acceptodds
Under review as a conference paper at ICLR 2027

Text-Attention Guided In-Context Forecasting for Multimodal Time Series Prediction

Abstract

Real-world time series forecasting often contains both numerical histories and textual context, such as news, reports, or event descriptions. Existing multimodal forecasting methods typically incorporate text through LLM-based prompting or representation-level fusion, which presents a modelling challenge when textual signals are heterogeneous, noisy and unevenly informative. We propose TAG-ICF, a text-attention guided in-context learning (ICL) forecasting framework that uses a frozen, pretrained ICL backbone for numerical prediction while using text to modulate attention over historical examples. Textual inputs are analyzed using LLMs and embedded as lagged features to compute query-context relevance scores, serving as an attention gate that prioritizes contextually relevant historical patterns. TAG-ICF consistently outperforms all multimodal baselines across twelve datasets, yielding superior average ranks of 1.58 and 1.58 in MSE and MAE, respectively. Our framework surpasses numerical-only ICL on nearly all datasets and delivers up to 15.57% MSE improvement relative to the strongest competing methods in diverse settings.

Primary area:
learning on time series and dynamical systems
Keywords:
In-context Learning, Tabular Foundation Model, Time Series Forecasting, Multimodal Prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ahkVOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses how to use noisy textual context, such as news and reports, to improve time series forecasting. Its proposed method, TAG-ICF, uses text to guide attention toward relevant historical examples. The numerical forecasting backbone stays frozen, while a small text-guided attention module is trained. The key insight is that text can help determine which past observations are useful for the current forecast. Experiments show strong overall performance, with text improving the same numerical backbone. However, the benefits depend on text quality, temporal alignment, and the underlying forecasting model.

Critical strengths

TAG-ICF uses text-derived similarity to modulate query–context attention within a frozen numerical ICL backbone. This provides a well-motivated separation of roles: text guides the weighting of historical examples, while the pretrained backbone handles numerical prediction.

Critical weaknesses
  • The contribution appears incremental, combining text-based relevance weighting with a frozen ICL backbone. The paper offers limited new insight into when or why this combination is effective.
  • The experiments do not compare different ways of using text while keeping the numerical backbone and text inputs the same. This leaves it unclear whether TAG’s attention mechanism offers an advantage over simpler ways of incorporating text.
Questions for authors
  • Were different ways of incorporating text evaluated using the same numerical backbone and processed text? If so, how do the results isolate the benefit of TAG’s attention mechanism?
  • What does TAG’s attention-level formulation enable beyond existing approaches that use text to select or weight historical examples? Which findings support this distinction?
3 comments

We thank the reviewer for taking the time to review our paper, and are glad that they found the text guidance in TAG-ICF well motivated.

W1

Novelty and mechanistic insight

We thank the reviewer for the comment. The core technical contribution of TAG-ICF is fusion through a text-derived attention guidance mechanism, which uses textual information to reweight historical numerical contexts rather than simply fusing text with numerical inputs. This also differs from approaches that bring text into their frozen or fine-tuned backbones through representation-level fusion or LLM-based prompting.

On why the combination is effective, Figure 1 gives a qualitative example from one of the benchmark datasets. Text guidance changes the backbone’s attention over the historical examples, which corrects the forecast direction. This internal attention change is consistent with the intuition behind the method.

On when it is effective, our experiments cover twelve datasets from diverse domains, where text improves over the same backbone without text on 11 of 12 (Tables 2-3). We also provide ablation studies on the ICL backbone, the text embedding model, text preprocessing, text alignment, etc. (Sections 4.3-4.4, Appendix D), which show that the gain carries over to another backbone and other embedding models, and that it comes from the semantic content of the text. We will add these results to the revision.

Q1/W2

Comparison with other fusion on the same backbone

We thank the reviewer for raising this point. Early and late fusion on the same backbone were our starting point. In early-stage experiments, neither could reliably extract a useful signal from the text, and this observation is what motivated similarity-based attention adjustment. We have now run both variants systematically with hyperparameter optimization, using the same TabDPT backbone, inputs, and preprocessing as TAG-ICF in the paper:

  • Early fusion: a trained MLP projects the text embeddings into K numerical features, which are concatenated with the time series features as input to the frozen TabDPT.
  • Late fusion: the frozen TabDPT's hidden states serve as the numerical representation, and a trained MLP fuses them with the projected text embeddings.
Dataset (MSE / MAE)No TextEarly fusionLate fusionTAG-ICF
Economy0.576 / 0.6040.841 / 0.7230.629 / 0.6310.537 / 0.578
Environment0.337 / 0.4000.354 / 0.4170.384 / 0.4160.326 / 0.391
Security8.491 / 2.51710.157 / 2.7527.720 / 2.3517.681 / 2.343
Social Good0.131 / 0.2620.138 / 0.2740.195 / 0.3320.108 / 0.242
Electricity0.708 / 0.6300.786 / 0.6680.718 / 0.6310.646 / 0.602

Out of the three text-integration mechanisms, our proposed attention guidance is the most effective: TAG-ICF achieves the lowest MSE and MAE on all five datasets. Neither early nor late fusion consistently turns the same GPT-5.2-processed text into forecasting gains. Early fusion is worse than the no-text baseline on every dataset and both metrics. Late fusion improves over the baseline only on Security, which is also the only dataset where it comes close to TAG-ICF. We will add these results in the revision.

Q2

Attention-level fusion vs. text-based selection/weighting

To our knowledge, TAG-ICF is the first method to fuse text by directly updating an ICL backbone’s query–context attention scores. The closest prior work is context retrieval and selection for ICL models, which acts as a preprocessing step that decides which examples enter the context [1, 2] rather than updating the attention scores. We also believe text-based context selection is not ideal for forecasting, because it breaks the temporal continuity of the context. TAG-ICF instead keeps the full contiguous context and reweights it softly inside the model.

The experiment below supports this argument. We compare against a text-semantic retrieval baseline that selects the top-K context examples, using the same backbone and text embeddings. Simple text-based retrieval does not help forecasting: for both values of K, it is worse than the numerical-only model with full context on all five datasets and both metrics, and TAG-ICF performs much better.

Dataset (MSE / MAE)Text retrieval, top 50Text retrieval, top 100No text (full context)TAG-ICF
Economy0.842 / 0.6930.671 / 0.6600.576 / 0.6040.537 / 0.578
Environment0.439 / 0.4750.376 / 0.4330.337 / 0.4000.326 / 0.391
Security20.038 / 3.36613.545 / 2.9898.491 / 2.5177.681 / 2.343
Social Good0.182 / 0.2930.143 / 0.2730.131 / 0.2620.108 / 0.242
Electricity1.035 / 0.7820.879 / 0.7060.708 / 0.6300.646 / 0.602

In addition, ConTextTab, one of our baselines, feeds text embeddings into an ICL model as input features, so its attention scores combine textual and numerical information implicitly. It shows little or no gain from text on most datasets, whereas TAG-ICF improves on nearly all of them (Tables 2-3). We will include the above discussion and results in the revision.

[1] Thomas et al. Retrieval & Fine-Tuning for In-Context Tabular Models. NeurIPS 2024.

[2] Xu et al. Mixture of In-Context Prompters for Tabular PFNs. ICLR 2025.

Dear Reviewer:

We hope this message finds you well, and thank you again for your initial review. As the discussion deadline approaches, we would appreciate your feedback on our updated responses and additional results.

If our response has addressed your concerns, we kindly ask you to consider raising your score. If any remain, we are happy to discuss which specific criteria still need to be met to support an increase.

Thank you again for your time and consideration.

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

This paper proposes TAG-ICF, a multimodal time series forecasting framework that keeps a pretrained tabular ICL backbone frozen and uses LLM processed text embeddings to reweight its attention over historical examples through a learned gate. Experiments on twelve datasets show that it achieves the best average rank among prompting and fusion based multimodal baselines.

Critical strengths
  • The method trains only a text projection and attention gates on top of a frozen backbone, which preserves the numerical inductive bias of the pretrained model and suits low data settings.

  • The experiments cover twelve datasets from diverse domains and seven multimodal baselines. They also include ablations on text shuffling, text dropping, embedding models and LLM preprocessing.

Critical weaknesses
  • Unclear source of the performance gains: The gains cannot be clearly attributed to the proposed guidance mechanism. As shown in Tables 4 and 15, on Environment, Social Good and Security most or even all of the improvement comes from the GPT-5.2 preprocessing, and with raw text TAG-ICF even underperforms the text-free model on some metrics. Such LLM-based preprocessing is already common in prior work such as TimeCAP and TiMi. On Economy and Electricity (MSE), where raw text retains most of the gain, the shuffled text variant performs nearly as well, suggesting the improvement is largely unrelated to text semantics.

  • Limited technical novelty: The TAG module mixes the numerical attention with a text similarity attention through a learned scalar gate per layer and head, which essentially adds a text retrieval head to a frozen TabDPT. Moreover, the gate is input independent and thus cannot adapt to how informative the text is, which contradicts the paper's own motivation that textual signals are unevenly informative.

  • Missing comparison with fusion on the same backbone: The paper claims that guidance is preferable to fusion but never compares the two on the same backbone, for example by feeding the same text embeddings into TabDPT as additional features or fusing them with the numerical representations through concatenation or cross attention.

Questions for authors

See weaknesses.

3 comments

We thank the reviewer for taking the time to review our paper, and are glad that they found our TAG-ICF architecture well suited to low-data settings and our experiments and ablations comprehensive.

W1

Unclear source of the performance gains

We thank the reviewer for closely examining our ablations. We first clarify that no hyperparameter optimization (HPO) was performed for the raw-text ablation in the original paper, which led to suboptimal results on some datasets. We have now conducted HPO for raw text as well, and the results are shown in the table below:

Dataset (MSE / MAE)No TextRaw text (w/o HPO)Raw text (HPO)ShuffledTAG-ICF
Economy0.576 / 0.6040.547 / 0.5810.548 / 0.5770.549 / 0.5820.537 / 0.578
Environment0.337 / 0.4000.333 / 0.3980.330 / 0.3950.335 / 0.3980.326 / 0.391
Security8.491 / 2.5178.635 / 2.4477.486 / 2.2128.230 / 2.4157.681 / 2.343
Social Good0.131 / 0.2620.129 / 0.2770.091 / 0.2390.131 / 0.2830.108 / 0.242
Electricity0.708 / 0.6300.669 / 0.6130.659 / 0.6090.676 / 0.6210.646 / 0.602

We agree that on the Economy dataset, the shuffled and tuned raw-text results taken together suggest the improvement may be structural, coming from the trained component and not from text semantics. However, on the other four datasets, tuned raw text improves over both the no-text model and the shuffled variant on both metrics.

For the Environment dataset. TAG-ICF’s improvement on this dataset is already modest in the original paper (3.3% MSE); with tuned raw text it is 2.1%. We include this as an example dataset with weakly informative text, where similar and modest gains from both raw and preprocessed text are expected.

For Social goods and Security, after tuning, TAG-ICF with raw text improves over the no-text model on both datasets, and performs even better than with GPT-5.2 preprocessed text. This further shows that the gain does not come from GPT-5.2 preprocessing, which brings no extra benefit and can even hurt performance on these datasets.

For Electricity, tuned raw text clearly improves over both the no-text model and the shuffled variant on both MSE and MAE, suggesting that text semantics contribute to the gains. In addition, GPT-5.2 preprocessed text yields a further modest improvement over tuned raw text, reducing MSE from 0.659 to 0.646 and MAE from 0.609 to 0.602. It is reasonable that GPT-5.2 preprocessing helps on some datasets, such as this one, and not on others, such as Security and Social Good.

We will update the results and clarify the points above in the revised paper.

W2

Limited technical novelty

We thank the reviewer for these comments. To our knowledge, TAG-ICF is the first to fuse text by directly updating the internal query-context attention scores of a frozen ICL backbone, which differs from the fusion methods discussed in our related work. This design is also supported by [1], which shows that these scores are central to the performance of tabular ICL models. The fusion does carry a sense of similarity-based retrieval, but we see this as a strength: it makes the design intuitive and aligned with real-world forecasting, where one looks to past situations that resemble the current one. The closest line of work is retrieval for tabular ICL models [2, 3], which acts as preprocessing outside the model, selecting which examples enter the context. In contrast, TAG-ICF uses a learnable textual projection and is softer and more fine-grained, with gates learned per layer and head. Compared with simple direct context retrieval, it also keeps the full contiguous context, preserving temporal continuity.

To address the reviewer’s comment that the gate is input independent: although the gate is static, the textual attention module itself can account for uninformative text through the way the textual attention is computed for each query. When the text for a query is uninformative, only weakly predictive, or unrelated to the target, its similarity to all or most historical examples in the context is expected to be about the same. The textual attention is then close to uniform, so the main group of ICL examples used for the forecast stays unchanged, with at most slight shifts in individual attention scores. In contrast, fusion-based methods that extract features from text can pick up spurious signals from noisy text and distort predictions. The design is therefore consistent with our motivation, even without a dynamic gate.

In this work we use a simple and interpretable convex gate to show that text can guide attention effectively. We agree that input-dependent (instance-aware) gating could improve this further. We already note this as future work in Sec. 3.3 and Sec. 5, and we will state it more clearly in the revision.

W3

Missing comparison with fusion on the same backbone

We thank the reviewer for raising this point. Early and late fusion on the same backbone were our starting point. In early-stage experiments, neither could reliably extract a useful signal from the text, and this observation is what motivated similarity-based attention adjustment. We have now run both variants systematically with hyperparameter optimization, using the same TabDPT backbone, inputs, and preprocessing as TAG-ICF in the paper:

  • Early fusion: a trained MLP projects the text embeddings into K numerical features, which are concatenated with the time series features as input to the frozen TabDPT.

  • Late fusion: the frozen TabDPT's hidden states serve as the numerical representation, and a trained MLP fuses them with the projected text embeddings.

Dataset (MSE / MAE)No TextEarly fusionLate fusionTAG-ICF
Economy0.576 / 0.6040.841 / 0.7230.629 / 0.6310.537 / 0.578
Environment0.337 / 0.4000.354 / 0.4170.384 / 0.4160.326 / 0.391
Security8.491 / 2.51710.157 / 2.7527.720 / 2.3517.681 / 2.343
Social Good0.131 / 0.2620.138 / 0.2740.195 / 0.3320.108 / 0.242
Electricity0.708 / 0.6300.786 / 0.6680.718 / 0.6310.646 / 0.602

Out of the three text-integration mechanisms, our proposed attention guidance is the most effective: TAG-ICF achieves the lowest MSE and MAE on all five datasets. Neither early nor late fusion consistently turns the same GPT-5.2-processed text into forecasting gains. Early fusion is worse than the no-text baseline on every dataset and both metrics. Late fusion improves over the baseline only on Security, which is also the only dataset where it comes close to TAG-ICF. We will add these results in the revision.

[1] Rubachev et al. On Finetuning Tabular Foundation Models. arXiv:2506.08982, 2025.

[2] Thomas et al. Retrieval & Fine-Tuning for In-Context Tabular Models. NeurIPS 2024.

[3]Xu et al. Mixture of In-Context Prompters for Tabular PFNs. ICLR 2025.

Dear Reviewer:

We hope this message finds you well, and thank you again for your initial review. As the discussion deadline approaches, we would appreciate your feedback on our updated responses and additional results.

If our response has addressed your concerns, we kindly ask you to consider raising your score. If any remain, we are happy to discuss which specific criteria still need to be met to support an increase.

Thank you again for your time and consideration.

Reviewer_MuJ9OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes TAG-ICF (Text-Attention Guided In-Context Forecasting) to better utilize textual context in multimodal time series forecasting. The method uses a frozen pretrained tabular foundation model (mainly TabDPT) for numerical forecasting, while textual information guides the attention over historical examples.

Critical strengths
  • The paper presents a clear and intuitive idea: instead of directly using text for numerical prediction, textual information can help identify historical examples that are more relevant to the current forecasting context.
  • The experiments cover 12 datasets across diverse domains, including economy, agriculture, energy, health, traffic, and finance.
  • The method keeps the pretrained tabular foundation model frozen and only trains the text projection and attention gating parameters, making the adaptation relatively lightweight.
Critical weaknesses
  • The core mechanism combines text-based attention and numerical attention through convex interpolation. This is essentially a similarity-based weighting of historical examples. It is unclear whether the improvements come from the proposed attention-level integration or simply from retrieving semantically similar historical examples. Comparisons with simpler text-based retrieval methods would help clarify the contribution.

  • Although TAG-ICF achieves the best average rank, stronger numerical-only foundation models still outperform it on some datasets. This raises questions about the effectiveness of text guidance compared with simply using a stronger numerical backbone.

  • One motivation of the paper is that real-world textual information can be heterogeneous, noisy, and unevenly informative. However, the attention gating parameters are static after training and do not adapt to the quality or relevance of the current textual input. This may limit the method when text provides little useful information.

  • TAG-ICF relies on a relatively heavy textual processing pipeline, using GPT-5.2 to analyze raw articles and Qwen3-Embedding-8B to generate text embeddings. It is unclear whether the improvements mainly come from better semantic matching or from useful directional predictions already generated by GPT-5.2 during preprocessing. Further analysis would help distinguish these effects.

Questions for authors

Please see weaknesses

4 comments

We thank the reviewer for taking the time to review our paper and for recognizing the clarity and intuitiveness of our idea, the breadth of our experiments, and the lightweight nature of our method.

W1

Attention-level integration vs. simple text-based retrieval

We thank the reviewer for this question. TAG-ICF differs from text-based retrieval in three ways, which together lead to better performance, as the experiment below shows. First, retrieval uses fixed cosine similarity, whereas TAG-ICF learns a projection Wtext with the forecasting loss, so similarity reflects forecasting relevance. Second, retrieval breaks the temporal continuity of the context, whereas TAG-ICF keeps the full contiguous context and only adjusts the attention scores over it. Third, TAG-ICF is softer and more fine-grained than simple retrieval: it retains a (1-) share of numerical attention, with gates learned per layer and head.

We conducted a retrieval experiment in which the context consists of the top-K examples ranked by cosine similarity between the text embeddings of the query and of each context example (H=6), using the same TabDPT backbone, preprocessing, and numerical and textual inputs as TAG-ICF in the submitted paper:

Dataset (MSE / MAE)Text retrieval, top 50Text retrieval, top 100No text (full context)TAG-ICF
Economy0.842 / 0.6930.671 / 0.6600.576 / 0.6040.537 / 0.578
Environment0.439 / 0.4750.376 / 0.4330.337 / 0.4000.326 / 0.391
Security20.038 / 3.36613.545 / 2.9898.491 / 2.5177.681 / 2.343
Social Good0.182 / 0.2930.143 / 0.2730.131 / 0.2620.108 / 0.242
Electricity1.035 / 0.7820.879 / 0.7060.708 / 0.6300.646 / 0.602

As the results above show, simple text-based retrieval does not help forecasting: for both values of K it is even worse than the numerical-only model with full context on all five datasets and both metrics. This supports our arguments above. We will add this comparison, together with the discussion of how TAG-ICF differs from text-based retrieval, to the revision.

W2

Strong numerical-only foundation models

We agree that TAG-ICF does not beat Chronos-2 on every dataset, but the remaining gaps come from the numerical backbone, not the text guidance. Against the same backbone without text, TAG improves MSE on 11 of 12 datasets. On the four datasets where Chronos-2 has lower MSE, frozen TabDPT is already behind before text is added, and text guidance narrows the gap each time. We chose TabDPT for its balance of computational efficiency and performance. With TabPFNv3 as an alternative backbone, TAG improves MSE on 4 of 5 datasets. On Social Good, where TAG-ICF with TabDPT falls well behind Chronos-2 in Table 1, TAG-ICF with TabPFNv3 outperforms it (0.051 / 0.158 vs. 0.064 / 0.171 in MSE / MAE). This is evidence that as ICL backbones keep improving, which is happening quickly, TAG-ICF can be applied on top of them directly and improve further.

More broadly, Chronos-2 has a better overall average rank than every existing multimodal baseline on these benchmarks, which shows how hard it is to turn real-world text into gains over a strong numerical model. TAG-ICF is the only multimodal method with a better overall rank than Chronos-2. We will expand this discussion in the revision.

W3

Static gating

We thank the reviewer for this point. Although the gate is static, our attention module can account for unevenly informative text, where the text of some data points is uninformative, through the way the textual attention is computed for each query: it matches text by similarity and does not extract signals from it. When the text for a data point carries no predictive signal, or only a weak one, or is unrelated to the target, this similarity to all or most historical examples in the context is expected to be about the same. The textual attention is then close to neutral (near-uniform), which leaves the ranking of the context examples unchanged, with at most slight shifts in individual attention scores.

On the other hand, fusion-based methods that extract features from text may be more vulnerable to uninformative or noisy text as spurious signals extracted from such text can distort predictions. This helps to explain why, in Tables 2 and 3, adding text hurts the performance of several other multimodal baselines on a number of datasets, while TAG-ICF benefits from text on nearly all of them. In this work we use a simple and interpretable convex gate to show that text can guide attention effectively. As noted in the paper (Sec. 3.3 and Sec. 5), we agree that input-dependent (instance-aware) gating could further push the boundary of this approach, and we leave it for future work. We will make this clearer in the revision.

W4: Do the gains come from semantic matching or from GPT-5.2's preprocessing?

We thank the reviewer for this question and address it with two experiments.

(1) TAG-ICF with raw text. The raw-text results in the paper's appendix (Table 15) already show gains over the no-text model on most datasets, but they were obtained without separate hyperparameter tuning and were therefore suboptimal. For a fair comparison, we conducted dedicated hyperparameter tuning for raw text:

Dataset (MSE / MAE)No TextRaw text (HPO)TAG-ICF (GPT-5.2 preprocessed text)
Economy0.576 / 0.6040.548 / 0.5770.537 / 0.578
Environment0.337 / 0.4000.330 / 0.3950.326 / 0.391
Security8.491 / 2.5177.486 / 2.2127.681 / 2.343
Social Good0.131 / 0.2620.091 / 0.2390.108 / 0.242
Electricity0.708 / 0.6300.659 / 0.6090.646 / 0.602

Based on the results above, raw text improves over the no-text model on all five datasets in both MSE and MAE. On Security and Social Good, raw text even outperforms GPT-5.2 preprocessed text on both metrics, so preprocessing brings no benefit there and can even hurt in some cases. On the other three datasets, preprocessing adds at most a small further gain. Overall, most of TAG-ICF's improvement is retained without any GPT-5.2 preprocessing, so the gains are not driven by GPT-5.2's directional predictions.

(2) Same text, different fusion. GPT-5.2 preprocessed text is still just free-form text, and whether it translates into forecasting gains still depends on the downstream model's ability to use it. We therefore compared TAG-ICF against early- and late-fusion variants on the same TabDPT backbone, using identical GPT-5.2 preprocessed text:

  • Early fusion: a trained MLP projects the text embeddings into K numerical features, which are concatenated with the time series features as input to the frozen TabDPT.
  • Late fusion: the frozen TabDPT's hidden states serve as the numerical representation, and a trained MLP fuses them with the projected text embeddings.
Dataset (MSE / MAE)No TextEarly fusionLate fusionTAG-ICF
Economy0.576 / 0.6040.841 / 0.7230.629 / 0.6310.537 / 0.578
Environment0.337 / 0.4000.354 / 0.4170.384 / 0.4160.326 / 0.391
Security8.491 / 2.51710.157 / 2.7527.720 / 2.3517.681 / 2.343
Social Good0.131 / 0.2620.138 / 0.2740.195 / 0.3320.108 / 0.242
Electricity NSW0.708 / 0.6300.786 / 0.6680.718 / 0.6310.646 / 0.602

Neither early nor late fusion turns the same GPT-5.2 preprocessed text into consistent forecasting gains. Early fusion is worse than the no-text model on both metrics across all five datasets, and late fusion improves over no text only on Security while being worse on both metrics on the other four. TAG-ICF achieves the lowest MSE and MAE on every dataset, suggesting that the fusion mechanism, rather than text preprocessing, is important for effective multimodal forecasting.

Dear Reviewer:

We hope this message finds you well, and thank you again for your initial review. As the discussion deadline approaches, we would appreciate your feedback on our updated responses and additional results.

If our response has addressed your concerns, we kindly ask you to consider raising your score. If any remain, we are happy to discuss which specific criteria still need to be met to support an increase.

Thank you again for your time and consideration.

I have read the authors response and prefer to keep my score.

Comments on the paper

1 comment
Area_Chair_wjWHOpenReview ↗

Reminder to Respond to the Authors’ Rebuttal

Dear Reviewer,

Thank you for your contribution to the review process. Please review the authors’ rebuttal and provide a follow-up response as soon as possible.

Your timely feedback will help us complete the discussion and reach an informed decision.

Thank you for your time and effort.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.