Questions Regarding Paper Details and Implementation

[LioN-YH]

Dear authors,
Thank you for your excellent work! I have a few questions regarding the paper and experimental setup, and I would greatly appreciate it if you could kindly clarify them.

• Text-Augmented Learner:
  The code comments mention that "Each variable in the time series will have its own prompt." However, the actual implementation appears inconsistent with this description. Instead, the code computes statistics (e.g., min, max, mean, and trend) across multiple variables to generate prompts. I'm concerned about whether such prompts carry meaningful and semantically coherent information.

B, T, n_vars = x_enc.shape  # Get batch size, sequence length, and number of variables

# Initialize a list to store prompts for each batch
prompts = []

# Calculate overall statistics for each batch
for b in range(B):
    # Calculate statistics for the current batch
    min_value = torch.min(x_enc[b]).item()  # Overall minimum value for the batch
    max_value = torch.max(x_enc[b]).item()  # Overall maximum value for the batch
    median_value = torch.median(x_enc[b]).item()  # Overall median value for the batch
    trend = x_enc[b].diff(dim=0).sum().item()  # Overall trend for the batch

    # Determine the overall trend direction
    trend_direction = "upward" if trend > 0 else "downward"

• Channel Strategy:
  I would like to inquire about the channel strategy used in the current model framework—specifically, whether a channel-independent strategy is adopted. After carefully examining both the code and the paper, it appears that Retrieval-Augmented Learner does employ channel independence, whereas the Vision-Augmented Learner and Text-Augmented Learner do not. This inconsistency has left me somewhat confused.

• Experimental Details and Hyperparameter Settings:
  From the scripts in the repo, I have noticed that d_model, use_mem_gate, and dropout are important hyperparameters. I have a question about the experimental setup in the paper: whether the experiments in the paper (including short-term forecasting, long-term forecasting, few-shot forecasting, zero-shot forecasting, and ablation studies) were conducted under a unified set of hyperparameters, or whether hyperparameter tuning was performed separately for each group of experiments to report relatively optimal performance. Could you please provide a detailed explanation?
  I ask because the released script TimeVLM_long_1.0p.sh uses different hyperparameter configurations for different datasets. When I used this script for testing, the model's actual performance fell short of the results reported in the paper. This discrepancy is particularly noticeable on datasets like ETTh1 (for horizons 336/720) and Electricity. Could you provide the exact hyperparameters needed to reproduce the paper's results?

Thank you very much for your time and consideration. I truly appreciate your contribution to the community, and I look forward to your response!!!

[siruzhong]

Thank you for your detailed questions about our Time-VLM work. Regarding your observations: (1) You're correct about the Text-Augmented Learner implementation - the current design calculates overall statistics rather than per-variable statistics for computational efficiency and semantic coherence. (2) Your observation about channel independence is accurate. We have tested per-channel approaches where each patch creates an image and text, but it increased power consumption significantly without improving performance. Time-VLM's core motivation is to create the vision-language prior space to complement time series when data is scarce, providing an overall space for alignment rather than individual channel processing. (3) For hyperparameter settings, our best parameters were extracted via search - you can refer to https://github.com/CityMind-Lab/ICML25-TimeVLM/blob/main/scripts/TimeVLM_search_parallel.sh. We have updated partial parameters for ETTh1 and Electricity, you can check https://github.com/CityMind-Lab/ICML25-TimeVLM/blob/main/scripts/TimeVLM_long_1.0p.sh. The experiments used dataset-specific hyperparameter tuning rather than unified configurations to ensure optimal performance for each dataset-prediction length combination.