RECURSIVE AGENTIC REASONING

Recent advances in test-time reasoning—such as iterative refinement, problem decomposition, and repeated sampling—are often studied in isolation, making it difficult to compare their effectiveness. This paper introduces a unified framework that treats these methods as recursion operators over an agent's reasoning trace: GROW (deepening a single path), PRUNE (decomposing and recomposing the problem), and BRANCH (sampling and selecting alternative paths). By implementing these operators within a shared evaluation harness, the authors provide a controlled, apples-to-apples comparison of how different computational shapes impact model performance.

The study evaluates these operators across five benchmarks and three frontier models, totaling 14 distinct model-benchmark settings. To ensure fairness, the authors implement a paired evaluation protocol, scoring all operators on the exact same set of items and excluding infrastructure-related failures. This methodology addresses two common pitfalls in evaluation: unpaired scoring, which can lead to comparing methods on different problem sets, and misclassifying infrastructure failures as model errors, which unfairly penalizes more compute-intensive methods.

The results demonstrate that BRANCH is the most effective operator, improving accuracy in all 14 settings with an average gain of 5.98 percentage points. In contrast, GROW and PRUNE show inconsistent performance and are occasionally detrimental. A key analytical insight is that BRANCH's success is strongly correlated with the baseline's rate of empty, budget-exhausted outputs (r = 0.72). This indicates that a significant portion of the performance gain comes from recovering answers that a single-pass generation fails to produce due to token limits, rather than solely from exploring multiple reasoning paths. The authors also find that adaptive operator routing is not supported by their data, as BRANCH consistently dominates at this level of abstraction.