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

2026-04-16 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

Researchers from The University of Tokyo introduce C-voting, a novel confidence-based test-time scaling strategy for recurrent neural network models, particularly those with latent recurrent processing. C-voting addresses the limitation of energy-based voting (E-voting) by not requiring an explicit energy function, making it broadly applicable. The method initializes latent states with multiple random candidates and selects the trajectory that maximizes the average of top-1 prediction probabilities, reflecting model confidence. C-voting achieves a 4.9% higher accuracy on Sudoku-hard tasks compared to E-voting. The team also developed ItrSA++, a lightweight recurrent model with approximately 3 million parameters, which, when combined with C-voting, significantly outperforms the Hierarchical Reasoning Model (HRM) on Sudoku-extreme (95.2% vs. 55.0%) and Maze-hard (78.6% vs. 74.5%) tasks, and AKOrN on Sudoku-hard (94.4% vs. 89.5%).

Key takeaway

For AI Scientists and Machine Learning Engineers developing recurrent models for complex reasoning tasks, C-voting offers a robust test-time scaling strategy. You should integrate C-voting to enhance model performance without additional training, especially for models without explicit energy functions. This approach can significantly improve accuracy on benchmarks like Sudoku and Maze, even with lightweight architectures like ItrSA++, making it a valuable technique for optimizing inference.

Key insights

C-voting enhances recurrent model performance by selecting the most confident prediction from multiple random latent trajectories.

Principles

• Confidence-based voting is model-agnostic.
• Randomized initialization improves generalization.
• Test-time scaling can compensate for iteration limits.

Method

C-voting samples multiple initial latent states, updates them recurrently, calculates the average top-1 probability for each trajectory, and selects the candidate with the highest average confidence for the final prediction.

In practice

• Apply C-voting to recurrent models lacking explicit energy functions.
• Use randomized latent state initialization for performance gains.
• Consider ItrSA++ for efficient, high-performance reasoning tasks.

Topics

• Confidence-based Voting
• Test-time Scaling
• Recurrent Models
• ItrSA++ Architecture
• Energy-based Voting

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

Related on AIssential

• C-voting: Confidence-Based Test-Time Voting without Explicit Energy Functions — C-voting enhances recurrent neural network performance by selecting the most confident latent trajectory without needing explicit energy functions.
• Entropy Centroids as Intrinsic Rewards for Test-Time Scaling — Temporal patterns of LLM uncertainty, captured by Entropy Centroids, predict response quality without external rewards.
• How Can A Model 10,000× Smaller Outsmart ChatGPT? — Iterative reasoning with small, recursive models can outperform massive, feed-forward networks in complex logical tasks.
• The Smallest Model Won One of My Tests, and Other Things Benchmarks Won’t Tell You — Real-world AI performance diverges from benchmarks, revealing critical differences in obedience, calibration, and consequence awareness.
• Simple Approximation and Derivative Free Inference-Time Scaling for Diffusion Models via Sequential Monte Carlo on Path Measures — URGE improves diffusion model sample quality via derivative-free pathwise importance reweighting and resampling.
• 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 cs.AI updates on arXiv.org.