a12354/AnomSeer

AnomSeer: Reinforcing Multimodal LLMs to Reason for Time-Series Anomaly Detection


@ ICML 2026

 

AnomSeer is a reinforcement-learning post-training framework for time-series anomaly detection (TSAD) with multimodal LLMs. It unifies anomaly classification, localization, and explanation by grounding MLLM reasoning in fine-grained, classical TSAD evidence through two components:

• ExpCoT — expert chain-of-thought traces generated from classical statistical methods (FFT, Matrix Profile, gradient analysis) that provide structured, quantitatively verifiable supervision.
• TimerPO — a novel RL algorithm that uses Optimal Transport to measure semantic alignment between model reasoning and ExpCoT, then applies orthogonal projection to inject this signal without interfering with the primary detection objective.

With Qwen2.5-VL-3B/7B-Instruct as backbone, AnomSeer outperforms GPT-4o and Gemini-2.5-Pro on anomaly classification and localization across AnomLLM, VisualTimeAnomaly, and TSB-UAD benchmarks.

---

Table of Contents

• Installation
• Data Preparation
  • AnomLLM (Training Data)
  • VisualTimeAnomaly
  • TSB-UAD
  • Preprocess into Parquet
• Training
• Evaluation
• Key Design Choices
• Citation
• Acknowledgements

---

Installation

conda create -n anomseer python=3.11 -y
conda activate anomseer

pip3 install -r requirements.txt
pip3 install flash-attn==2.7.4.post1 --no-build-isolation
pip3 install -e .

pip3 install -r requirements_anomseer.txt
pip3 install git+https://github.com/ahstat/affiliation-metrics-py.git

Requires CUDA 12.4 (for torch 2.6.0 and flash-attn). flash-attn must be installed separately with --no-build-isolation.

---

Data Preparation

AnomSeer is trained on AnomLLM (synthetic) and evaluated on AnomLLM, VisualTimeAnomaly, and TSB-UAD.

1. AnomLLM (Training Data)

Clone AnomLLM, set up its environment, and generate the synthetic dataset:

git clone https://github.com/Rose-STL-Lab/AnomLLM.git
cd AnomLLM
conda env create --file environment.yml && conda activate anomllm && poetry install --no-root
bash synthesize.sh

2. VisualTimeAnomaly (Evaluation)

Download from: https://github.com/mllm-ts/VisualTimeAnomaly

3. TSB-UAD (Evaluation)

Download from: https://github.com/decisionintelligence/TAB

4. Preprocess into Parquet

python multimodal_data_processing/anom.py \
    --root_dir /path/to/AnomLLM/data/synthetic \
    --out_dir ./data/anol_processed_mllm_data

This produces train_full.parquet (3,200 samples) and test_full.parquet under ./data/anol_processed_mllm_data/.

---

Training

Edit MODEL_PATH in example/timerpo_trainer/run_anomseer.sh to select the backbone:

MODEL_PATH=Qwen/Qwen2.5-VL-3B-Instruct   # 3B variant
# MODEL_PATH=Qwen/Qwen2.5-VL-7B-Instruct  # 7B variant

Then launch training (requires 4 GPUs by default):

bash example/timerpo_trainer/run_anomseer.sh

Checkpoints are saved to ./checkpoints/.

To override the data paths without editing the script:

TRAIN_FILE=./data/anol_processed_mllm_data/train_full.parquet \
VAL_FILE=./data/anol_processed_mllm_data/test_full.parquet \
bash example/timerpo_trainer/run_anomseer.sh

---

Evaluation

Set EVAL=True inside example/timerpo_trainer/run_anomseer.sh and point VAL_FILE to the desired benchmark:

EVAL=True
VAL_FILE=./data/anol_processed_mllm_data/test_full.parquet

Then run:

bash example/timerpo_trainer/run_anomseer.sh

Metrics reported in the logs:

• Affinity-Precision / Affinity-Recall / Affinity-F1 — temporal localisation quality
• Classification Accuracy — anomaly type identification

To evaluate on VisualTimeAnomaly or TSB-UAD, preprocess those datasets with multimodal_data_processing/anom.py into the same parquet format and set VAL_FILE accordingly.

---

Citation

@inproceedings{zhang2026anomseer,
  title     = {AnomSeer: Reinforcing Multimodal LLMs to Reason for Time-Series Anomaly Detection},
  author    = {Zhang, Junru and Feng, Lang and Shi, Haoran and Guo, Xu and Yu, Han and Dong, Yabo and Xu, Duanqing},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026}
}

Acknowledgements

We thank the veRL project for foundational RL infrastructure and AnomLLM for the synthetic TSAD benchmark and data generation code.