When asked about a long report, a language model often fails because it processes the whole text at once, making it hard to connect far-away clues.
LongCapHarness breaks the question into small steps, handles each step with a focused model call, and automatically builds this step-by-step plan by checking against correct answers on a small set, without extra data or model retraining.
On two standard long-text tests it beat methods that require collecting extra training data, and its development cost was less than half. This shows that existing models can reason over long texts if guided properly.