Announcement
Meta researchers say Kimi-K2.6 can learn its own memory playbook
Meta researchers say Self-Improving Memory discovers better memory prompts, then trains the model on those behaviors.
TLDR
On three Continual Learning Bench tasks, Meta researchers report optimized memory prompts scored 56.0 versus 44.6 for the best hand-written versions. Fine-tuning on optimized runs, with generic prompts substituted, produced a model scoring 43.1 versus 37.6 for training on generic-prompt runs, they say. But 37 of 44 proposed revisions failed to beat previous best scores in their search. Self-improvement still needs a tough editor.
Combined views
21.5K
7 Sources, first seen ago

Comments