Report
Evolving memory reportedly beats test-time training on five math-discovery tasks
A post says TTME kept model weights fixed and used 82.8%–84.6% fewer FLOPs than test-time training on three tasks.
TLDR
A post describing a UC Santa Barbara paper says its TTME approach stores lessons from earlier attempts instead of updating model weights during a search. It says TTME beat the reported test-time training scores on five math tasks across two frameworks and three models. With Qwen3-8B on three tasks, the post reports 82.8%–84.6% fewer FLOPs and 47.7%–58.4% less runtime than test-time training.
Combined views
1.7K
4 Sources, first seen ago
