Research map / Time series / Anomaly detection
AnomSeer: Reinforcing Multimodal LLMs to Reason for Time-Series Anomaly Detection
Junru Zhang, Lang Feng, Haoran Shi, Xu Guo, Han Yu, Yabo Dong, Duanqing Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ExpCoT (classical-analysis expert traces) instead of GPT-4o generated CoT as reasoning reference (ablation) | AnomSeer-3B · AnomLLM frequency anomalies | Affinity-F1 (%): 49.8 -> 58.9 | Improved |
| use ExpCoT instead of GPT-4o CoT (ablation) | AnomSeer-3B · AnomLLM trend anomalies | Affinity-F1 (%): 79.5 -> 84.9 | Improved |
| use ExpCoT instead of GPT-4o CoT (ablation) | AnomSeer-3B · AnomLLM range/point anomalies | Affinity-F1 (%): range 84.4 -> 85.6; point 86.1 -> 87.8 | Improved |
| add orthogonal projection of the reasoning advantage (ablation) | AnomSeer-3B · AnomLLM frequency anomalies | Affinity-F1 (%): 53.5 -> 58.9 | Improved |
| add orthogonal projection of the reasoning advantage (ablation) | AnomSeer-3B · AnomLLM trend/range/point anomalies | Affinity-F1 (%): trend 81.1 -> 84.9; range 83.5 -> 85.6; point 85.4 -> 87.8 | Improved |
| add ExpCoT, time-series reasoning advantage and orthogonalization on top of vanilla GRPO (ablation) | GRPO on Qwen2.5-VL-3B · AnomLLM point anomalies | Affinity-F1 (%): 80.6 -> 87.8 | Improved |
| add TimerPO components to vanilla GRPO (ablation) | GRPO on Qwen2.5-VL-3B · AnomLLM frequency anomalies | Affinity-F1 (%): 50.4 -> 58.9 | Improved |
| increase SFT data tenfold (3.2k to 32k instances) | Qwen2.5-VL-3B SFT · AnomLLM test | average Affinity-F1 (%): 32.8 -> 39.4 | Improved |
| use AnomSeer-3B RL instead of SFT on 32k instances | Qwen2.5-VL-3B · AnomLLM test | average Affinity-F1 (%): 39.4 -> 79.3 | Improved |
| scale AnomSeer from 3B to 7B backbone | AnomSeer · AnomLLM test | average Affinity-F1 (%): 79.3 -> 84.4 | Improved |
| use AnomSeer-3B instead of TimeMaster-3B (SFT+GRPO) | Qwen2.5-VL-3B · AnomLLM test | average Affinity-F1 (%): 71.9 -> 79.3 | Improved |
| vary reasoning advantage weight alpha (ablation) | AnomSeer-3B · AnomLLM | Affinity-F1 | No clear effect |
Closest papers by meaning
- Can LLMs Understand Time Series Anomalies? (ICLR 2025 · similarity 0.90)
- Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.88)
- AD-FM: Multimodal LLMs for Anomaly Detection via Multi-Stage Reasoning and Fine-Grained Reward Optimization (AAAI 2026 · similarity 0.88)
- Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability. (NeurIPS 2026 · similarity 0.87)
- The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.87)
- Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
