Language models can improve coding answers without checking target correctness

The authors report that training on self-generated programming problems outperformed competing methods as checks on the target answer weakened, provided the starting model could sometimes solve it.

Top University

Akira Yoshiyama · Akmal Ashirmatov · Jonas Hübotter · Ido Hakimi · Andreas Krause

ETH Zurich · Tufa Labs

Research Digest·6 November 2026·2 min read

Illustration: Language models can improve coding answers without checking target correctness

In one line

Test-Time Self-Play beats verifier-based methods for test-time scaling by training a solver on generated synthetic problems, improving by 35% with no verifier.

Before you read this

Three ideas this paper builds on

test-time scalingAllocating extra computation during inference to improve output quality.105 papers →synthetic data generationCreating training examples artificially, often using an LLM, to overcome real-data scarcity or privacy constraints.76 papers →verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.305 papers →

Read it as

When a large language model faces a hard problem, extra computation at deployment can help find a better answer. Most methods need a verifier, a checker that decides whether an answer is right; in many real situations none exists.

Test-Time Self-Play (TTSP) needs no verifier. The model plays two roles: one generates new synthetic problems in the same style, the other solves them. A Python interpreter tells the solver whether its solution runs correctly, so training needs no judge for the original problem.

On competitive programming tasks, TTSP raised the average success rate by 35 percentage points. Its advantage grows as the verifier weakens, except when the base model almost never solves the original problem. Extra compute helps even without a checker.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

test-time scaling asymmetric self-play curriculum learning synthetic data generation verification group relative policy optimization