TTPO paper: Qwen3-1.7B lifts accuracy from 38.0% to 45.2% without ground-truth labels
TheTuringPost · x · 2026-09-02
Turing Post highlights TTPO (Test-Time Policy Optimization), a paper showing models can learn at test time without any ground-truth answers.
How it works:
- The model never sees true answers — it uses a majority vote across its own attempts as pseudo-labels
- Solutions agreeing with the majority are treated as likely useful; disagreeing ones as likely wrong — but TTPO penalizes only the most suspicious tokens, not whole solutions
- This asymmetric approach reduces the risk of reinforcing wrong answers versus treating the majority as definitely correct
Results:
- Qwen3-1.7B's average accuracy improved from 38.0% to 45.2% with no answer labels
- Gains transferred across math benchmarks, suggesting broader learning beyond one dataset
Why it matters: models can keep learning even where no answer key exists.
Related event: TTPO Enables Test-Time Learning Without Ground Truth(2 posts)→
More from Models
- MOSS SFT vs Flux 3 Klein vs SD 3.0: Same-Prompt Comparison Exposes SD3's Age — ZootAllures9111 · 2026-10-18
- Quick take: GPT-6 matches Opus 5.5 at coding, far ahead at computer use — _vztu · 2026-10-18
- Open models now above Sol-level, outpacing expectations as community awaits V4.1 Pro — yacineMTB · 2026-10-18
- Codex took 3 days, Opus 5.5 nailed it in an afternoon — banned user's video protest — dotey · 2026-10-18
- Yacine MTB: Users will remember how Anthropic treated them once intelligence proliferates — yacineMTB · 2026-10-18
- Mistral's Timestral forecasting models start at 5M params and run on plain CPUs — lmoroney · 2026-10-18