LLM Adapters for Agentic Workflows- MSC/PHD Student Position- Research Lab

Summary: Join IBM Research in Beer Sheva to build Granite Switch, an open-source AI model. You will train and evaluate LLM adapters for agentic workflows, design benchmarks, and contribute to open-source projects while collaborating with experienced researchers.

At IBM Research, you will work alongside experienced researchers on real technological challenges at the cutting edge of AI. You'll gain hands-on experience, build in-demand technical and professional skills, and contribute to meaningful projects in a collaborative, team-based environment.

We are seeking a student to join the IBM Research lab in Beer Sheva and help us build the next generation of AI: Granite Switch. Build AI models like you build software.

Our work is open source: [1] https://github.com/generative-computing/granite-switch

In this role, you will train and evaluate LLM adapters and apply them to real customer use cases in agentic workloads. You will also develop new ways to switch between adapters quickly and efficiently.

Your role and responsibilities:

• Work as part of a research team
• Break customer agentic workflows down into discrete tasks and tools, and design a specialized adapter for each one
• Build datasets for adapter training, including synthetic data generation, data curation and labeling pipelines
• Train, fine-tune and evaluate LLM adapters on security-domain tasks
• Design benchmarks and evaluation methods that compare Granite Switch with large frontier models on accuracy, latency, cost and data privacy
• Build end-to-end proof-of-concept demos showing that Granite Switch solves customer problems better than the alternatives
• Analyze results, find where adapters fall short, and iterate on data, training and adapter composition
• Present findings to the team and to customers, and contribute to publications, patents and technical blog posts
• Contribute code, adapters and benchmarks to Granite Switch and related open-source projects (vLLM, Hugging Face Transformers)

References

Visible links
1. https://github.com/generative-computing/granite-switch