Carnot-EBM/carnot-thinkprm-v3

ThinkPRM v3

Model Description

ThinkPRM v3 is a step-level Process Reward Model trained to verify reasoning traces produced by an upstream LLM. It is one verifier in Carnot's broader k=15 verifier ensemble (the "Tier 0a step-level probe"); the ensemble's joint AND-composition is what produces Carnot's headline FoVer numbers, not this checkpoint alone.

This model card describes the standalone ThinkPRM v3 checkpoint. For the full verifier-ensemble framework, see github.com/Carnot-EBM/carnot-ebm.

• Model ID: Carnot-EBM/carnot-thinkprm-v3
• Architecture: Process Reward Model adapter; step-level classifier over hidden-state features
• Framework: Python / PyTorch + safetensors
• License: Apache 2.0
• Status: Research artifact. Not a safety-critical production component.

Intended Use

ThinkPRM v3 is intended as an adapter for step-level verification within reasoning pipelines -- for example, scoring intermediate steps of a chain-of-thought trace as part of Carnot's verifier ensemble.

It is not a hallucination detector on its own. Carnot's verification guarantees come from AND-composing multiple verifiers with different inductive biases (formal-claim, schema, semantic, energy-ranking, ...). A single PRM is a single axis of evidence.

Training Data

The model was trained on the FoVer dataset -- a curated corpus of formally-verified reasoning steps with positive/negative pairs.

Training Procedure

Contrastive energy minimization over (correct, violating) step pairs from the FoVer corpus, using Carnot's standard verifier-training harness.

Evaluation

The standalone v3 checkpoint reports AUROC = 0.85 on a held-out split, measured at the time of training. For current ensemble-level FoVer numbers, see the Carnot technical report at docs/arxiv-paper/main.pdf; the AUROC of the full ensemble in production conditions is reported there and is the authoritative number for citation purposes.

Limitations

• Single-axis evidence. A PRM verifies that individual reasoning steps look plausible against a learned step-correctness distribution. It does not verify formal claims, schemas, or symbolic constraints; those signals come from sibling verifiers in the ensemble.
• In-distribution calibration. The reported AUROC was measured on a FoVer-distributed held-out split. Out-of-distribution behavior (e.g., on highly-optimized SOTA outputs) is constrained by an energy-ordering inversion documented in the Carnot technical report; this checkpoint inherits that constraint.
• Not a safety component. Do not use this model in a safety-critical loop without independent validation appropriate to your domain.

Usage

# pip install carnot-ebm
from huggingface_hub import hf_hub_download
import safetensors.torch

path = hf_hub_download(
    repo_id="Carnot-EBM/carnot-thinkprm-v3",
    filename="checkpoint.safetensors",
)
state_dict = safetensors.torch.load_file(path)
# Load into your PRM adapter and run as a step-level verifier.

For the ensemble-level verifier pipeline rather than this single checkpoint, install the framework:

pip install carnot-ebm

from carnot.pipeline import VerifyRepairPipeline

pipeline = VerifyRepairPipeline()
result = pipeline.verify("What is 15 + 27?", "15 + 27 = 42")

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}
}