A small language model just learned to run its own research process.
Researchers fine-tuned Granite 4.1 3B, a 3-billion-parameter model, on 13,194 example actions pulled from accepted teacher search traces, teaching it to pick from seven moves in a retrieval-augmented question-answering pipeline: decompose, search, reformulate, extract, synthesize, verify, or stop. Using LoRA, a lightweight fine-tuning method, the model reached a macro-F1 of 0.6536 on 1,646 held-out examples, versus 0.1736 for the same model with no training and 0.5399 for a simple TF-IDF logistic-regression baseline. When the fine-tuned model handled both the decision-making and the final-answer writing across 149 held-out trajectories, exact-match accuracy rose from 0.7530 to 0.7946 and token F1 climbed from 0.7783 to 0.8295, compared with the untrained base model doing both jobs. Most of that gain came from the model recording more evidence along the way, not from sharper final answers when it only handled the decisions.
Most retrieval-augmented systems lean on large, expensive models to judge when to keep searching versus when to answer, since reasoning about your own uncertainty is a hard problem. This result suggests a much smaller, cheaper model can learn that judgment if it is trained on real examples of good search behavior rather than just final answers. That is the unglamorous plumbing work that could make agentic search tools cheaper to run without reaching for a bigger model every time.
Still, this is one arXiv preprint tested on 149 trajectories, not an industry benchmark, so read the specific numbers as a proof of concept rather than a verdict on small models generally.