ACL2025 Time Series AI paper notes paper summaries Time-Series Forecasting LLM Sentiment Analysis Multimodal/VLM Few-/Zero-Shot Learning Question Answering

📈 Time Series¶

💬 ACL2025 · 7 paper notes

📌 Same area in other venues: 📷 CVPR2026 (7) · 🔬 ICLR2026 (121) · 💬 ACL2026 (8) · 🧪 ICML2026 (45) · 🤖 AAAI2026 (31) · 🧠 NeurIPS2025 (54)

🔥 Top topics: Time-Series Forecasting ×6 · LLM ×2

ANRE: Analogical Replay for Temporal Knowledge Graph Forecasting: This paper proposes the ANRE (Analogical Replay) method, which retrieves structurally similar "analogical events" from historical knowledge graph snapshots as reasoning clues to assist future event forecasting in temporal knowledge graphs, achieving significant performance improvements across multiple benchmark datasets.
Context-Aware Sentiment Forecasting via LLM-based Multi-Perspective Role-Playing Agents: Proposed a multi-perspective role-playing framework (MPR) based on LLMs. By using subjective agents to simulate user posting and an objective agent (a fine-tuned "psychologist" LLM) to audit behavioral consistency, it forecasts social media users' future emotional responses to real-time events through iterative rectification, significantly outperforming traditional methods at both macro and micro levels.
CTPD: Cross-Modal Temporal Pattern Discovery for Enhanced Multimodal Electronic Health Records Analysis: A CTPD framework is proposed, which utilizes Slot Attention to discover cross-modality shared temporal prototype patterns from multimodal EHR data (irregular time series and clinical notes). Temporal semantics of both modalities are aligned via a TP-NCE contrastive loss, achieving SOTA performance on mortality prediction and phenotype classification tasks on MIMIC-III.
G2S: A General-to-Specific Learning Framework for Temporal Knowledge Graph Forecasting with Large Language Models: This paper proposes the G2S framework, which decouples general patterns (temporal structural regularities) from scenario-specific information (concrete entities/relations) in temporal knowledge graph (TKG) forecasting. By first learning general patterns on anonymized temporal structures and then injecting scenario-specific information, G2S effectively enhances the generalization capability of LLMs in TKG forecasting.
LETS-C: Leveraging Text Embedding for Time Series Classification: LETS-C is proposed: it digitizes time series into text strings, encodes them using a text embedding model, merges them with the original time series via element-wise addition, and feeds them into a lightweight CNN-MLP classification head. With only 14.5% of the trainable parameters, it achieves SOTA, outperforming 27 baselines including OneFitsAll (fine-tuned GPT-2) on 10 UEA multivariate time series datasets.
Revisiting LLMs as Zero-Shot Time-Series Forecasters: Small Noise Can Break Large Models: This paper systematically evaluates the effectiveness of LLMs as zero-shot time-series forecasters and discovers that LLMs are extremely sensitive to input noise—even a small amount of noise can lead to a drastic performance degradation, making them underperform even simple domain-specific models (such as DLinear). The authors suggest that future research should focus on fine-tuning LLMs to better handle numerical sequences.
Time-MQA: Time Series Multi-Task Question Answering with Context Enhancement: Proposes the Time-MQA framework and the TSQA dataset (~200k QA pairs), unifying time series forecasting, imputation, anomaly detection, classification, and open-ended reasoning QA under a natural language question answering paradigm, and endowing LLMs with time series understanding and reasoning capabilities through continual pre-training.