Sakana says multiple diffusion models plus MCTS beat test-time scaling on coding and math

SakanaAILabs · x · 2026-07-21

UnMaskFork uses multiple diffusion language models and MCTS to improve coding and math

Sakana AI says its ICML 2026 paper, UnMaskFork, explores whether test-time scaling can work for masked diffusion language models (MDLMs).

• Instead of increasing randomness with temperature, the method creates diversity through model switching.
• Multiple MDLMs collaborate to unmask a single answer, while Monte Carlo Tree Search searches for promising generation paths.
• The approach requires no extra training and no model changes; it works at inference time by combining pre-trained models.
• The team says the method consistently outperforms existing test-time scaling baselines on coding benchmarks and also scales well on math tasks.
• Sakana frames the work as part of its broader “collective intelligence of AI” line of research, alongside AB-MCTS and Sakana Fugu.

Original post →

More from Models

• Google introduces three new Gemini models, including a 350 tok/s Flash-Lite — ammaar · 2026-07-22
• Moonshot’s Kimi K3 sets a new open-weights ECI record at 156 — scaling01 · 2026-07-22
• A post says six companies now beat Google’s best LLM, including two open-source models — soham_btw · 2026-07-22
• Google says Gemini 3.5 Flash-Lite is its smallest and fastest Gemini model — reach_vb · 2026-07-22
• Gemini 3.6 Flash matches 3.5 Flash on Artificial Analysis benchmark — airesearch12 · 2026-07-22
• Testing Qwen Image Model: One Prompt Generates Maps Pleasing Both India and Pakistan — NirantK · 2026-07-22

Models channel →