acceptodds
Under review as a conference paper at ICLR 2027

TSLM-ICL: Toward Effective In-Context Learning for Time Series Language Models

Abstract

Time series language models (TSLMs) combine temporal signals with natural-language information, enabling tasks such as perception, reasoning, and decision making. In-context learning (ICL) offers a training-free way to adapt frozen TSLMs through demonstrations, yet existing time-series ICL methods are largely task-specific and do not address the heterogeneous structure of multimodal time-series data. We introduce **TSLM-ICL**, a training-free, model-agnostic retrieval framework for multimodal demonstration selection. TSLM-ICL represents samples using token-level temporal and textual embeddings, compares variable-sized representations via optimal transport, and selects relevant demonstrations under a fixed context budget. Across five public multimodal time-series benchmarks and six frozen TSLMs, TSLM-ICL improves accuracy by up to 16% over zero-shot prompting and consistently outperforms random, text-only, TimeRAG, and TS-RAG retrieval, achieving the best average accuracy in 14 of 15 dataset–budget configurations.

Primary area:
foundation or frontier models, including LLMs
Keywords:
ICL, time series language models, retrieval, optimal transport

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nCdwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TSLM-ICL, a training-free method for selecting demonstrations for time-series language models. It combines time-series token embeddings with a pooled text embedding and uses optimal transport to rank candidate examples. Experiments on five benchmarks and six frozen models show improvements over random, text-only, and adapted time-series retrieval baselines. The ablations suggest that temporal features account for most of the gains, while optimal transport provides greater benefits on multivariate tasks.

Critical strengths
  1. The method provides a practical, training-free approach to demonstration selection for time-series language models. It preserves token-level temporal information and accommodates multivariate sequences of varying lengths without updating the target model.

  2. The evaluation covers five benchmarks and six models with different input representations, showing overall gains over the compared retrieval baselines. The ablations help identify where these gains arise: temporal features contribute most, while token-level optimal transport provides additional benefits on multivariate tasks.

Critical weaknesses

The variance-based motivation for OT in §4.2 conflates loss of distributional geometry with loss of scalar variance. For unit-normalized tokens, Eq. (14) gives σ2=1−∣eˉ∣22\sigma^2=1-\\|\bar e\\|_2^2. Thus, the pooled mean before further normalization exactly determines the scalar variance, and Euclidean comparison remains sensitive to this norm information, unlike cosine similarity. Eqs. (3)–(4) quantify reconstruction residuals, but do not establish that this scalar energy is unavailable from the mean or requires token-level comparison to recover. Moreover, Eq. (7)’s premise—identical means but different scalar variances—is unattainable under these assumptions.

For example, consider uniform distributions on

\qquad B=\left\\{\frac{(1,0,1)}{\sqrt2},\frac{(1,0,-1)}{\sqrt2}\right\\}.$$ Both have mean $(1/\sqrt2,0,0)$ and scalar variance $1/2$, yet every cross-set distance equals one, giving $W_1(A,B)=1$. OT therefore distinguishes geometry beyond the mean and scalar variance.
Questions for authors
  1. Given Eq. (14), the pooled mean before further normalization exactly determines the scalar variance. How do the authors reconcile this with §4.2’s claim that pooling followed by Euclidean comparison discards that variance? Please revise the energy-loss interpretation of Eqs. (3)–(4) and the unattainable premise of Eq. (7), distinguishing scalar variance from distributional geometry.

  2. Appendix C.3 provides validation for exact OT, but the method uses entropically regularized transport. Beyond the example in Figure 10, can the authors report the mean, upper quantiles, and maximum cross-modal transport mass across retrieved pairs at (ε=0.03\varepsilon=0.03), to substantiate the claim that cross-modal leakage is negligible in practice?

4 comments

We thank the reviewer for the careful reading and for recognizing the practical, training-free design, which preserves token-level temporal information across variable lengths and channel counts, and the breadth of the evaluation and ablations. We address each point below.


W1: The variance-based motivation for OT in §4.2 conflates loss of distributional geometry with loss of scalar variance. For unit-normalized tokens, Eq. (14) gives σ² = 1 − ‖ē‖²₂..... Eqs. (3)–(4) quantify reconstruction residuals, but do not establish that this scalar energy is unavailable from the mean or requires token-level comparison to recover. Moreover, Eq. (7)'s premise—identical means but different scalar variances—is unattainable under these assumptions...

We would like to thank again for pointing this out, this constructive feedback is essential for the improvements and the correctness of our work, below we address this concern.


Q1: Given Eq. (14), the pooled mean before further normalization exactly determines the scalar variance. How do the authors reconcile this with §4.2's claim that pooling followed by Euclidean comparison discards that variance? Please revise the energy-loss interpretation of Eqs. (3)–(4) and the unattainable premise of Eq. (7), distinguishing scalar variance from distributional geometry.

We agree. For ℓ2-normalized tokens, ‖ē‖² + σ² = 1, so the unnormalized mean fully determines the scalar variance σ², and a Euclidean comparison of means is not blind to it. Pooling does lose the distributional geometry: the covariance Σ and, beyond it, the full token distribution. The given example shows this well. A and B have the same mean and the same σ² = 1/2, but different covariances (spread along e₂ vs. e₃), and W₁(A, B) = 1.

Thanks to the shared comments, we revised the first part of §4.2 around this distinction:

  1. Energy-loss interpretation removed. We dropped the claim that pooling discards a fixed fraction of the energy, i.e., old remark 1 (σ² via the mean pairwise cosine) and the cos 30°/45° floors of old Eq. (4). The pairwise-cosine identity is kept only as a side remark in Appendix C.1.
  2. Covariance instead of scalar variance. §4.2 now defines the mean and covariance of a token set, and uses the identity 1 = ‖ē‖² + σ² with σ² = tr(Σ) only to bound what the mean cannot see. New Remark 1 gives ‖Σ‖_F ≤ σ² = 1 − ‖ē‖². So σ² is not the lost quantity. It caps the covariance structure that pooling discards, and for two sets ‖Σ_A − Σ_B‖_F ≤ σ_A² + σ_B². In particular, Σ = 0 iff σ² = 0, i.e., only near-collinear token sets lose nothing under pooling.
  3. Premise for old Eq. (7). The OT remark now assumes equal means and different covariances (Σ_A* ≠ Σ_B) instead of different scalar variances. This case is attainable, e.g., {u, −u} vs. {v, −v} for orthogonal unit u, v, and the given A, B. Both pairs have equal means and equal σ², yet OT separates them: every cross-set distance in {u, −u} vs. {v, −v} is √2, so W₁ = √2 (and W₁(A, B) = 1). OT's strict positivity there therefore reflects geometry beyond the mean and scalar variance.
  4. Empirical measure. Figure 2 now plots σ² directly, as the cap on the covariance the mean cannot see. For Chronos-2, 96% of token sets have σ² > 0.1, i.e., substantial room for covariance differences that pooling cannot distinguish.

The main claims of Remark 2 and Proposition 1 are unchanged, and so are their proofs in the appendix. Remark 2 still states that W₁ is lower-bounded by the pooled Euclidean distance and is strictly positive whenever the two token distributions differ. Its proof only uses that W₁ is a metric on empirical distributions, so it never relied on the scalar variance. We only restated the example premise of Eq. (7) with covariances, which makes it attainable. Proposition 1 (the dual embedding space, which keeps transport mass within each modality) does not involve the variance at all.


Q2: Appendix C.3 provides validation for exact OT, but the method uses entropically regularized transport. Beyond the example in Figure 10, can the authors report the mean, upper quantiles, and maximum cross-modal transport mass across retrieved pairs at (ε = 0.03), to substantiate the claim that cross-modal leakage is negligible in practice?

Yes. Using the method's own Sinkhorn solver at ε = 0.03, we computed γ_ε for every query's top-8 retrieved demonstrations over all 39 subtasks. We report the cross-modal mass, i.e., the share of γ_ε's total mass that sits on time series↔text entries.

Cross-modal transport mass (% of total mass), top-8 retrieved pairs, ε = 0.03:

BenchmarkPairsMeanp90p99Max
TSExam5,9680.003%0.013%0.025%0.51%
TSR-Bench33,0000.0006%0.0001%0.021%0.05%
Time-MQA48,0000.0002%<0.0001%0.007%0.13%
MTBench110,1840.0003%<0.0001%0.009%0.15%
TimeOmni32,4320.001%0.0007%0.025%0.17%
All229,5840.0005%0.0001%0.020%0.51%
Random pairs (reference)96,0000.058%0.035%1.27%27.2%

Leakage on retrieved pairs is negligible. Even the worst retrieved pair moves 0.51% of the mass across modalities, and the text point keeps at least 99.5% of its mass α on the other sample's text point. Cross-modal entries contribute at most 1% of any retrieved pair's OT score (mean 0.001%). Leakage is larger only on random, dissimilar pairs (max 27%), i.e., pairs that are never retrieved, so the modality-preserving structure of Proposition 1 holds where it affects retrieval. This table and its discussion are now in Appendix C.3 (new Table), with a one-line pointer in §4.4 (TSLM-ICL), where Proposition 1 is introduced.


We thank the reviewer again for the careful and constructive review. It led us to sharpen the motivation in §4.2 around covariance rather than scalar variance, and to quantify the entropic plan's cross-modal leakage on every retrieved pair. We hope these revisions and the new results address the reviewer's concerns. The reviewer rated the paper as informative ("I learned something interesting or useful"), and we believe this reflects the potential of the work. With the soundness concerns now resolved, we kindly ask the reviewer to reconsider their assessment.

Thank you for providing the requested leakage statistics. However, the reported pair counts are inconsistent with the evaluation population specified in the paper. Tables 6–7 report 19,611 samples across 39 subtasks, so retrieving eight demonstrations for every query yields 156,888 directed query–demonstration pairs, rather than 229,584. Specifically, the expected versus reported counts are 24,000 versus 48,000 for Time-MQA, 68,640 versus 110,184 for MTBench, and 25,280 versus 32,432 for TimeOmni. Appendix B.2 also states that retrieval results are shared across all six models.

A related inconsistency already appears in Appendix C.3: it reports approximately 115,000 selected pairs at k=4k=4, whereas 19,611 queries would yield 78,444 pairs.

What exact query populations and counting procedures were used for these two validations? Please provide the query counts per subtask, clarify whether additional samples or repeated records were included, and reconcile the counts with Tables 6–7. If the leakage statistics were computed on a different population, please report them for the documented evaluation set.

We apologize for the confusion and thank the reviewer for engaging in the discussion and for the careful check. The difference in counts comes only from scope. The validations were run on all 39 subtasks we prepared, in addition to 8 subtasks (together 28,698 queries) which are excluded from the evaluation in the paper, as described in Appendix D:

  • the forecasting and imputation subtasks of MTBench, Time-MQA and TimeOmni, where outputs are numeric sequences rather than answers that can be scored by accuracy
  • Time-MQA open-ended QA, whose answers are free-form

These subtasks were omitted because at k=0 their answers are numeric sequences (forecasting, imputation) or free-form text (open-ended QA), which cannot be scored by accuracy, and because the models perform poorly on them, often producing malformed or wrong-length outputs that are truncated by token limits, unrelated to ICL.

These 8 subtasks account exactly for the differences the reviewer identified in Time-MQA, MTBench and TimeOmni, and for the ≈115k pairs at k = 4 in Appendix C.3 (28,698 × 4). For consistency with Tables 6–7, we now report both validations on the 39 evaluated subtasks.

Shared retrieval. Yes. To ensure a fair comparison, all six TSLMs receive exactly the same selected demonstrations for each query, so pairs are counted once per query, not once per model.

Cross-modal transport mass (% of total mass), top-8 retrieved pairs, ε = 0.03, 39 evaluated subtasks (19,611 queries):

BenchmarkPairsMeanp90p99Max
TSExam5,9680.003%0.013%0.025%0.51%
TSR-Bench33,0000.0006%0.0001%0.021%0.05%
Time-MQA24,000<0.0001%<0.0001%<0.0001%0.02%
MTBench68,6400.0004%<0.0001%0.016%0.15%
TimeOmni25,2800.001%0.001%0.026%0.17%
All156,8880.0006%0.0001%0.022%0.51%
Random pairs (reference)78,0000.067%0.042%1.47%27.2%

The conclusions are unchanged. The worst retrieved pair moves 0.51% of the mass across modalities, and cross-modal entries contribute at most 1% of any retrieved pair's OT score. In Appendix C.3, the condition now holds for all 78,444 top-4 pairs, for every top-1 pair, and for 99.99% of top-8 pairs. We will update Table [ot_leakage] and Appendix C.3 accordingly.

Reviewer_Vn5pOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes TSLM-ICL, a training-free framework that combines pretrained time-series and text embeddings with token-level optimal transport to retrieve demonstrations for frozen time-series language models. Evaluation across five benchmarks, six models, and three demonstration budgets shows generally improved accuracy over random, text-only, and existing time-series retrieval baselines.

Critical strengths

The broad evaluation across five benchmarks and six frozen models, supported by component ablations and encoder-matched comparisons, provides useful evidence that the framework’s retrieval benefits extend across multiple tasks and architectures.

Critical weaknesses
  1. Limited methodological and theoretical novelty. The framework combines pretrained encoders, standard optimal transport, and Top-k selection, with little innovation beyond assembling established components. The mathematical analysis mainly restates basic properties of pooling and Wasserstein distance; it motivates the design but does not establish a substantive new theoretical result. Together, these limitations make the contribution highly incremental.

  2. The empirical gains do not sufficiently compensate for the limited novelty. Although the evaluation is broad, average improvements over TS-RAG are only 1.8–2.4 percentage points, while the isolated benefit of OT is particularly small on predominantly univariate tasks. Moreover, the comparisons include pretrained time-series encoders but no retrievers trained specifically for retrieval or demonstration selection. The experiments therefore provide limited evidence that the proposed combination offers a substantial advantage over stronger retrieval approaches.

Questions for authors
  1. What methodological or theoretical novelty does TSLM-ICL establish beyond combining pretrained encoders, standard OT, and Top-k retrieval?
  2. Given the small OT ablation gains, particularly on univariate tasks, what evidence justifies its added computational cost over pooled retrieval?
  3. Why are retrieval-trained methods absent from the comparisons, and what evidence establishes competitiveness with such approaches?
2 comments

We thank the reviewer for the careful reading and for recognizing the breadth of the evaluation across five benchmarks and six frozen models, and the component ablations and encoder-matched comparisons that support it. We address each point below.


W1: Limited methodological and theoretical novelty....

We respectfully disagree, and clarify both contributions below.

Methodological contribution. To the best of our knowledge, no prior work provides a tool for in-context learning with time-series language models, i.e., for selecting demonstrations for a frozen TSLM. Existing time-series retrieval and ICL methods target forecasting and assume fixed lengths, fixed or single variates, or no text. TSLM-ICL is the first training-free demonstration selector for frozen TSLMs that handles variable lengths, variable variate counts and text in one similarity.

Theoretical contribution. Our analysis formalizes why token-level transport is the right comparison for TSLM demonstration retrieval, to the best of our knowledge, no time-series work formalizes this issue and provilde a suitable solution. The key result is Remark 2 (§4.2):

  • What it shows: OT never loses the information that mean pooling keeps. It also separates samples that pooling cannot tell apart, i.e., token sets with the same mean but a different covariance, such as {u, −u} vs. {v, −v} (W₁ = √2, pooled distance 0).

  • Extension to the multimodal setting (Proposition 1, App. C.3): with time-series and text tokens in disjoint subspaces and non-uniform masses, one joint OT still satisfies

$ W₁ ≥ α‖ē^{text,*} − ē^{text}‖ + (1−α)‖ē^{ts,*} − ē^{ts}‖ $

without mixing modalities. Proposition 1 is, to our knowledge, a new formulation. It addresses exactly our problem: comparing time series and text in a single transport, rather than scoring each modality separately and fusing the scores. It guarantees that this joint comparison keeps the per-modality pooled information while transport mass stays within each modality. We placed it in the appendix for space and because it extends Remark 2, but it is what makes the ts-multimodal method well-founded.


W2: The empirical gains do not sufficiently compensate for the limited novelty. Although the evaluation is broad, average improvements over TS-RAG are only 1.8–2.4 percentage points, while the isolated benefit of OT is particularly small on predominantly univariate tasks. Moreover, the comparisons include pretrained time-series encoders but no retrievers trained specifically for retrieval or demonstration selection. The experiments therefore provide limited evidence that the proposed combination offers a substantial advantage over stronger retrieval approaches.

Consistency of the gains. The gains are averages over 39 subtasks and six TSLMs, and they are consistent: TSLM-ICL is best in 14 of 15 dataset–budget settings.

Stronger retrievers. Trained time-series retrievers are not designed for TSLMs. They mostly target forecasting and come with restrictions: fixed input shapes or lookback windows, univariate or fixed variate counts, no text, or no selection of demonstrations across samples. Nevertheless, following this recommendation and reviewer qJBR's, we adapted TRACE (NeurIPS 2025), a retriever trained for multimodal time-series retrieval, to our benchmarks during the rebuttal and fine-tuned it on them. TSLM-ICL beats it at every k, by 1.9 points on average, without any training. The results are in Q3 and in the main table of the revision.

OT on univariate tasks: see Q2.


Q1: What methodological or theoretical novelty does TSLM-ICL establish beyond combining pretrained encoders, standard OT, and Top-k retrieval?

See W1. Theoretically, Remark 2 shows that OT over token sets keeps all the information of mean pooling and also separates samples whose token covariances differ, which pooling cannot. Proposition 1 extends this to the joint time series–text token set without mixing modalities. Methodologically, TSLM-ICL is the first training-free demonstration selector for frozen TSLMs over variable-length, variable-variate multimodal inputs.


Q2: Given the small OT ablation gains, particularly on univariate tasks, what evidence justifies its added computational cost over pooled retrieval?

The aggregate univariate gain is small mainly because OT and pooling often retrieve the same demonstrations, which dilutes the average. As shown in App. "Where OT and Pooling Retrieve Differently": whenever the selections differ, OT's demonstrations give higher accuracy. This holds for univariate and multivariate subtasks, at every k, against pooled cosine and Euclidean, and for all six TSLMs, and the advantage grows with the disagreement.

On univariate subtasks alone, among queries where OT and pooling select completely different demonstrations, OT's accuracy is higher than pooling's:

Univariate, all demonstrations differk=1k=4k=8
vs. pooled cosine+1.4+4.3+0.8
vs. pooled Euclidean+1.2+3.6+2.5
# (model, query) pairs, cosine / Euclidean36.2k / 37.2k2.1k / 2.2k356 / 364

(Accuracy points; 6 TSLMs.)


Q3: Why are retrieval-trained methods absent from the comparisons, and what evidence establishes competitiveness with such approaches?

We initially compared against training-free retrievers, as TSLM-ICL itself is training-free and requires no training data from the target benchmarks. Trained time-series retrievers mostly target forecasting and assume fixed lengths or variate counts, so they require adaptation to our setting. Following this comment, during the rebuttal we added TRACE (Chen et al., NeurIPS 2025), a multimodal time series–text retriever trained for retrieval. We adapted it to variable lengths and variate counts, and fine-tuned it on our benchmarks with its own objective. Each benchmark's model was trained only on the other four, to avoid leakage.

Accuracy averaged over all completed subtasks (6 TSLMs). We are still waiting for a few runs (MTBench at k=8, one model); all other TRACE results are final, so these numbers are close to final:

Methodk=1k=4k=8Avg.
TSLM-ICL (ours, training-free)0.4880.5330.5420.519
TRACE (fine-tuned)0.4710.5140.5210.500

TSLM-ICL is best at every k, by 1.9 points on average, without any training. It is also ahead on every dataset and k completed so far, including MTBench at k=1 and k=4. TRACE is now in Tables 1–2 and in Appendix "Retrieval Methods".


We thank the reviewer again for the constructive review, which led us to add the TRACE baseline, the univariate OT analysis, and a clearer statement of Proposition 1. We hope these address the reviewer's concerns and kindly ask the reviewer to reconsider their assessment.

Reviewer_qJBROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposed TSLM-ICL: it represents samples using token-level temporal and textual embeddings, compares variable-sized representations via optimal transport, and selects relevant demonstrations under a fixed context budget. The improvements seem promising, with accuracy by up to 16% over zero-shot prompting.

Critical strengths

S1: The paper is well-positioned that this is the only ICL TSFM considered both time series and text (although some important reference seems missing, see the W1)

S2: The paper is well written, and the appendix is well prepared.

S3: Simple idea with decent mathematical discussions.

Critical weaknesses

W1: some important references are missing, e.g., [R1] Jiecheng Lu, Yan Sun, Shihao Yang, "In-context Time Series Predictor", ICLR 2025 [R2] Shangqing Xu, Harshavardhan Kamarthi, Haoxin Liu, B. Aditya Prakash, "In-context Pre-trained Time-Series Foundation Models adapt to Unseen Tasks" [R3] Etienne Le Naour, Tahar Nabil, Adrien Petralia, "TS-ICL: A Flexible Time-Indexed Foundation Model for Time Series via In-Context Learning"

W2: Related to this, the main SOTA baselines are actually only TimeRAG (first released in Dec 2024) and TS-RAG (first released in Mar 2025), both are 2 years ago / 1.5 years old baseline. This weakens the real performance gain of this paper over the uncompared SOTA models.

W3: In Eq. [4], the module of optimal transport comparison, the floor / thres "0.1340" or "0.2929" comes from cos 30◦ or cos 45◦ solely based on the empirical results of Chronos-2 only. This is probably a critical limitation: why has to be Chrones to be the "rule"? how generaliazble the result will be? what if we use another time-series foundation model to set the empirical thres?

Questions for authors

Q1: See W2. Q2: In Table 2, there are five benchmarks, and each with different tasks. Which tasks are actually being evaluated? Q3: In Table 3, "w/o text" does not drop the performance that much. So it seems like it's not that necessary to add the text into TS-ICL? In such a case, the claimed "novelty" of being the first one to consider text is weakened.

3 comments

We thank the reviewer for the careful reading and for recognizing the positioning of the work, the quality of the writing and appendix, and the simplicity of the approach. We address each point below.


W1: some important references are missing...

Thanks for pointing these works out, we will add them to the related work. Moreover, we added TS-ICL to some of the experiments as shown in figures 2 and 4.


W2: Related to this, the main SOTA baselines are actually only TimeRAG (first released in Dec 2024) and TS-RAG (first released in Mar 2025), both are 2 years ago / 1.5 years old baseline. This weakens the real performance gain of this paper over the uncompared SOTA models.

Time series language models (TSLMs) are a young and growing field, and unfortunately there no relevant direct works, only adaptations from time series forecasting. TimeRAG and TS-RAG are among the few training-free alternatives to TSLM-ICL that were found relevant. Even they required adaptation to our benchmarks as discussed in (Appendix "Retrieval Methods").

Moreover, The works cited (W1) use a different problem formulation. They are models that consume in-context examples, or they are typically restricted to fixed variate counts or fixed lookback windows. None selects demonstrations across samples for a frozen TSLM. TSLM-ICL lifts these restrictions, which is why directly comparable methods are scarce.

Nevertheless, we added a recent, retrieval-trained baseline: TRACE (Chen et al., NeurIPS 2025), a multimodal time series retriever. We adapted its fixed input size to variable lengths and channel counts, used both series and text, and fine-tuned it on our benchmarks. To avoid leakage, each benchmark's model was trained only on the other four.

Accuracy averaged over all completed subtasks (6 TSLMs). We are still waiting for a few runs (MTBench at k=8, one model); all other TRACE results are final, so these numbers are close to final:

Methodk=1k=4k=8Avg.
TSLM-ICL (ours, training-free)0.4880.5330.5420.519
TRACE (fine-tuned)0.4710.5140.5210.500

TSLM-ICL is best at every k, 1.9 points above TRACE on average, without any training. It is also ahead on every dataset and k completed so far, including MTBench at k=1 and k=4. TRACE is now in Tables 1–2.


W3: In Eq. [4], the module of optimal transport comparison, the floor / thres "0.1340" or "0.2929" comes from cos 30◦ or cos 45◦ solely based on the empirical results of Chronos-2 only...

We made two changes in the revision. First, we added a plot for each prominent multivariate time series foundation model: TimesFM-3, TiRex-2 and TS-ICL (the work cited as [R3]), alongside Chronos-2.

The finding holds for every encoder: most token sets are far from collinear, and each model is at least as dispersed as Chronos-2. Comparing token sets rather than their means is therefore not specific to Chronos-2.

Encoderc̄ < cos 30°c̄ < cos 45°
Chronos-2 (ours)95%64%
TimesFM-399%95%
TiRex-292%85%
TS-ICL [R3]100%100%

Following reviewer nCdw, Section 4.2 now measures dispersion directly as the total variance σ² = 1 − ‖ē‖², which bounds the covariance Σ that pooling discards (Remark 1). The cos 30°/45° discussion moved to Appendix C.1.


Q2: In Table 2, there are five benchmarks, and each with different tasks. Which tasks are actually being evaluated?

Appendix D.3 details every task. In short, we evaluate all accuracy-scored subtasks of the five benchmarks: 39 subtasks with 19,611 samples. These range from perception (anomaly detection, pattern and noise recognition, similarity) through multi-step reasoning (abductive, causal, deductive, inductive, etiological, numerical, temporal-relation) to decision making and context-grounded prediction (trading decisions, news-conditioned stock trends, scenario understanding). Each subtask is its own leave-one-out retrieval pool and reporting unit.


Q3: In Table 3, "w/o text" does not drop the performance that much. So it seems like it's not that necessary to add the text into TS-ICL?...

Benchmarks and tasks differ widely in how much task-relevant text they carry. It is true that, on average, removing text changes performance only moderately. The effect, however, varies across tasks, and with 39 subtasks the text-relevant ones are diluted in the average. Here are 10 tasks where the text component is relevant.

Accuracy gain from text (points), averaged over 6 TSLMs and k ∈ {1, 4, 8}:

SubtaskGain
TSR-Bench quantitative decision+23.0
TSExam noise understanding+5.4
MTBench finance trend (long)+4.3
TSR-Bench noise understanding+4.2
TSR-Bench event prediction+3.1
TSR-Bench inductive reasoning+3.0
TSR-Bench forecasting (MCQ)+2.1
TSExam pattern recognition+1.9
TSExam causality analysis+1.7
MTBench weather trend (future)+1.2

We will add this per-subtask table to the appendix.


We thank the reviewer again for the constructive review, which led us to add the TRACE baseline, the missing references, the multi-encoder dispersion analysis, and the per-subtask text analysis. We would be happy to address any further concerns. If none remain, and given that the reviewer found the paper informative, we kindly ask the reviewer to reconsider their score.

Dear Reviewer qJBR,

With the discussion period nearing its end, we hope that our responses have adequately addressed all of your concerns and clarified the points raised in your review. If any concerns remain or further clarification would be helpful, we would be happy to address them.

Thank you again for your time and constructive feedback. We sincerely appreciate your consideration of our work.

Comments on the paper

1 comment

Revisions and rebuttal answers summary

  • New retrieval-trained baseline, TRACE (NeurIPS 2025), in Tables 1–2 and Appendix "Retrieval Methods", to compare against a recent, trained retriever beyond TimeRAG/TS-RAG (qJBR, Vn5p).
  • TS-ICL [R3] and other time series foundation models (TimesFM-3, TiRex-2) in the dispersion analysis (Fig. 2, App. C.1), and TS-ICL in the encoder ablation (Fig. 4), to show that the token-set motivation is not specific to Chronos-2 (qJBR).
  • Revised §4.2 motivation for OT: pooling discards the covariance Σ (bounded by σ² = 1 − ‖ē‖², Remark 1), not the scalar variance. The energy-loss interpretation was removed and the premise of Remark 2 restated with covariances (nCdw).
  • Cross-modal leakage of the entropic plan (mean, quantiles, max over all retrieved pairs at ε = 0.03) added to App. C.3: at most 0.51% of the mass (nCdw).
  • Per-query OT vs. pooling numbers on univariate tasks: when selections differ, OT is more accurate at every k (Vn5p).
  • New appendix, "Where Text Matters", with a per-subtask, per-model table of the gain from text, showing that a small average drop hides large task-specific gains (qJBR).
  • Missing references (ICTSP, ICTP, TS-ICL, TRACE) added to the related work (qJBR).

Update (18/11/2026 6:24 UTC) - new revision with the full results and corrections from the rebuttal was uploaded.

All changes are highlighted in red in the revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.