LongZero trains long-context reasoning with RL on data the model writes itself

The paper builds on two lines of work. Reinforcement learning is already an effective way to strengthen long-context reasoning, but it needs training data that are long, diverse, challenging and verifiable. RL from zero data, where a model learns from tasks it proposes for itself, has achieved excellent performance by the authors' account, but they say existing methods of that kind target short-context tasks. LongZero is their attempt to carry the zero-data approach over to long contexts.

The authors list four properties that long-context RL data needs: it must be long, diverse, challenging and verifiable. They then rule out the three usual ways of getting it. Human annotation of such data is expensive. Naturally occurring data are too scarce to sustain RL. Existing synthesis methods depend on external corpora or stronger teacher models.

The third objection matters most for teams training strong models. A teacher-based pipeline needs a model stronger than the one being trained, and the authors argue that both teachers and suitable corpora get harder to find as the trained model improves. The paper states that dependency directly.

The core design choice reverses the usual order of synthesis. The authors contrast LongZero with methods that write questions and answers over existing documents. In LongZero the model first produces the question, the answer and the supporting evidence in a compact structured form. Only after that does a separate step expand the evidence into a long multi-document context.

A likely reason for this ordering is verifiability, though the abstract does not spell it out. The answer and its evidence exist before the long context does, so the label is fixed when the task is built and does not depend on a model reading a long document correctly. The abstract does not explain how the reward is computed from that answer. It could use exact match, normalized string match or a model-based judge, and the source does not say which.

The abstract leaves several construction details open. It does not say how long the expanded contexts are or how many documents each contains. It does not say whether the expansion adds distractor documents that do not support the answer, or whether the expanding model is the policy being trained or a frozen copy. Each of these choices affects how hard the tasks are and whether the expansion can introduce text that contradicts the answer written in step 1.

LongZero mixes two kinds of task, both written by the model: multi-hop reasoning tasks and multi-round retrieval tasks. The abstract names these task types but does not define them or give examples. It also does not give the mixing ratio between them, the number of hops or rounds, or how difficulty is controlled as training goes on.

Optimization uses GRPO. The abstract does not say why GRPO was chosen over other RL algorithms. It reports no GRPO hyperparameters, group sizes or training-step counts.

The abstract ends partway through the sentence "We evaluate LongZero on two LLM". It does not name the two models, their sizes or their context windows. It names no long-context benchmarks. It gives no scores, no deltas against the base models and no comparison with corpus-based or teacher-based synthesis pipelines.

That leaves the second half of the angle unanswered: the source gives no figure for how much LongZero improves long-context benchmarks compared with corpus-based synthesis. Any such figure would have to come from the full paper. Anyone citing results should check them against the paper's own tables, not secondary summaries.

Engineers deciding whether to reproduce or adapt LongZero can take a short list of questions from the abstract to the full paper. The method's main draw is that it needs no external data. The same independence creates specific failure modes worth checking.

If the full paper answers these questions, the practical result is a long-context RL data pipeline that needs no external corpora or stronger teacher models. For teams whose models already match the best available teachers, that removes a dependency the authors say gets harder to satisfy over time. Until the evaluation section is available, LongZero is a described method whose effect size is not reported.