Alibaba and ZJU Propose TTPO for Label-Free Test-Time Self-Improvement in LLMs

Alibaba and Zhejiang University proposed TTPO, a test-time policy optimization method that distills consistent reasoning paths and penalizes inconsistent ones under majority-vote pseudo-labels. It enables label-free test-time training, boosting Qwen3-1.7B by about 7 points on math reasoning and matching supervised distillation.

2026-08-28 ~ 2026-08-29 · 3 related posts

• TTPO: Test-Time Policy Optimization for Label-Free Math Reasoning — Aozhe Wang · 2026-08-28
• Alibaba and ZJU Propose TTPO for Label-Free Test-Time Training — burkov · 2026-08-29
• TTPO: Test-Time Policy Optimization Boosts Qwen3-1.7B from 38.0% to 45.2% Without Labels — JFPuget · 2026-08-29