paper / mlittmancs / May 12
DRAGO is a continual model-based RL framework designed to maintain and grow world models across a sequence of tasks that share state space and dynamics but differ in reward functions. It combats catastrophic forgetting through two mechanisms: generative synthetic experience replay (avoiding raw data storage) and an intrinsic reward signal that steers the agent back toward states relevant to prior tasks. The combination enables incremental world model development without task-specific replay buffers, achieving superior performance across continual learning benchmarks.
reinforcement-learningcontinual-learningworld-modelscatastrophic-forgettingmodel-based-rlknowledge-retentiongenerative-models
“DRAGO enables knowledge retention across sequential tasks in continual model-based RL without storing raw past experience data.”
paper / mlittmancs / May 12
Despite LLMs exhibiting unpredictable performance swings across prompt phrasings, this research identifies a consistent internal mechanism — "lexical task heads" — specific attention heads whose outputs literally encode the task being performed. These heads are shared across both instruction-based and few-shot example-based prompting styles, suggesting a unified internal task representation underlies surface-level behavioral variability. Prompt-to-prompt performance differences are explained by the degree of activation of these heads, with failures linked to competing task representations that dilute the target task signal — offering a mechanistic account of what has traditionally appeared as arbitrary model brittleness.
llm-interpretabilityprompt-sensitivitymechanistic-interpretabilityattention-headsfew-shot-promptinglarge-language-modelsbehavioral-analysis
“LLMs engage shared underlying attention mechanisms (lexical task heads) across structurally different prompt styles for the same task.”