Semantics or Structure? Auditing Text Sensitivity in Multimodal Time-Series Forecasting
研究通过文本扰动和归因分析，探讨了多模态时间序列预测中模型是否对文本语义敏感，发现文本内容并非提升预测性能的关键因素。
0 citationsRead paper
Selected work
Representative Papers
Mix, Don't Pick: Why Synthetic Corpus Composition Matters for Time Series Foundation Model Pretraining
This work addresses the lack of a unified criterion for selecting synthetic data generators in pretraining time series foundation models, where the optimal generator varies across model architectures, leading to unstable performance. Rather than treating this as a single-generator selection problem, the study reframes it as a corpus composition challenge and proposes constructing pretraining corpora by equally mixing multiple synthetic generators, further refined through integration with real data. Experiments training Chronos-T5-Mini and Moirai-Small from scratch demonstrate that this mixed-generation strategy matches or surpasses the best individual generator on both architectures; when combined with real data, it achieves overall superior pretraining performance. These results validate that effective corpus composition strategies must be tailored to specific model families.
0 citationsRead paper
GITCO: Gated Inference-Time Context Optimization in TSFMs
This work addresses the vulnerability of patch-based time series foundation models (TSFMs) to contextual contamination from structurally anomalous patches during inference, which degrades zero-shot forecasting performance. To mitigate this issue without updating model parameters, the authors propose GITCO—a lightweight inference-time framework comprising three components: a Gate, a Router, and a Critic. GITCO introduces, for the first time, a context sensitivity profile to dynamically identify and suppress harmful patches by selectively modulating their attention weights. Evaluated on TimesFM 2.5 across 53 datasets in the GIFT-Eval benchmark, GITCO reduces average MASE by 1.95%, achieving 89.9% of the theoretical improvement ceiling and substantially enhancing zero-shot prediction accuracy.
0 citationsRead paper
LLM-as-a-Judge for Time Series Explanations
This work addresses the absence of a general framework for evaluating the faithfulness of natural language explanations generated by large language models (LLMs) for time series data without relying on reference texts. It proposes the first reference-free evaluation method, leveraging LLMs to perform ternary correctness judgments based on pattern recognition, numerical accuracy, and answer faithfulness. By integrating a synthetic benchmark with a data-driven mechanism, the approach enables stable and reliable scoring and ranking of generated explanations without requiring ground-truth labels. Experimental results demonstrate that, despite substantial variability in LLMs’ generation performance (accuracy ranging from 0.00 to 0.96), their capacity for faithful evaluation remains highly robust, thereby validating the effectiveness and innovative potential of using LLMs as evaluators of time series explanations.
0 citationsRead paper
Domain-Adaptive and Scalable Dense Retrieval for Content-Based Recommendation
This work addresses the keyword sparsity problem in e-commerce recommendation arising from lexical mismatches between users’ natural language intents and item metadata. To this end, the authors formulate content-based recommendation as a dense retrieval task and propose an end-to-end reproducible, domain-adaptive dual-tower semantic matching architecture. The encoders are jointly fine-tuned via supervised contrastive learning and multi-negative ranking loss. For efficient CPU deployment, the system integrates FAISS HNSW indexing, ONNX Runtime, and INT8 dynamic quantization. Evaluated on a dataset comprising 826,402 items, the approach improves Recall@10 from 0.26 (BM25 baseline) to 0.66, reduces model size to one-quarter of the original, and achieves a median inference latency of only 6.1 milliseconds.
0 citationsRead paper
Recent publications
Latest Papers
Semantics or Structure? Auditing Text Sensitivity in Multimodal Time-Series Forecasting
研究通过文本扰动和归因分析，探讨了多模态时间序列预测中模型是否对文本语义敏感，发现文本内容并非提升预测性能的关键因素。
0 citationsRead paper
Mix, Don't Pick: Why Synthetic Corpus Composition Matters for Time Series Foundation Model Pretraining
This work addresses the lack of a unified criterion for selecting synthetic data generators in pretraining time series foundation models, where the optimal generator varies across model architectures, leading to unstable performance. Rather than treating this as a single-generator selection problem, the study reframes it as a corpus composition challenge and proposes constructing pretraining corpora by equally mixing multiple synthetic generators, further refined through integration with real data. Experiments training Chronos-T5-Mini and Moirai-Small from scratch demonstrate that this mixed-generation strategy matches or surpasses the best individual generator on both architectures; when combined with real data, it achieves overall superior pretraining performance. These results validate that effective corpus composition strategies must be tailored to specific model families.
0 citationsRead paper
GITCO: Gated Inference-Time Context Optimization in TSFMs
This work addresses the vulnerability of patch-based time series foundation models (TSFMs) to contextual contamination from structurally anomalous patches during inference, which degrades zero-shot forecasting performance. To mitigate this issue without updating model parameters, the authors propose GITCO—a lightweight inference-time framework comprising three components: a Gate, a Router, and a Critic. GITCO introduces, for the first time, a context sensitivity profile to dynamically identify and suppress harmful patches by selectively modulating their attention weights. Evaluated on TimesFM 2.5 across 53 datasets in the GIFT-Eval benchmark, GITCO reduces average MASE by 1.95%, achieving 89.9% of the theoretical improvement ceiling and substantially enhancing zero-shot prediction accuracy.
0 citationsRead paper
LLM-as-a-Judge for Time Series Explanations
This work addresses the absence of a general framework for evaluating the faithfulness of natural language explanations generated by large language models (LLMs) for time series data without relying on reference texts. It proposes the first reference-free evaluation method, leveraging LLMs to perform ternary correctness judgments based on pattern recognition, numerical accuracy, and answer faithfulness. By integrating a synthetic benchmark with a data-driven mechanism, the approach enables stable and reliable scoring and ranking of generated explanations without requiring ground-truth labels. Experimental results demonstrate that, despite substantial variability in LLMs’ generation performance (accuracy ranging from 0.00 to 0.96), their capacity for faithful evaluation remains highly robust, thereby validating the effectiveness and innovative potential of using LLMs as evaluators of time series explanations.
0 citationsRead paper
Domain-Adaptive and Scalable Dense Retrieval for Content-Based Recommendation
This work addresses the keyword sparsity problem in e-commerce recommendation arising from lexical mismatches between users’ natural language intents and item metadata. To this end, the authors formulate content-based recommendation as a dense retrieval task and propose an end-to-end reproducible, domain-adaptive dual-tower semantic matching architecture. The encoders are jointly fine-tuned via supervised contrastive learning and multi-negative ranking loss. For efficient CPU deployment, the system integrates FAISS HNSW indexing, ONNX Runtime, and INT8 dynamic quantization. Evaluated on a dataset comprising 826,402 items, the approach improves Recall@10 from 0.26 (BM25 baseline) to 0.66, reduces model size to one-quarter of the original, and achieves a median inference latency of only 6.1 milliseconds.
0 citationsRead paper