Small AI teams get uneven returns from extra calls, preprint reports

Giving small language-model teams 30 calls instead of three improved arithmetic accuracy much more than performance on several other tests, according to a preprint submitted to arXiv on 28 September 2026. The authors report gains of up to 17 percentage points on GSM8K and GSMHard, both arithmetic word-problem benchmarks. On ARC, GPQA and MMLU, gains topped out at four points for every architecture tested.

Researchers Blaz Bertalanic and Carolina Fortuna compared eight agent-orchestration architectures using five instruction-tuned models with 7–9 billion parameters. Their sweep covered five short-answer benchmarks and one executable-code benchmark, with budgets reaching 30 calls. Even that shared budget was not a uniform measure of expense: the authors report a 2.1-fold difference in token cost at equal call counts.

The paper’s explanation separates generating possible answers from turning those candidates into a final answer. Its proposed decomposition attributes changes in accuracy to two components: changes in candidate coverage and changes in the downstream transformation. More opportunities to propose an answer, in other words, need not translate into more correct final answers.

According to the authors, arithmetic tasks left room to expand coverage, and a critic-guided process could turn that opportunity into gains. The multiple-choice tests presented a different problem: coverage either saturated or the downstream process failed to convert it into better results.

That helps explain why the choice of team design mattered. Proposer-Critic improved fastest and outperformed every other architecture in aggregate at the largest call budget, the authors report. Yet it ranked among the weakest designs on other tasks. No architecture won across the board.

These findings are reported in an arXiv preprint, and this account is limited to its abstract and submission record. Without the full methods and tables, it is not possible here to assess how the experiments were configured or how broadly the reported differences hold.