Undergraduate thesis · Lead technical researcher
WebFARM: A Multimodal Framework for Failure Detection, Adaptive Recovery and Episodic Memory in Autonomous Web Agents
Supervisor: Asst. Prof. Nuzhat Tabassum, Department of Computer Science, AIUB
- Led the technical research: methodology and experimental design, framework architecture and implementation, model integration, experiment execution, and results analysis.
- Developed and evaluated multimodal approaches using Qwen2-VL-2B, Qwen2.5-VL-7B, and InternVL3.5-8B on Web-Gold-40K, a dataset of 40,017 real-browser interaction steps.
- Evaluated Browser Use and SeeAct agents across 22 websites and 42 tasks.
- Best F1
- 0.839
- Recovery prediction
- 93.49%
- Browser Use gain
- +13.1pp
- SeeAct gain
- +20.2pp
Preliminary results; task-completion gains measured over the unmodified agents.