Retain-set-free LLM unlearning, developed under the IARPA Bengal program. Demotes memorized logits at high-surprisal tokens via self-distillation — no curated retain set required.

In-context memory and adaptation for LLM agents.

Persona/emotion vectors as intrinsic reward for multi-agent LLM cooperation — contagion and cause-attribution on GSM8K.

A meta markdown file that evolves on its own.