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.
PRO- C-voting picks the recurrent trajectory with highest average top-1 confidence, no energy function needed.
- Beats AKOrN's E-voting by 4.9 points on Sudoku-Hard (94.4% vs 89.5%) on the same checkpoint.
- New 3M-parameter ItrSA++ hits 95.2% on Sudoku-Extreme where HRM reaches 55.0%.
- Also beats HRM on Maze-Hard 78.6% vs 74.5% using roughly one-ninth the parameters.
- Fails when the model is confidently wrong, as seen in Maze-Hard error distributions.
- Anonymized code released for HRM, AKOrN and ItrSA++ integrations.
Recurrent reasoning models like HRM and AKOrN 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, 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.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
