TTME beats test-time training on math discovery with 84% fewer FLOPs
rohanpaul_ai · x · 2026-10-29
A new UC Santa Barbara paper (arXiv:2610.32046) introduces Test-Time Memory Evolution (TTME), letting LLMs skip costly test-time training (TTT) on math discovery tasks.
- TTT's problem: running RL weight updates on scored attempts during search is slow, expensive, and hard to scale to bigger models.
- TTME's approach: freeze the weights and store lessons in external memory — practical coding rules distilled from earlier attempts plus the ideas behind score gains — to guide later attempts.
- Results: across 5 math tasks, 2 frameworks, and 3 models, TTME beat reported TTT scores every time; with Qwen3-8B it cut FLOPs by 82.8–84.6% and runtime by 47.7–58.4% on 3 tasks.
Takeaway: if your LLM search loop has an automatic checker, try a memory of lessons before paying for mid-search weight updates. Code is open-sourced.
Related event: TTME Replaces Test-Time Training with Memory Evolution, Cutting FLOPs 84%(3 posts)→
More from coding & agent
- No papers needed: a 4-step recipe to start mech-interp with a local SLM and Claude — burny_tech · 2026-10-29
- Dev rants: coding agents taint codebases with leftover comments after every fix — blelbach · 2026-10-29
- PSA: Claude Code deletes your transcripts after 30 days — ask to change it — mike64_t · 2026-10-29
- Agents Use Your Codebase as Memory — and That Changes How You Should Write It — auto_grad_ · 2026-10-29
- Dev builds fully automated AI content farm, gains ~100 followers in 5 days — cneuralnetwork · 2026-10-29
- Redditor seeks open-source VLM to reverse-engineer cinematic prompts from stills, says ChatGPT falls short — Complete-Box-3030 · 2026-10-29