IBM DeepScanner
Research infrastructure for analyzing anomalous patterns in LLM activations
- Status
- Internal Tool
- My role
- AI Research Scientist Intern
- Period
- June 2025 - September 2025
Overview
DeepScanner applies non-parametric subset-scanning methods to internal neural-network representations to study anomalous behavior, distribution shifts, and representation structure.
At IBM Research, I improved the codebase, implemented and ran experiments, and contributed to infrastructure for activation-space analysis across local and distributed GPU environments.
The work also contributed to research on retrieving from a vector database using multiple query vectors.
Methods and Tools
- Python
- PyTorch
- Ray (Distributed Computing)
- IBM Cloud Cognitive Cluster
- Hugging Face Transformers
- Weights & Biases
Project milestones
- June - September 2025
Research Infrastructure
Improved and used modular workflows for activation extraction, scanning, and reproducible large-scale experiments
- 2026
Workshop Publication
Contributed to Retrieval With Multiple Query Vectors Through Anomalous Pattern Detection, a paper at the AAAI Workshop on Frontiers in Information Retrieval