How Training Teaches One Model Two Kinds of Answers

Qwen3's architecture and interface define what the family can offer, but post-training determines whether those pieces cooperate in use. This chapter traces how the report builds deliberate reasoning, adds direct-response behavior, and broadens the model without losing sight of difficult tasks. The important judgment is not whether one training stage is “best”; it is how each transition changes the model's state, what capability persists, and where the authors accept a trade-off between specialized reasoning and general versatility.

Build reasoning before trying to generalize it

The first post-training problem is to make a model do more than produce a plausible answer. Qwen3's cold-start stage begins with math, code, logical reasoning, and general STEM problems paired with verified answers or code-based tests. The emphasis on verification is structural. If a response can be checked, training can distinguish a correct solution from a fluent but wrong one. If a prompt contains several loosely related subquestions or asks for unconstrained text generation, the signal is harder to interpret. The report therefore describes filtering queries that are not easily verifiable and excluding questions that the preceding model can answer correctly without chain-of-thought reasoning.

That filter defines what the first training state is meant to learn. The goal is not to maximize the number of examples or to force every request into a visible reasoning path. It is to expose the model to problems where deeper reasoning is necessary and where the outcome can be checked. The team also annotates domains to keep the mixture balanced, reserves validation queries, and generates candidate responses with QwQ-32B. Responses are removed when the final answer is wrong, the reasoning repeats itself, the solution looks like guesswork, the thinking and summary conflict, the language or style shifts inappropriately, or the example may overlap with validation material.

These criteria are more than data-cleaning details. They define the conditions under which the model is allowed to treat a long reasoning path as useful. A verbose answer is not automatically a reasoning example. A correct final answer reached by guessing is not the same as a reliable reasoning pattern. A path whose summary contradicts its own intermediate work teaches an unstable relationship between deliberation and conclusion. The cold-start stage therefore tries to establish patterns of problem decomposition and checking before asking reinforcement learning to improve them.

The authors deliberately keep this preparatory stage limited. They say the objective is to instill foundational reasoning patterns without overemphasizing immediate performance, and that minimizing both samples and training steps helps preserve room for later improvement. In system terms, the cold start initializes a reasoning-capable state; it is not supposed to finish the optimization.

Reasoning reinforcement learning then changes the feedback loop. The report collects 3,995 query-verifier pairs that were not used for cold start, are learnable for the current model, are as challenging as possible, and cover a broad range of subdomains. It applies GRPO to update the model.