---
title: "University of Tokyo's ItrSA++ Beats HRM on Extreme Puzzles With 9x Fewer Parameters"
subtitle: "A confidence-driven voting scheme lets any recurrent reasoner scale at test time, and a 3M-parameter model built around it beats HRM by 40 points on Sudoku-Extreme."
url: "https://alphasignal.ai/news/university-of-tokyo-s-itrsa-beats-hrm-on-extreme-puzzles-with-9x-fewer"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-04-15T00:00:00.000Z"
updated: "2026-08-23T22:59:27.858Z"
type: "Paper"
topics:
  - "Reasoning"
  - "Llms"
subtopics:
  - "Test Time Compute"
  - "Math Reasoning"
  - "Small Models"
upvotes: 1777
source_url: "https://arxiv.org/abs/2604.13521"
comments: 0
truncated: true
---

# University of Tokyo's ItrSA++ Beats HRM on Extreme Puzzles With 9x Fewer Parameters

> A confidence-driven voting scheme lets any recurrent reasoner scale at test time, and a 3M-parameter model built around it beats HRM by 40 points on Sudoku-Extreme.

Recurrent reasoning models like [HRM](https://arxiv.org/abs/2506.21734) and [AKOrN](https://openreview.net/forum?id=nwDRD4AMoN) have shown that iterating the same block over a latent state can crack puzzles that stump frontier LLMs. A new paper from the University of Tokyo, [C-Voting](https://arxiv.org/abs/2604.13521), tackles the next question: once you have such a model, what is the best way to spend extra compute at inference? The authors propose a simple selection rule that beats the previous state-of-the-art voting method and, paired with a stripped-down architecture called ItrSA++, sets new numbers on Sudoku-Extreme and Maze-Hard using roughly a ninth of HRM's parameters.

## Why voting matters for latent recurrent models

Latent recurrent models apply the same layer to a hidden state over and over. Test-time scaling can be realized in two ways: running more iterations at inference than in training, or selecting the best trajectory from randomly initialized runs. The first path saturates quickly and is sequential. The second, voting, parallelizes cleanly and keeps improving even when depth stops helping.

AKOrN introduced **energy-based voting** (E-voting): sample many random initial latent states, run the recurrence, and keep the trajectory whose final state has the lowest energy. Using 4096 candidates gave roughly a 40% jump in board accuracy on Sudoku. The catch is that it only works when the model is literally defined as gradient descent on a scalar energy function. HRM, recurrent transformers, and most other reasoning architectures have no such function, which shuts them out of this powerful scaling axis.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [C-Voting: Confidence-Based Test-Time Voting without Explicit Energy Functions](https://arxiv.org/html/2604.13521)
- [Hierarchical Reasoning Model](https://arxiv.org/abs/2506.21734)
- [Artificial Kuramoto Oscillatory Neurons](https://openreview.net/forum?id=nwDRD4AMoN)
- [HRM official repository](https://github.com/sapientinc/HRM)