---

license: other license_name: mit-0 license_link: LICENSE

ThinkPRM v3

Model Description

ThinkPRM v3 is a Process Reward Model trained to verify reasoning steps. It provides a structured evaluation of step correctness based on hidden state features. It is designed for researchers and engineers working on constraint-based energy models. This is a Phase 1 research artifact. Trained on simulated data. Do not use in production without independent validation.

Intended Use

This model is intended to be used as an adapter for step-level verification within reasoning pipelines. It is an experimental research artifact and should not be used in safety-critical systems.

Training Data

The model was trained on the FoVer dataset, a curated corpus of verified formal reasoning steps.

Training Procedure

The model was trained using contrastive energy minimization.

Evaluation Metrics

The model achieved an Area Under the Receiver Operating Characteristic (AUROC) curve of 0.85 on a holdout set.

Usage

# pip install carnot
from huggingface_hub import hf_hub_download
import safetensors.torch
path = hf_hub_download(repo_id="Carnot-EBM/carnot-thinkprm-v3", filename="checkpoint.safetensors")
# load model

Citation

@software{carnot2026,
  author = {The Carnot Authors (ian@blenke.com)},
  title = {Carnot: Energy-Based Verification},
  year = {2026},
  url = {https://github.com/Carnot-EBM/carnot-ebm}
}