Core NLP & Model Operations
Unlearning, text watermarking, continual learning, and domain adaptation work.
Core NLP & Model Ops
This theme bundles the infrastructure that keeps language models reliable once they leave the lab: unlearning, watermarking, continual learning, and domain adaptation for high-stakes verticals.
Featured Work
- Beyond the Unlearning Mirage - Dynamic evaluations that stress-test multi-hop leakage and fuse in watermarking + continual-learning tactics.
- Domain-Specific Evaluations With Real Consequences - Finance, medical, and cybersecurity benchmarks that measure impact instead of just BLEU.
Highlights
- Unlearning & Safety
- Dynamic probe generation (single hop → multi-hop, alias chains)
- Activation-pathway tracing to confirm concepts are truly removed
- Pip package + leaderboard for reproducible stress tests
- Provenance & Watermarking
- Taxonomy covering robustness under editing/fine-tuning
- Guidance for platform operators to audit leaked generations
- Continual Learning & Domain Adaptation
- Power-law learning environments that mitigate catastrophic interference
- Domain-specific adapters for finance, clinical narratives, and more
How to use this work
- Plug your unlearning method into the evaluator before deployment.
- Cite the watermarking taxonomy when rolling out provenance policies.
- Adopt the domain benchmarks when showcasing vertical LLM claims.
Questions or collab ideas? Drop a note in Discord or email-this stack is meant to be extended.