C-voting: Confidence-Based Test-Time Voting without Explicit Energy Functions

2026-04-15 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Researchers introduce C-voting, a novel test-time scaling strategy for neural network models employing latent recurrent processing. This method is designed for recurrent models that generate multiple latent candidate trajectories, selecting the trajectory that maximizes the average of top-1 prediction probabilities to reflect model confidence. C-voting offers a significant advantage by not requiring an explicit energy function, unlike prior energy-based voting strategies. When applied to Sudoku-hard tasks, C-voting achieves 4.9% higher accuracy compared to energy-based methods. The paper also presents ItrSA++, a simple attention-based recurrent model with randomized initial values, which, when combined with C-voting, demonstrates superior performance over the Hierarchical Reasoning Model (HRM) on Sudoku-extreme (95.2% vs. 55.0%) and Maze (78.6% vs. 74.5%) tasks.

Key takeaway

For research scientists developing or deploying recurrent neural networks for reasoning tasks, C-voting offers a robust method to enhance model accuracy at test time. You should consider integrating C-voting, especially for models without explicit energy functions, as it has shown significant performance gains on challenging benchmarks like Sudoku and Maze solving. This approach could simplify model design while improving performance.

Key insights

C-voting enhances recurrent neural network performance by selecting the most confident latent trajectory without needing explicit energy functions.

Principles

• Test-time scaling improves recurrent model performance.
• Confidence-based selection can outperform energy-based methods.

Method

C-voting initializes latent states with multiple candidates, then selects the trajectory maximizing the average of top-1 prediction probabilities.

In practice

• Apply C-voting to recurrent models lacking energy functions.
• Combine C-voting with ItrSA++ for complex reasoning tasks.

Topics

• C-voting
• Test-Time Scaling
• Recurrent Neural Networks
• Latent Candidate Trajectories
• ItrSA++

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• C-voting: Confidence-Based Test-Time Voting without Explicit Energy Functions — C-voting enhances recurrent model performance by selecting the most confident prediction from multiple random latent trajectories.
• Pulse-Driven Neural Architecture: Learnable Oscillatory Dynamics for Robust Continuous-Time Sequence Processing — Biologically-inspired oscillatory dynamics significantly improve neural network robustness to temporal input gaps.
• Recurrent neural networks approximate continuous functions — A single, fixed ReLU RNN can approximate any continuous function by running longer, enabled by a Turing machine with neural units.
• Are Candidate Models Really Needed for Active Learning? — Randomly initialized deep learning models can effectively select informative samples for active learning without candidate models.
• Titans: Learning to Memorize at Test Time — Training a dedicated neural memory module at inference time improves long-context sequence modeling on novel data.
• Intrinsic Selection and Particle Resampling for Inference-Time Scaling Beyond Domain Verifiability — Intrinsic statistics provide a robust signal for solution quality in Inference-Time Scaling for non-verifiable domains.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.