Distilling VLMs for On-Device Fire Understanding

This is my current research at Aalto University with Prof. Siavash Khajavi. Large vision-language models can look at a scene and decide whether a fire is actually happening, which cuts down on false alarms. They are also far too big for a fire sensor.

Approach

I built a teacher–student distillation setup that trains compact 0.5B–3B students from a 7B Qwen2.5 teacher using a combination of KL-divergence and cross-entropy supervision. The students are evaluated on the DetectiumFire benchmark and then on a commercial Detectium sensor.

Results

Metric	Value
Parameter reduction (teacher → student)	14×
Burning-object accuracy retained	95.2% of teacher
Fire / non-fire accuracy after deployment	99.5%
Peak memory on Raspberry Pi 5 (CPU only)	1.9 GB
End-to-end latency	≈ 4 s
False alarms on 1,452 test images	2

Paper and next steps

The paper is on arXiv (Kazzazi et al., 2026). I am now looking at prompt optimization and reasoning strategies for vision-language-action models.

References

2026

arXiv
Distilling Vision-Language Models for On-Device Fire Understanding

Mohammad Kazzazi, Zixuan Liu, and Siavash Khajavi

arXiv preprint arXiv:2609.05782, Sep 2026

Abs DOI arXiv Bib HTML PDF

Vision-language models (VLMs) offer a promising alternative to conventional fire detection systems by reasoning about the semantic context of a scene and thus reducing false alarms, yet their large model size makes deployment on embedded fire sensors impractical. We develop a teacher-student knowledge distillation framework in which large VLMs fine-tuned for fire understanding are distilled into lightweight students, then deploy the distilled models on a commercial Detectium fire detection sensor and evaluate accuracy, latency, and memory. Compact students preserve most of their teachers’ fire-understanding capability, and Qwen2.5-0.5B provides the strongest overall deployment trade-off.
@article{kazzazi2026distill, title = {Distilling Vision-Language Models for On-Device Fire Understanding}, author = {Kazzazi, Mohammad and Liu, Zixuan and Khajavi, Siavash}, journal = {arXiv preprint arXiv:2609.05782}, year = {2026}, month = sep, archiveprefix = {arXiv}, primaryclass = {cs.AI}, doi = {10.48550/arXiv.2609.05782}, url = {https://arxiv.org/abs/2609.05782}, }