Today for AI
HOT RADAR
llmHEAT 13.9°

AI CURATED HOT STORY · 10/29/2026

UCSB Paper: TTME Replaces Test-Time Training with Evolving Memory for Math Discovery

1 reports archived1 independent sourcesupdated 10/29/2026, 13:15:00
Synthesis & Latest Updates
1 Sources Cross-Validated

A new paper from UC Santa Barbara introduces TTME, a framework that replaces costly test-time training (TTT) with an 'evolving memory' mechanism. By keeping model weights fixed and storing practical coding rules and score-gain insights from previous attempts, TTME outperforms TTT across multiple math tasks while reducing FLOPs by ~82% and runtime by nearly half.

LATEST/A new paper from UC Santa Barbara introduces TTME, a framework that replaces costly test-time training (TTT) with an 'evolving memory' mechanism. By keeping model weights fixed and storing practical coding rules and score-gain insights from previous attempts, TTME outperforms TTT across multiple math tasks while reducing FLOPs by ~82% and runtime by nearly half.

Coverage timeline

Total 1 reports · Latest first
  1. Rohan Paul (@rohanpaul_ai)T2·78 pts

    UCSB Paper: TTME Replaces Test-Time Training with Evolving Memory for Math Discovery

    • TTME avoids the expensive RL weight updates of test-time training by using fixed weights and an external memory module.
    • TTME outperformed reported test-time training baselines across 5 math tasks and 3 model architectures.
    • Benchmarking with Qwen3-8B showed TTME reduced FLOPs by 82.8%-84.6% and runtime by 47.7%-58.4% compared to TTT.