README.md

metadata

library_name: transformers
tags:
  - reward-model
  - prm
  - generative reward model
  - process supervision
  - chain-of-thought
  - verification
  - math reasoning
  - code verification

Model Card for ThinkPRM-1.5B

ThinkPRM-1.5B is a generative Process Reward Model (PRM) based on the R1-Distill-Qwen-1.5B architecture. It is fine-tuned to perform step-by-step verification of reasoning processes (like mathematical solutions) by generating an explicit verification chain-of-thought (CoT) that involves labeling every step. It is designed to be highly data-efficient, requiring significantly less supervision data than traditional discriminative PRMs while achieving strong performance.

Here's an example of the model output:

Model Details

Model Description

ThinkPRM-1.5B provides step-level verification scores by generating natural language critiques and correctness judgments for each step in a given solution prefix. It leverages the underlying reasoning capabilities of the base Large Reasoning Model (LRM) and enhances them through fine-tuning on a small (1K examples) dataset of synthetically generated verification CoTs. These synthetic CoTs were produced by prompting QwQ-32B-Preview and filtered against ground-truth step labels from the PRM800K dataset to ensure quality.

The model uses a standard language modeling objective, making it interpretable and allowing it to scale process verification compute by generating longer or multiple verification CoTs. It demonstrated superior performance compared to LLM-as-a-judge and discriminative PRM baselines (based on the same R1-Distill-Qwen-1.5B model but trained on ~100x more labels) on benchmarks including ProcessBench, MATH-500, AIME '24, GPQA-Diamond, and LiveCodeBench.

• Finetuned from model [optional]: R1-Distill-Qwen-1.5B

Model Sources [optional]

• Repository: Github
• Paper: Process Reward Models that Think (arXiv:2504.16828)

Direct Use

ThinkPRM-1.5B is intended for verifying the correctness of step-by-step reasoning processes. Primary uses include:

• Scoring Solutions: Assigning step-level or overall scores to candidate solutions for ranking in Best-of-N sampling or guiding tree search in reasoning tasks.
• Generating Verification Rationales/CoTs: Producing detailed chain-of-thought verifications that explain why a particular step is correct or incorrect, aiding interpretability.
• Standalone Verification: Evaluating the correctness of a given problem-solution pair.

The model has been evaluated on mathematical reasoning (MATH, AIME), scientific QA (GPQA), and code generation (LiveCodeBench). See our paper for more details.

Limitations

• Overconfidence: Generative PRMs like ThinkPRM can sometimes produce scores clustered near 0 or 1, potentially not reflecting true uncertainty
• Step Label Interference: The autoregressive nature might cause an early incorrect step judgment to negatively bias the evaluation of subsequent steps.
• Sensitivity to Formatting/Prompting: Performance might be sensitive to the exact format of the input solution and the prompt used for verification (though fine-tuning likely reduces this compared to LLM-as-a-judge).

How to Get Started with the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
from vllm import LLM, SamplingParams

model_id = "launch/ThinkPRM-1.5B" # Replace with actual model ID on Hub
tokenizer = AutoTokenizer.from_pretrained(model_id)
llm = LLM(model=model_id, max_model_len=16384)

# Example problem and solution
problem = "Solve for x: 2x + 3 = 7"
prefix = "Step 1: Subtract 3 from both sides: 2x = 4\nStep 2: Divide by 2: x = 2"

# Format the prompt -- this is the prompt used for the results in the paper.
prompt = f"""You are given a math problem and a proposed step-by-step solution:

[Math Problem]

{problem}

[Solution]

{prefix}

Review and critique each step in the proposed solution to determine whether each step is correct. If the solution is incomplete, only verify the provided steps
"""

prompt = tokenizer.apply_chat_template([
    {'role': "user", "content": prompt}
], tokenize=False, add_generation_prompt=True) # construct prompt 

# Set sampling parameters
sampling_params = SamplingParams(
    temperature=0.0,
    max_tokens=4096,
    stop=None
)

# Generate the verification
outputs = llm.generate(prompt, sampling_params)
verification_cot = outputs[0].outputs[0].text

print(verification_cot)
# Alright, I've got this math problem to solve: 2x + 3 = 7. And there's a proposed solution in multiple steps that I need to review and critique.
# ...
# Yes, that checks out. So, the solution is correct.

# Alternatively, I could think about it differently. Maybe I could first divide both sides by 2, but that would give me x + 1.5 = 3.5, and then subtract 1.5 from both sides to get x = 2. But that's more complicated than the given solution.

# The given solution is straightforward and efficient.

# So, in conclusion, both steps are correct, and the solution is accurate.

# But according to the instructions, since both steps are correct, I don't need to go into more detail. I just need to state that each step is correct.

# So, my output should be:

# Step 1: The step is to subtract 3 from both sides, resulting in 2x = 4. This is a correct operation to isolate the term with x. The step is \boxed{correct}
#
## Step 2: The step is to divide both sides by 2, resulting in x = 2. This is the correct operation to solve for x. The step is \boxed{correct}

# </think>
# Is the solution correct? Yes