Reproduce “Parallel Test-Time Scaling for Latent Reasoning Models” on OpenResearch

About this paper

Runyang You, Yongqi Li, Meng Liu, Wenjie Wang, Liqiang Nie, Wenjie Li

This research enables parallel test-time scaling for latent reasoning models, which previously lacked stochastic sampling and aggregation mechanisms in continuous latent spaces. It introduces methods for generating diverse latent trajectories and a novel Latent Reward Model (LatentRM) for their evaluation, demonstrating accuracy gains without requiring model retraining.

Key insights

• Both MC-dropout and AGN effectively scale latent reasoning models, with MC-dropout consistently achieving higher coverage by leveraging epistemic uncertainty for exploration.
• The LatentRM successfully discriminates promising latent trajectories, enabling significant accuracy improvements through Best-of-N selection and Beam Search compared to simple Majority Voting.
• Latent reasoning augmented by parallel TTS achieves a superior balance of accuracy and computational efficiency compared to explicit Chain-of-Thought, requiring fewer decoding steps for comparable performance.

Summary by alphaXiv.