Research map / Large language models / Agents, code and math
A2Seek: Towards Reasoning-Centric Benchmark for Aerial Anomaly Understanding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces A2Seek, a large drone-view anomaly dataset with reasoning annotations, and A2Seek-R1, a GoT-guided SFT plus A-GRPO reinforcement fine-tuning framework with a seeking mechanism for aerial anomaly understanding.
Problem
Video anomaly detection from UAV (drone) viewpoints is hard due to dynamic viewpoints, scale variation, and small anomalies, and existing datasets and methods are built for fixed ground-level cameras without spatial localization or causal reasoning annotations.
Approach
A reasoning-centric aerial anomaly benchmark with graph-of-thought (trigger/diagnose/reasoning/reflection/seeking) annotations and bounding boxes, plus a two-stage training pipeline: GoT-guided SFT followed by A-GRPO with format, accuracy, localization, seeking, and length rewards, where the seeking stage predicts a region to crop and re-inspect, mimicking UAV viewpoint adjustment.
Key results
- A2Seek-R1 reaches 53.14% average AP_c on A2Seek versus 46.42% for GoT-SFT, a 6.72-point gain, and 34.03% for zero-shot.
- A2Seek-R1 improves mIoU to 26.03% versus 20.81% for GoT-SFT and 3.50% zero-shot.
- On Qwen2.5-VL-7B, A2Seek-R1 raises AP_c from 44.65% to 57.07%; on InternVL-3.0-2B, from 20.03% to 48.76%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply A2Seek-R1 instead of GoT-SFT | Qwen2.5-VL-3B-Instruct · A2Seek | AP_c: 46.42 -> 53.14 | Helped |
| apply A2Seek-R1 instead of GoT-SFT | Qwen2.5-VL-3B-Instruct · A2Seek | mIoU: 20.81 -> 26.03 | Helped |
| apply A2Seek-R1 | QwenVL-2.5-7B-Instruct · A2Seek | AP_c: +12.42 | Helped |
| apply A2Seek-R1 | InternVL-3.0-2B-Instruct · A2Seek | AP_c: 20.03% to 48.76% | Helped |
| add CoT to ANS-only SFT (ablation) | A2Seek-R1 SFT · A2Seek | AP_c: 31.10 -> 38.24 | Helped |
| replace CoT with GoT in SFT (ablation) | A2Seek-R1 SFT · A2Seek | AP_c: 38.24 -> 46.42 | Helped |
| add accuracy and location rewards to ANS+GoT SFT (ablation) | A2Seek-R1 · A2Seek | mIoU: 20.81 -> 24.03 | Helped |
| add seeking reward to accuracy and location rewards (ablation) | A2Seek-R1 · A2Seek | AP_c: 51.78 -> 53.14 | Helped |
| add seeking reward to accuracy and location rewards (ablation) | A2Seek-R1 · A2Seek | mIoU: 24.03 -> 26.03 | Helped |
| use A2Seek-R1's longer test-time reasoning token budget | A2Seek-R1 · A2Seek | Samples/s: 0.9174 -> 0.3267 | Hurt |
Surprising findings
- RL-Zero (GRPO with only format and accuracy rewards, no SFT) averages just 12.42% AP_c, far below zero-shot.
- Zero-shot and ANS-SFT get high AP at IoU 0.00 mainly by predicting overly broad regions, with near-zero AP at IoU 0.90.
Limitations
- Complex reasoning chains and strategy learning limit inference speed and deployment efficiency.
- Focuses on spatial-level anomaly perception; long-duration temporal anomalies are less well modeled.
- Privacy and fairness concerns for public-safety applications.
Experimental setup
Datasets: A2Seek
Benchmarks: A2Seek
Metrics: AP_c mIoU AP at IoU thresholds BLEU METEOR ROUGE CIDEr samples/s
Baselines: GRPO RL-Zero ANS-SFT CoT-SFT GoT-SFT SSRL HSTforU ANDT AnomalyRuler LAVAD Holmes-VAU LLaVA-Video InternVL-3.0
Builds on: GRPO Chain-of-thought Qwen2.5-VL Grounded-SAM-2
Closest papers by meaning
- Is your VLM Sky-Ready? A Comprehensive Spatial Intelligence Benchmark for UAV Navigation (CVPR 2026 · similarity 0.85)
- Game4Loc: A UAV Geo-Localization Benchmark from Game Data (AAAI 2025 · similarity 0.84)
- Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery (NeurIPS 2026 · similarity 0.84)
- SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery (NeurIPS 2026 · similarity 0.84)
- Self-Prompting Analogical Reasoning for UAV Object Detection (AAAI 2025 · similarity 0.84)
- PixDLM: A Dual-Path Multimodal Language Model for UAV Reasoning Segmentation (CVPR 2026 · similarity 0.83)
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.