A new paper has language models plan several steps ahead before answering, instead of guessing in one shot.
The method, called ReHoPER, is inference-only, meaning it needs no retraining, no labeled data, and no task-specific prompt engineering. A model generates a batch of candidate follow-up questions, picks one to answer, then replans based on what it just learned, repeating that cycle in a rolling window until it settles on a final answer. The same generic instructions ran across every dataset and model tested, including iLLC, a new benchmark the authors built specifically to measure compositional reasoning, where a correct answer depends on chaining several facts or steps together. The paper reports ReHoPER beating strong baselines overall, with the widest margins on the most compositional tasks.
Most reasoning upgrades for language models come from fine-tuning or prompts hand-tuned for one specific task, and those rarely transfer elsewhere. ReHoPER's pitch is that one fixed instruction set works across different problems and different models, which, if it holds up, makes it a cheap add-on for systems already running in production rather than a retraining project.
It joins a growing pile of inference-time reasoning tricks that try to buy accuracy without touching the model's weights, and like those methods, the real test is whether the gains survive contact with benchmarks the authors did not build themselves.