Teaching open LLMs to think harder, not just grow bigger

What it does

Search and Learn is a collection of scripts and YAML recipes that apply search algorithms to open LLMs at inference time. It targets “verifiable” domains like math and code, where a reward model can score each reasoning step and guide the model toward better answers without retraining the base model. The repository also includes code to train process reward models that provide this step-by-step feedback.

The interesting bit

The project explicitly takes aim at the “Bitter Lesson” idea that general methods scale with compute—but shifts that compute from training to inference. Rather than chasing billion-dollar pretraining clusters, it asks what happens if you simply let existing open models search longer and verify their own reasoning steps, much like the closed o1 model but built in the open.

Key highlights

• Supports Best-of-N sampling, beam search, and Diverse Verifier Tree Search (DVTS) via declarative YAML recipes.
• Includes training pipelines for process reward models (PRMs) that score each step of a reasoning chain.
• Focuses on verifiable problems—math and code—where automated verification makes search practical.
• Built to replicate the scaling results from the accompanying Hugging Face blog post and the underlying DeepMind paper.
• Designed to be simple and minimal: mostly scripts and recipes directories with little abstraction overhead.

Verdict

Worth a look if you’re researching inference-time scaling or trying to squeeze stronger reasoning out of open models you already have. Skip it if you need general-purpose chat improvements or production-ready inference infrastructure.