IBM DeepScanner
A scalable framework for detecting anomalies in LLM activations
About the Project
DeepScanner supports the analysis of internal language-model activations using non-parametric subset scanning. I contributed to workflows for extracting activations and investigating anomalous patterns associated with model behavior.
I engineered the distributed workflow for this framework, optimizing the pipeline to run parallel experiments on the IBM Cloud Cognitive Cluster. This significantly reduced the time required to analyze large-scale models like Llama 3 and Granite.
The work involved modular analysis across model layers and reproducible evaluation workflows. An anomalous activation pattern is a signal to investigate, rather than proof of a particular behavior.
Project Details
Milestones
- Architecture Design - June 2025
Designed the modular architecture for extracting and analyzing model activations
- Distributed Pipeline - July 2025
Implemented Ray-based distributed processing to handle large-scale model data
- Optimization - August 2025
Optimized data loading and tensor processing, achieving a 3x speedup in analysis time
- Paper Submission - August 2025
Results contributed to a non-archival paper later presented at the AAAI 2026 Workshop on New Frontiers in Information Retrieval