Speaker: Yasin Almalioglu

Abstract: As Large Language Models (LLMs) transition from single-turn assistants to the core engines of complex agentic systems, the practice of manual "prompt engineering" has become a significant bottleneck. This seminar explores the shift toward automated prompt optimization, specifically focusing on the DSPy framework and the MIPRO (Multi-stage Instruction and Prompt REcovery) strategy. We examine the challenges of credit assignment and combinatorial search spaces in multi-stage pipelines and discuss how grounded instruction generation and Bayesian surrogate models can significantly outperform human-engineered baselines. Furthermore, we address the unique requirements of multimodal LLMs (VLMs), such as Qwen3-VL, where precise grounding is essential. The session concludes with a discussion on the trade-offs between system performance and prompt interpretability in algorithmic optimization.

https://us06web.zoom.us/j/84636366540
MEETING ID: 84636366540
PASSCODE: 25862