Tiny in Parameters, Not in Compute: What a Recursive Reasoning Model's Puzzle Score Credits to Recursion, Why the Ablations Split by Task, and the Matched Comparisons That Would Assign the Rest

Abstract

Life of Research

Recursive reasoning models, which apply one small network many times to a latent state, report accuracies on Sudoku, maze and ARC-AGI puzzles above those of most frontier language models in their comparison tables, with 5 to 27 million parameters. The result is usually credited to recursion: iterating a latent state is said to buy the depth that reasoning needs. A 2025-2026 line of ablations, mechanistic studies and controlled re-evaluations disputes that credit. In this analysis the author reviews and synthesises published results to show that the disagreement dissolves once two things are separated: what the word recursion names, and which task the score comes from. Recursion names three functions - computation through iterated depth, regularisation through parameter sharing, and a source of candidate trajectories for a learned selector - and a reported score also carries per-task identity conditioning, training on the evaluation tasks' demonstrations, augmentation with voting, and an optimisation recipe. On Sudoku-Extreme the evidence credits iteration directly: one-pass models of equal size solve almost no puzzles, and running the recursion past its training horizon recovers most failures, though at matched effective depth part of the tied model's advantage is regularisation. On ARC-AGI the measured contribution of in-step recursion is small, the score falls to zero without the task identifier, and the evidence fits regularisation at least as well as depth. Test-time gains from stochastic restarts are selection gains whose size, on the benchmarks where both are reported, tracks how well the model's halting head verifies an answer. Parameter count is the wrong denominator for all of it. The paper consolidates the measured values in a table and a figure, states a crediting procedure as an algorithm, and specifies the matched comparisons that would assign what remains.

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 109 papers by Pranay Mahendrakar, or his ORCID record.