The honesty problem in test-time compute
A lot of "reasoning engine" repos share two flaws: they never actually call a model, and they publish benchmark tables nobody can reproduce. HyperReason v1 was, by its own README's admission, exactly that.
v2.0.1 is a full honest rebuild: the engine genuinely drives real LLMs and publishes transparent evaluations — including runs where MCTS underperformed greedy decoding. No fake fallbacks; every number traces to code.
What it is
An adaptive-entropy Monte Carlo tree search engine for LLM test-time compute, on PyPI:
pip install hypermcts # imports as hyper_reason
Supported providers out of the box:
| Provider | Backend |
|---|---|
| Z.AI / GLM |
GLMBackend (glm-4.6, glm-4-flash, glm-4-plus) |
| OpenAI / DeepSeek / Groq |
OpenAIBackend (gpt-4o, deepseek-reasoner, vLLM, LM Studio) |
| Ollama (local) |
OllamaBackend (llama3:8b, deepseek-r1:8b, qwen2.5:7b, mistral) |
| HuggingFace (local) | TransformersBackend |
| Mock (offline) |
MockBackend — deterministic heuristic sampler, no API key needed |
Minimal usage
from hyper_reason import wrap_model, GLMBackend
model = wrap_model(GLMBackend(model="glm-4-flash"))
# run the AE-MCTS search on your problem
There's also a CLI, a live web demo (rudra496.github.io/hyper-reason), and an offline mock backend so you can try the whole pipeline without any API key.
The honest eval contract
- Raw JSONL eval traces are published in the repo (
eval/runs/) — wins and losses - The eval contract (
eval/CONFIG.md) documents what was measured and how - Citable via Zenodo: 10.5281/zenodo.22118810
That's the actual point of the project: if MCTS doesn't beat greedy on your problem, the traces will tell you — which is more than most engines will admit.
Links
- Source (MIT): github.com/rudra496/hyper-reason
- PyPI: pypi.org/project/hypermcts
- Live demo: rudra496.github.io/hyper-reason
Top comments (0)