A unified chassis for tree search, agents, and inference scaling

What it does

LLM Reasoners is a Python library that reframes complex LLM reasoning as a modular search problem. It provides a standard object model—WorldModel, SearchConfig, and SearchAlgorithm—and implements more than a dozen published methods, from Chain-of-Thought to Monte-Carlo Tree Search and Process Reward Model guided decoding. You inherit the base classes to describe your domain, then plug in an existing search strategy.

The interesting bit

The authors treat reasoning algorithms as interchangeable parts rather than monolithic scripts. They also integrated SGLang to accelerate inference and ship a built-in visualizer that renders search trees in a browser, which turns opaque MCTS rollouts into something you can actually debug.

Key highlights

• Implements a wide catalog of methods: RAP/MCTS, Tree-of-Thoughts (BFS and DFS), Beam Search, Grace Decoding, PromptAgent, DRPO, StructChem, and others.
• Abstracts every method into three reusable primitives: a world model, a reward/search config, and a search algorithm.
• Backends include Hugging Face Transformers, OpenAI API, Exllama, fairscale, llama.cpp, and SGLang for structured generation.
• Includes a web-based visualizer meant to diagnose reasoning traces with minimal code.
• Claims faithful reproduction of original paper results for ToT, Guided Decoding, and GRACE on GSM8k and Game of 24.

Caveats

• The README reads more like a news feed and collection of example notebooks than formal documentation, so the boundary between stable API and experimental recipe is unclear.
• The “100x speed-up” attributed to SGLang appears only in a news bullet without supporting benchmark tables.
• Most algorithms live in individual examples/ directories, so heavy reading of the source is required to understand the abstraction.

Verdict

Researchers who need to compare or compose reasoning algorithms—especially tree search and inference-time scaling—will find this a useful shared substrate. If you just need a single, well-documented prompting utility, the breadth here may be overkill.