin, a training-free and calibration-free method for 4-bit quantization of time series foundation models. The main idea is to use spectral truncation to keep the most important directions in the attention projection weights, apply random rotation before quantization, and filter attention scores to reduce quantization noise. The authors evaluate SpecSpin on three pretrained models across six forecasting benchmarks and four prediction horizons. The results show that SpecSpin achieves better average forecasting accuracy than the compared 4-bit baselines, while substantially reducing model storage requirements. The paper claims that SpecSpin is useful for deployment, but the experiments only use simulated quantization and theoretical computation estimates, not real 4-bit inference. The score filtering also adds extra operations and is not directly compatable with FlashAttention. Without actual inference latency and peak memory results, it is difficult to know whether SpecSpin really brings benefits in practical deployment.

Critical strengths
  1. The paper addresses an important practical problem of reducing the memory requirements of time series foundation models. SpecSpin does not require retraining or calibration data, which makes it easier to apply to existing pretrained models.
  2. The authors evaluate SpecSpin on three time series foundation models, six benchmark datasets, and four prediction horizons. The results show that SpecSpin achieves better overall average MSE than the compared 4-bit baselines on all three models.
Critical weaknesses
  1. The experimental comparisons are not strong enough to fully support the effectiveness of SpecSpin. The main baselines use uniform 4-bit quantization, while SpecSpin uses NF4, making it difficult to tell whether the improvements come from the proposed spectral operations or simply from a different quantization scheme. A direct NF4 baseline and comparisons with more established quantization methods would make the results more convincing.
  2. The paper emphasizes the deployment benefits of SpecSpin, but the experiments rely on simulated quantization and theoretical computation estimates rather than actual 4-bit inference. The proposed score filtering also introduces additional operations and is not directly compatible with FlashAttention. Without real inference latency and peak memory measurements, it is hard to judge how practical the method is for deployment.
Questions for authors
  1. The authors show that spectral truncation improves forecasting accuracy, but it is unclear whether this improvement is actually related to quantization. Could the authors apply the same spectral truncation to the original BF16 model without quantization?
  2. SpecSpin uses NF4 quantization, while the main baselines use uniform 4-bit quantization. Could the authors include a baseline that uses NF4 alone, without spectral truncation or rotation? Otherwise, it is difficult to tell how much of the improvement comes from the proposed method rather than the choice of quantization scheme.
  3. According to Table 3, removing the rotation step has almost no effect on forecasting performance, with MSE changes of at most 0.1%. Given this result, what is the actual benefit of including rotation in SpecSpin? It would be helpful to see whether this step provides more noticeable improvements on other backbones or under more aggressive quantization settings, such as 3-bit or 2-bit.
  4. The authors describe SpecSpin as calibration-free, but the paper mentions that the retention ratios are tuned for different backbones and sometimes for individual datasets. How are these ratios selected for a completely unseen dataset? Could the authors evaluate SpecSpin using the same fixed settings across all datasets to show whether the method can generalize without dataset-specific tuning?
3 comments

Thank you for your careful review, and for recognizing the practical importance of compressing time-series foundation models, the training- and calibration-free design and the breadth of the evaluation across three backbones, six benchmarks and four horizons. Below we quote each of your comments and respond to it. Changes in the revised PDF are marked in blue.

Summary of the new results (new App. E). Following your suggestions, we added NF4 alone and four established post-training quantizers, all quantizing the same layers, and ran SpecSpin with one fixed setting per backbone for every dataset and horizon:

  • Best 4-bit method on Chronos-2 and Moirai 2.0, without calibration data. SpecSpin is −1.6% and −5.3% vs. BF16, ahead of calibrated AWQ, GPTQ and QuaRot-style rotation (Table 10). On Moirai 2.0 it is below BF16 on all six datasets, and its gain grows to −7.8% at H = 720.
  • Robust. The margins over calibrated AWQ and rotation+GPTQ are significant on both backbones under a paired block bootstrap over test windows, and larger than the spread over Haar seeds (App. E.2).
  • Real savings. With real NF4 kernels, weights shrink 3.6–7.6× and peak memory 1.5–5.0× (App. E.5).

Weakness 1.

The experimental comparisons are not strong enough to fully support the effectiveness of SpecSpin. The main baselines use uniform 4-bit quantization, while SpecSpin uses NF4, making it difficult to tell whether the improvements come from the proposed spectral operations or simply from a different quantization scheme. A direct NF4 baseline and comparisons with more established quantization methods would make the results more convincing.

Response. Thank you for your comment. We added NF4 alone and four established methods (new App. E.1, Table 10; per horizon in Tables 11–13).

  • Setup. All 4-bit methods quantize exactly the layers SpecSpin quantizes, at group size 128, from one shared implementation:

    • RTN: round-to-nearest (the AWQ-4bit baseline of Table 1; on TimeMoE it now also quantizes the expert gates, routers and output heads).
    • AWQ: per-layer activation-aware scale search.
    • GPTQ.
    • QuaRot-style rotation + RTN and rotation + GPTQ.
    • NF4: SpecSpin with truncation, spin and filter all switched off.

    AWQ and GPTQ are calibrated on 32 windows of each dataset's training split; SpecSpin uses no data and one fixed setting per backbone (Question 4).

  • Results. Average MSE change vs. BF16 over 6 datasets × 4 horizons:

RTNAWQGPTQRot.+RTNRot.+GPTQNF4SpecSpin
Chronos-2−0.1%−0.1%−0.4%−0.1%+0.3%−0.1%−1.6%
Moirai 2.0−3.6%−4.0%−1.1%+0.1%−0.2%−5.1%−5.3%
TimeMoE−3.8%−4.0%−2.0%−2.0%−1.8%+0.6%+1.1%
  • Chronos-2: the gain comes from the spectral stages, not the grid. NF4 alone stays at the BF16 level (−0.1%), while SpecSpin reaches −1.6% and is below BF16 on 5 of 6 datasets; every baseline stays within 0.4% of BF16. Over test windows, SpecSpin is significantly ahead of AWQ, rotation+GPTQ and NF4 (new Table 15).
  • Moirai 2.0: best overall. SpecSpin is below BF16 on all six datasets and significantly ahead of AWQ (−1.4%), GPTQ (−4.2%) and rotation+GPTQ (−5.1%). As you anticipated, the NF4 grid accounts for much of the gain (NF4 alone: −5.1%), and spectral truncation adds a further 1.5 points on top of it (Question 1).
  • TimeMoE. SpecSpin is the best 4-bit method on ETTm2 and Exchange. With one setting shared by all datasets, its average is on par with NF4 alone; on this backbone, calibrated AWQ benefits most from its training data.

Weakness 2.

The paper emphasizes the deployment benefits of SpecSpin, but the experiments rely on simulated quantization and theoretical computation estimates rather than actual 4-bit inference. The proposed score filtering also introduces additional operations and is not directly compatible with FlashAttention. Without real inference latency and peak memory measurements, it is hard to judge how practical the method is for deployment.

Response. Thank you for your comment. We measured actual 4-bit inference with bitsandbytes NF4 kernels (new App. E.5, Table 19). SpecSpin is implemented as an index gather plus a shared Haar matrix in front of NF4 projections. Batch 1, L = 512, H = 96 (FP = released checkpoint precision):

H = 96Weights (MB): FP / NF4 / SpecSpinPeak memory (MB)Latency (ms)
Chronos-2 (FP32 ckpt.)455.8 / 79.5 / 79.6476 / 95 / 9969 / 85 / 106
TimeMoE (BF16 ckpt.)888.4 / 246.9 / 248.61291 / 610 / 103785 / 187 / 359
Moirai 2.0 (FP32 ckpt.)43.6 / 5.7 / 6.297 / 50 / 53101 / 80 / 132
  • Memory. Real 4-bit kernels cut the weights on the GPU by 3.6–7.6× and the peak memory of a forecast by 1.5–5.0× (H = 96 and 720). SpecSpin's truncation and shared rotation add at most 2 MB.
  • Latency. At batch 1, off-the-shelf NF4 kernels are dominated by dequantization at these model sizes. SpecSpin's weight stages need no custom kernel; its latency is set by the score filter, which currently runs on the unfused attention path (8–92% over NF4 alone). The +1.8–2.3% in Sec. 2.7 compares the filter with no filter on that same unfused path; App. E.5 gives the end-to-end numbers, and we will make this distinction explicit in Sec. 2.7. A fused DCT–filter–softmax kernel would remove this cost and make the filter compatible with FlashAttention-style fusion; we leave it to future work.
  • SpecSpin is therefore best suited to deployments bound by storage or memory. Since the GPU was shared with other jobs, latencies are indicative; weights and peak memory are exact.

Question 1.

The authors show that spectral truncation improves forecasting accuracy, but it is unclear whether this improvement is actually related to quantization. Could the authors apply the same spectral truncation to the original BF16 model without quantization?

Response. Thank you for this question. We applied the same truncation, with and without the score filter, to the unquantized model, as part of a full component study (new App. E.3, Table 16, which covers all three backbones; per dataset in Table 17). Average MSE change vs. BF16:

Moirai 2.0Chronos-2
BF16 + truncation−0.2%−1.2%
NF4−5.1%−0.1%
NF4 + truncation−6.6%−1.1%
  • Moirai 2.0: the gain is tied to quantization. Truncation leaves the BF16 model unchanged (−0.2%) but lowers the NF4 error by 1.5 points, consistent with removing low-energy directions before they are rounded.
  • Chronos-2: it helps both. Truncation gives a similar gain with and without quantization (−1.2% vs. −1.1%), so it also acts as a regularizer of the full-precision model.

Question 2.

SpecSpin uses NF4 quantization, while the main baselines use uniform 4-bit quantization. Could the authors include a baseline that uses NF4 alone, without spectral truncation or rotation? Otherwise, it is difficult to tell how much of the improvement comes from the proposed method rather than the choice of quantization scheme.

Response. Thank you for your comment. NF4 alone is now a row of Table 10 (see Weakness 1). In short:

  • On Chronos-2, the gain comes from the spectral stages (NF4 alone −0.1%, SpecSpin −1.6%; significant over test windows).
  • On Moirai 2.0, the NF4 grid accounts for much of the gain (NF4 alone −5.1%), and truncation adds a further 1.5 points (NF4 + truncation −6.6%).
  • On TimeMoE, SpecSpin is on par with NF4 alone (+1.1% and +0.6%).

Question 3.

According to Table 3, removing the rotation step has almost no effect on forecasting performance, with MSE changes of at most 0.1%. Given this result, what is the actual benefit of including rotation in SpecSpin? It would be helpful to see whether this step provides more noticeable improvements on other backbones or under more aggressive quantization settings, such as 3-bit or 2-bit.

Response. Thank you for your comment. Following your suggestion, we tested the spin under more aggressive quantization and on other backbones. Its benefit is clearest at 3 bits and on TimeMoE and TimesFM.

  • 3 bits. We repeated the comparison with a 3-bit NormalFloat grid (NF3) (new App. E.4, Table 18). Average MSE change vs. BF16:
HNF3NF3 +SNF3 +T+FNF3 +T+S+F
Chronos-296+3.0%+0.6%−1.7%−1.3%
720−2.5%−3.5%−5.3%−4.6%
TimeMoE96+24.4%+15.0%+20.6%+21.6%
Moirai 2.096−1.2%−1.2%−0.8%−1.1%
720−10.9%−11.3%−11.7%−11.9%
  • On plain 3-bit weights, the spin lowers the error by 2.4 and 1.0 points on Chronos-2 and by 9.4 points on TimeMoE, compared with at most 1.2 points at 4 bits. This supports its role as an outlier-spreading step under coarse grids.

  • With truncation and the filter, 3-bit Chronos-2 stays 1.3–5.3% below BF16. On top of these stages the spin adds little (between −0.3 and +1.0 points).

  • Other backbones. At 4 bits, the spin is the stage that helps most on TimeMoE (−0.9 points on NF4; App. E.3), and on TimesFM (App. D.1) removing it raises the error on 5 of 6 datasets, by 1.5–5.7%. On Chronos-2 and Moirai 2.0, truncation alone already captures the gain.

    We therefore view truncation with NF4 as the core of SpecSpin, and the spin as a stage whose benefit is largest on autoregressive backbones and at low bit-widths.

  • 2 bits. We focused on 3 bits, since at 2 bits uniform quantization already diverges on Moirai and TimesFM (App. C.1).


Question 4.

The authors describe SpecSpin as calibration-free, but the paper mentions that the retention ratios are tuned for different backbones and sometimes for individual datasets. How are these ratios selected for a completely unseen dataset? Could the authors evaluate SpecSpin using the same fixed settings across all datasets to show whether the method can generalize without dataset-specific tuning?

Response. Thank you for your comment. Following your suggestion, all results in App. E use one fixed setting per backbone for every dataset and horizon:

  • Chronos-2: (ρ_W, ρ) = (0.90, 0.87).

  • TimeMoE: (0.95, 0.99).

  • Moirai 2.0: (0.90, 0.90).

  • Generalization. With these fixed settings, SpecSpin remains the best 4-bit method on Chronos-2 and Moirai 2.0 (Weakness 1), so its advantage there does not depend on dataset-specific tuning. On TimeMoE, the fixed setting keeps SpecSpin on par with NF4 alone, and a per-dataset filter setting (ρ = 0.40) gives larger gains on Exchange.

  • Unseen datasets. We recommend the backbone default. The parameter study (App. D.2) shows wide safe ranges. Averaged over its three datasets, every truncation level from ρ_W = 0.7 to 0.9 is at least as accurate as no truncation on Chronos-2 and Moirai 2.0. For the filter, the defaults are within 0.3% of no filter and ρ = 0.7 is within 1.5%; only strong filtering hurts (ρ ≤ 0.5 on Chronos-2, ρ = 0.3 on Moirai 2.0).

  • Scope. The Chronos-2 and Moirai 2.0 defaults were chosen with the parameter study on ETTh1, ETTh2 and Exchange, and the TimeMoE default on ETTh1. SpecSpin needs no calibration data; its two retention ratios are set once per backbone.


Thank you again for your review. We hope the added baselines, the component study and the real-kernel measurements address your concerns, and we would be glad to answer any further questions.

Comments on the paper

1 comment

Dear Reviewers, thank you once again for your constructive feedback. With the discussion period nearing its end, we would greatly appreciate knowing whether our rebuttal and new experimental results have resolved your concerns. We remain available to clarify any remaining questions before the deadline.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.