ESC

Type to search articles...

No articles found.

↑ ↓ Navigate ↵ Open
Esc Close
Blog Tags GitHub
All tags

test-time-training

1 post tagged with "test-time-training"

July 3, 2026

SIA: Self Improving AI with Harness & Weight Updates

提出 SIA 自改进循环——由 Feedback-Agent 同时更新任务专属 agent 的「harness(脚手架)」与「模型权重(LoRA)」,打破自改进 AI 的两大孤岛(仅改 scaffold / 仅改 weights)。在中文法律分类、GPU kernel 优化、单细胞 RNA 去噪三个对比领域上,双杠杆组合全面超越单独 scaffold 迭代:LawBench +25.1%、GPU kernel 快 12.4%、去噪 +20.4%(均超先前 SOTA)

reinforcement-learning multi-agent-systems llm-agent test-time-training llm-training

Navigation

  • Work

Resources

  • Lexington Themes.

Socials

  • @Mike_Andreuzza
© 2025 MicroStudio. All rights reserved.

MicroStudio is not affiliated with Stripe, Breeew, Astro, or Tailwind Labs, nor is it endorsed or sponsored by them.