Recurrent Harness for Agentic Continual Learning
ReHarness updates an agent’s stored reflections after each task, using the task trajectory as feedback for future online adaptation.
The paper frames this as a strictly streaming setup, without extra replay or validation data. Its key move is “backward revision”: after an interaction, the harness revises the reflections that were retrieved for that instance. The authors report gains across three continual-learning benchmarks and three model backbones, beating raw-context, memory, reflection, and prompt-optimization baselines. They also say the revisions are source-grounded and help correct stale or overbroad experience under distribution shifts. HF Daily Papers' note
The paper frames this as a strictly streaming setup, without extra replay or validation data. Its key move is “backward revision”: after an interaction, the harness revises the reflections that were retrieved for that instance. The authors report gains across three continual-learning benchmarks and three model backbones, beating raw-context, memory, reflection, and prompt-optimization baselines. They also say the revisions are source-grounded and help correct stale or overbroad experience under distribution shifts. HF Daily Papers' note
score 5