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.

More from Models